High-content phenotypic analysis of a C. elegans recombinant inbred population identifies genetic and molecular regulators of lifespan

Abstract

Lifespan is influenced by complex interactions between genetic and environmental factors. Studying those factors in model organisms of a single genetic background limits their translational value for humans. Here, we mapped lifespan determinants in 85 C. elegans recombinant intercross advanced inbred lines (RIAILs). We assessed molecular profiles – transcriptome, proteome, and lipidome – and life-history traits, including lifespan, development, growth dynamics, and reproduction. RIAILs exhibited large variations in lifespan, which positively correlated with developmental time. Among the top candidates obtained from multi-omics data integration and QTL mapping, we validated three longevity modulators, including rict-1, gfm-1, and mltn-1. We translated their relevance to humans using UK Biobank data and showed that variants in GFM1 are associated with an elevated risk of age-related heart failure. We organized our dataset as a resource (https://systems-genetics.org/cel_longevity) that allows interactive explorations for new longevity targets.

Introduction

An intricate interplay of genetic, epigenetic, and environmental factors determines the lifespan of an organism¹. Over the past few decades, extensive research has been carried out to decipher the underlying mechanisms governing longevity with gain- and loss-of-function (G/LOF) studies in different model organisms. However, a prevalent limitation in the evaluation of the effects of mutations and environmental perturbations is the predominant reliance on animal models with a single genetic background for analysis². This restricts the translational value and generalizability of these studies^3,4. Although such a strategy should ideally be employed in vertebrate models, the scope of the experimental testing in multiple genetic backgrounds combined with the ethical hurdles associated with such massive animal experimentation make this approach unrealistic. To overcome these constraints, the roundworm C. elegans has emerged as an attractive model for aging research, offering one of the best compromises between the simplicity of cell models and the complexity of vertebrate models⁵. In this regard, worm genetic reference populations (GRPs), such as the recombinant inbred lines (RILs)^6-9 and recombinant inbred advanced intercross lines (RIAILs)^10,11, have been increasingly used in the past years. These panels consist of inbred strains derived from crosses between two genetically divergent parental strains^11,12. With this study design, the recombination between the parental strains allows for fine mapping of quantitative trait genes (QTGs) — genes that explain the variation in certain quantitative traits¹³. Furthermore, the availability of genotype data, and the ability to reproduce identical individuals, allow for the in-depth interrogation of quantitative traits at the systems level in several environmental conditions and at multiple physiological levels.

Here, we used a worm GRP consisting of 85 genetically diverse RIAILs derived from crosses between two parental strains, i.e., QX1430 (with an N2 Bristol background) and CB4856 (Hawaii)^11,14. To investigate the alleles contributing to subtle variations in longevity-related phenotypes across this worm GRP, we measured their transcriptome, proteome, lipidome, and lifespan (Figure 1). In addition, we employed a high throughput fully automated microfluidic-based robotic phenotyping platform (see Nagi Bioscience SA https://nagibio.ch/), to collect other life-history phenotypes including body size, developmental dynamics, activity, as well as parameters related to worm reproduction and fertility. Integration of these omics and phenotypic data allowed the identification of a genetic locus associated with lifespan variations in these RIAILs. Within these loci, we identified gfm-1, rict-1, and mltn-1 as candidate longevity regulators and further validated gfm-1 and mltn-1 as bona-fide longevity regulators through loss-of-function studies. To assess the clinical significance of these candidate longevity genes in humans, we explored the UK Biobank data to show that variants in the human GFM1 correlate with age-related heart failure. While our study focused on longevity regulation, we generated an extensive map of the molecular and phenotypic landscape in the RIAILs population. This resource will be valuable for subsequent in silico hypothesis generation and we have made it publicly available through an interactive open-access web resource (https://systems-genetics.org/cel_longevity).

Figure 1. Overview of the study design.

85 recombinant intercross advanced inbred lines (RIAILs) derived from the crossing of QX1430 (N2 Bristol background, with deletions of confounder genes) and CB4856 strain (Hawaii) were used. Lifespan, early life-history traits, transcriptome, proteome, and lipidome were collected for each strain. We applied a systems genetics approach to study relations between different phenotypes and molecular traits to identify candidate lifespan genes. After prioritization of the candidate genes, we validated them through wet lab experiments and human population genetics (e.g. UK Biobank). We collected data from all RIAILs using three pipelines. In the first, worms were cultured and scored for their lifespans; in the second, they were cultured in a microfluidic device for ~100 h to collect early life-history traits, including body size, moving shapes, developmental parameters, reproduction, and fertility; and in the last, they were cultured to reach L4/young adulthood, and collected for multi-omics measurements (transcriptomics, proteomics, and lipidomics). Examples of variants: SNV: single-nucleotide variant. INDEL: insertions and deletions. GSEA: gene set enrichment analysis; Chr. I-X: chromosome I to X; LC-MS: liquid chromatography–mass spectrometry; L1- L4: larval stage 1 to 4; YA: young adulthood; GA: gravid adulthood. Max. Lifespan: maximum lifespan; Avg. Lifespan: average lifespan.

Results

85 RIAILs exhibit extensive variations in the lifespans and life-history traits

To determine the extent to which genetic background can influence longevity, we first assessed the lifespans of 85 RIAILs by manually scoring them on plates¹⁴. The range of average lifespan of RIAILs was from 13 to 21 days (Figure 2A). Although the majority of RIAIL average lifespans lay between those of the parentals, nine strains’ lifespans were shorter-lived than CB4856 and four strains lived longer than N2, suggesting the presence of transgressive segregation (Figures 2A and 2B). Also, we observed a similar pattern for early (25% dead and 75% alive), mid (50% mortality), and late (75% dead and 25% alive) time to mortality (Figure 2C), that is, the age in days at which 25%, 50%, or 75% of the worms died, respectively. Studies in different organisms have shown that diverse trade-offs dominate life-history traits^15-17. As a consequence, various organisms display correlations among different life-history traits, such as lifespan and fecundity¹⁸, development and lifespan¹⁹, body size and longevity^20-22. We therefore monitored a number of life-history traits across the early life stage of the RIAILs (approximately 100 h after egg hatching), including maximum body size, developmental time, sexual maturity (emergence of the 1^st egg), fertility (rate of egg accumulation), the emergence of the 1^st larvae, (following the emergence of the 1^st egg), the rate of progeny accumulation, and moving shapes, using an innovative whole-organism high-content screening technology (see https://nagibio.ch/) (Figures 2D-2E, S1, and Table S1). RIAIL strains displayed large variations in early life-history phenotypes, including developmental dynamics, reproduction (Figures 2D and S1B-S1D), and activity (Figure 2E). The N2 strain had a protracted growth period, characterized by delayed attainment of ½ maximal body size and greater overall body size when compared to the CB4856 strain (Figures 2D and S1A). Both strains had comparable timing of reproductive maturation and showed no disparities in various fertility measures (Figures 2D and S1C-S1D).

Figure 2. RIAILs exhibit extensive variation in lifespan and life-history traits.

(A) Bar plot showing the average lifespan of 85 RIAIL strains (60 worms/strain) and two wild-type parental strains (600 worms/strain). Grey bars: RIAILs; Orange bar: N2 (Bristol); Blue bar: CB4856 (Hawaii, HW). Examples of strains with different average lifespans are labeled. (B) Examples of differences in the lifespan of RIAILs (QX537, QX520, and QX597; grey) and parental N2 (orange) and CB4856 (blue) strains. (C) Violin plots of the RIAIL lifespan traits. Dots represent the average value of the trait for the two parental strains. (D) Violin plots of early life-history phenotypic traits. Dev. time: developmental time; Egg acc.: egg accumulation; Prog. acc.: progeny accumulation. Details on the number of replicates for each phenotypic trait of each RIAIL strain are provided in Table S1. (E) Violin plots of the activity life-history phenotypic traits. Shape 1: straight; Shape 2: active; Shape 3: swimming; Shape 4: supercoil. (F) Pearson correlation between lifespan and physiological traits. *: BH-adjusted p-values<0.05. LS: lifespan. Mort: mortality. (G) Correlation of 25% mortality with developmental time, body size, and the time to the 1^st egg, respectively. R: Pearson correlation coefficients. p: BH-adjusted p-values. (H) Scatter plots of time spent in 2 moving shapes and time to the 1^st progeny of each RIAIL strain. The p-values were calculated by Dirichlet regression. Related to Figure S1 and Table S1.

Longer lifespan is associated with slow development and late egg emergence

Early life-history traits can potentially provide insights into the developmental trajectory and long-term outcomes of organisms^22,23. To obtain an overview of associations between the phenotypic traits and lifespan traits, we pooled phenotypes and performed pairwise correlation analysis between all traits (Figure 2F). Worm developmental time (R = 0.37, BH-adjusted p < 0.001), reflecting the worm growth dynamics, and egg emergence (time to the moment when a worm lays its first egg, R = 0.42, BH-adjusted p < 0.001), reflecting sexual maturity, were most associated with 25% mortality (Figures 2F-2G). Body size, egg emergence, and progeny emergence (time until the first egg or progeny of a worm is detected) were strongly correlated with developmental time (R = 0.52, BH-adjusted p < 0.001, R = 0.52, BH-adjusted p < 0.001, and R = 0.58, BH-adjusted p < 0.001 respectively) (Figure 2F). In addition to the phenotypic readouts on worm development and reproduction, we also evaluated the most common shapes of worms in each population (Figures 2E and S1E). Four main categories of shapes were defined: two regular wild-type shapes in liquid (shape 2 - active; shape 3 - swimming) and two extreme shapes (shape 1 - straight; shape 4 - supercoiled) (Figure 2E). Since worms adopt different shapes over time, the shape metric reflects the percentage of time worms spent in each shape category. We observed a negative association between shape 1 (straight) and the duration before the first progeny appeared (Figure 2H). Conversely, there is a positive correlation between shape 2 (active movement) and the time it took for the first progeny to emerge (Figure 2H). However, none of the shapes directly correlated with lifespan traits (Figures 2F and S1F). Next, we determined the heritability (H²) for these phenotypic traits. This metric quantifies the genetic influence on the variability observed in lifespan and life-history traits (Figure S1G). Traits with high heritability increase the likelihood of identifying the contributing genomic regions, known as quantitative trait loci (QTL). We found heritability ranging from ~20% (e.g. shape 2 and time to 1^st progeny) to 60% (shape 4) (Figure S1G). The variable levels of heritability highlight the complexity of the genetic architecture underlying these traits. Both lifespan and life-history traits have moderate to high heritability, increasing the chance to identify significant QTL. In combination, these findings corroborate that delayed development and late egg emergence are an evolutionary cost of longevity.

Early life transcriptome unveils potential pathways influencing lifespan traits

To explore connections between the transcriptome at the early life (L4) stage and longevity, we performed an association analysis between the expression levels of ~11,000 transcripts and three measures of time to mortality: 25%, 50%, and 75% (Figures 3A and S2A-S2B). However, none of the individual transcripts showed a correlation with lifespan traits after multiple test corrections (Figures S2A-S2B, and Table S2). Therefore, we explored the possibility of disparities at the pathway level using gene set enrichment analysis (GSEA) to determine the collective impact of the transcripts (Figures 3B, S2D, and Table S2). We found 938 pathways significantly enriched for 25% mortality (81 positive and 857 negative), 701 (105 positive and 596 negative) for those associated with 50% mortality, and 58 (32 positive and 26 negative) for 75% mortality respectively (Figure S2D and Table S2). Given the early life transcriptome was more strongly associated with 25% mortality than other lifespan traits (Figure S2A), we investigated the enriched pathways associated with this metric. The majority of 25% mortality-enriched biological processes were negatively associated with lifespan and were primarily involved in chromosome organization, cytoskeleton organization, cellular lipid metabolism, cell division, DNA repair, and protein metabolic processes (Figure 3B). Among the top 30 enriched pathways, two pathways were positively associated with 25% mortality, namely neuropeptide signaling and G protein-coupled receptor signaling. Although it was not among the top 30, the geneset “determination of adult lifespan” was among those significantly inversely associated with early mortality (adj. p-value<0.05) (Figures 3C and S2C, Table S2). Taken together, the early life transcriptome showed significant associations with lifespan at the pathway level, particularly with 25% mortality. Our data indicate that various biological processes and pathways at the transcriptional level can influence early mortality and potentially affect longevity.

Figure 3. Quantitative assessment of associations between the transcriptome/proteome and lifespan.

(A) Schematic pipeline of mRNA-lifespan association mapping and gene set enrichment analysis (GSEA). (B) Graph representing the top 30 biological process gene sets enriched. GSEA of mRNA-25% mortality associations. Genes were ranked by the signed logarithm of the odds (LOD) score of mRNA-25% mortality association. Color represents positive (red) or negative (blue) normalized enrichment score (NES). All gene sets in the figure had a significant q-value. Gene sets in bold: overlapping gene sets at both mRNA and protein levels. (C) The running GSEA plot of “determination of adult lifespan” (GO:0008340). adj. p: adjusted p-value. (D) The schematic pipeline of protein-lifespan association mapping and GSEA. (E) Graph representing the top 30 biological process gene sets enriched. Gene set enrichment analysis of protein-25% mortality associations. Genes of proteins were ranked by the signed logarithm of the odds (LOD) score of protein-25% mortality association. Color represents positive (red) or negative (blue) normalized enrichment score (NES). All gene sets in the figure had a significant q-value. Gene sets in bold: overlapping gene sets at both mRNA and protein levels. Related to Figure S2 and Tables S2 - S3.

Quantitative assessment of correlations between protein pathways and different lifespan traits

As proteome analysis offers a more direct perspective on cellular function, complementing the information obtained through transcriptome analysis, we measured the protein profiles of RIAILs and detected >6,500 proteins following the removal of non-detectable peptides and rigorous quality control measures (Figure 3D and Table S3). Similar to the transcript level, none of the individual proteins showed an association with lifespan traits after multiple test corrections (Figures S2E-S2F). When investigating the top pathways enriched for 25% mortality, seven pathways were identified, such as vesicle-mediated transport, Golgi vesicle transport, actomyosin structure organization, and supramolecular fiber organization, that positively correlated with the 25% mortality (Figure 3E and Table S3). The negatively associated pathways were mostly related to DNA metabolism and cell cycle regulation (Figure 3E and Table S3). In addition, we examined the pathways enriched at both the mRNA and protein levels (Figures 3B and 3E). Gene sets involved in DNA damage response, DNA repair, and cell cycles overlapped at both mRNA and protein levels and were negatively correlated with lifespan traits. Consistent with these results, cell cycle, and associated genome integrity pathways were reported to be negatively associated with cellular turnover, a measure of cell and tissue longevity²⁴, whereas enhanced DNA repair capacity has been suggested in long-lived species^25,26.

Correlation of lipid profiles and lifespan traits

Perturbations in circulating lipid levels due to genetic, lifestyle, and environmental factors can heighten the risk of developing age-related disorders, such as cardiovascular and metabolic diseases²⁷. We hence integrated lifespan traits with lipid profiles measured in the RIAIL cohort (Figure 4A). The first two dimensions of a principal component analysis (18.8% and 11.3% of variance explained, respectively) did not visually segregate strains by lifespan (Figure 4B). Individual lipids did not exhibit significant lipid-lifespan correlations after applying multiple test corrections. We then explored distinct correlation profiles between lipids and different lifespan metrics (Figures 4C and 4D). Cardiolipins (CLs), comprised mainly of polyunsaturated acyl chains, were among the lipids that displayed a positive correlation with average lifespan and 75% mortality (Figure 4C and Table S4). The levels of CLs consistently decline in aged worms and rats^28,29, supporting the concept that higher levels of CLs may be advantageous for health and longevity. Phosphatidylinositols (PIs) were among the primary lipid classes positively associated with average lifespan and 75% mortality, while many triglycerides (TGs) correlated positively with 25% mortality (Figure 4C). In contrast, numerous phosphatidylethanolamines (PEs) and PE-derivatives (e.g. plasmanyl-PE and plasmenyl-PE) exhibited a negative correlation with all the lifespan traits (Figure 4D and Table S4). Over-representation analysis of the different lipid classes confirmed that TGs, CLs, and PIs were positively associated with lifespan traits, while PEs were negatively associated (adjusted p-value < 0.05) (Figure S3A).

Figure 4. Quantitative assessment of lipidome-lifespan associations.

(A) Diagram of lipid-lifespan association mapping and lipid-class over-representation. (B) Principal component analysis (PCA) representation of RIAIL strains based on all LC-MS measured lipids at L4/young adulthood. Color represents z-score of the average lifespan of each RIAIL strain. (C-D) Bar plots of the number of lipids with (non-adjusted p-value < 0.05) positive (C) or negative (D) association coefficient. CL: cardiolipin. PE: phosphatidylethanolamine. PE-derivatives: including Lyso-PE, plasmanyl-PE, plasmenyl-PE. PI: phosphatidylinositol. TG: triglycerides. PC: phosphatidylcholines. PC-derivatives: PC[OH] and plasmanyl-PC. Cer[AS]: ceramideAS. Cer[NS]: ceramideNS. Cer[AP]: ceramideAP. SP: Sphingolipid. Methyl-PA: methylphosphatidic acid. Related to Figure S3 and Table S4.

Lifespan variations of RIAILs are associated with the proportion of the N2 genome and independent of the mitochondrial haplotype

As the parental strain CB4856 worms have a shorter lifespan compared to the other parental strain N2 (Figures 2A-2B)¹⁴, we asked whether the allelic proportion of each parental genome in the RIAILs could partly explain the variations observed in the lifespan traits. Using Cox proportional hazards modeling, we found that the proportion of the N2 genome in the strain was significantly associated with a longer lifespan (HR = 0.46, p-value < 2e-16) (Figures 5A-5B, and S3B). The lifespan of CB4856 was previously found to be influenced by variants in the mitochondrial DNA (mtDNA)³⁰. We therefore separated the RIAILs by mitochondrial genotype but found no associations between the mitotype and the average lifespan (Figure S3C). A previous study using a different, small set of RIAILs found a positive correlation between the CB4856 mitotype and lifespan, as well as a negative correlation between N2 mitotype and lifespan³¹. However, we found no such correlation for either the CB4856 or N2 mitotype in these RIAILs (Figure S3D). Taken together, these data underscore the necessity for a more refined approach to pinpoint specific loci that determine lifespan.

Figure 5. Identification of a lifespan-modulating locus on Chromosome II and prioritization of candidate genes.

(A) Stacked barplot of Hawaii allele content (HAC) of CB4856 and N2 of 85 RIAIL strains and two wild-type parental strains. (B) Survival curve showing the effect of the N2 genome proportion in the strain on lifespan. Higher N2 genome proportion is associated with a longer lifespan. HR: hazard ratio; HAC (N2): proportion of N2 genome in RIAILs. (C) QTL mapping of lifespan and life-history traits identifies two significant QTL on Chr. II, and V for average lifespan, and time to 1^st progeny, respectively. The vertical axis shows the logarithm of the odds (LOD score). The horizontal axis represents the genomic position in mega-basepair (Mbp). Dashed and solid grey lines represent suggestive (p < 0.1) and significant (p < 0.05) thresholds, respectively. (D) Boxplot of the average lifespan of RIAILs with CB4856 or N2 genotype at position 13,121,591 on Chr. II. (E) Boxplot of time to 1^st progeny of RIAIL strains with CB4856 or N2 genotype at position 12,125,475 on Chr. V. The p-value represents the comparison of the two groups calculated using a two-tailed Student’s t-test. (F) Candidate genes are prioritized under the confidence region of the loci on Chr. II. Genes under the lifespan QTL peak were annotated if: a gene has one or more variants in CB4856; a gene has one or more missense or modifier variants in CB4856; the gene has been annotated in GenAge (https://genomics.senescence.info/genes/index.html) for involvement in the aging; the gene/protein showed a correlation with any lifespan traits; the gene/protein has an expression cis-eQTL or a protein cis-pQTL; the RNAi clone with the right sequence is available from at least one of the libraries (Ahringer and Vidal libraries). Grey color: not available (mRNA or protein not measured). A known longevity gene, rict-1, is highlighted in red. Related to Figures S3 - S5, and Table S5.

Identification of a lifespan QTL on Chromosome II

Next, we sought to leverage the genetic diversity of the RIAIL population to map their associations with lifespan traits and potentially uncover novel genetic regulators of lifespan. Through variant calling using RNA-seq data (Figures S3E-S3G), we generated a genetic map with 5,198 genetic markers for the RIAIL population (Figure S4). We then performed QTL mapping of lifespan traits and detected a significant QTL on Chr. II for average lifespan (Figures 5C and S5A-S5C). We detected a suggestive QTL in the same locus for the other mortalities (Figures 5C and Table S5). We observed a decrease in the LOD score for this locus, from 4.685 for average lifespan to 3.996 for 50% mortality, and finally to 3.177 for 75% mortality (Table S5). Upon examining the average lifespan of the RIAILs for the two genotypes at this locus, we found that strains with the CB4856 genotype have longer lifespans compared to strains carrying the N2 genotype (Figure 5D). In other words, the allele associated with a longer lifespan comes from the shorter-lived CB4856 strain, suggesting that complex gene-gene interactions overcome any single-locus effect on lifespan in the RIAILs.

We further explored other life-history traits and detected four significant QTLs: one for progeny emergence (the time when the 1^st progeny is observed) on Chr. V:12,125,475, one for egg emergence (the time when the 1^st egg is observed, indicating sexual maturity) on Chr. V:20,279,818, and two for shapes straight and supercoil, both on Chr. X (X: 11,549,662 and X:12,745,016 respectively) (Figures 5C, S5A-S5B, S5D, and Table S5). When examining the locus at V:12,125,475, we found that RIAIL strains with the N2 genotype exhibited a longer time for progeny emergence compared to those with the CB4856 genotype (Figure 5E). Furthermore, despite observing a correlation between developmental time and lifespan traits (Figures 2F-2G), we did not detect significant or shared QTL between the two traits, suggesting that this correlation does not necessarily imply a common genetic regulation.

Exploration of lifespan QTL identified eight candidate genes

The lifespan QTL encompassed eight genes. We first assessed their expressions across the RIAILs and observed no difference when categorizing them by the N2/CB4856 markers (Figure S5E). To prioritize the most likely candidate modulators of lifespan, we considered a wide range of factors (Figure 5F), namely whether there were any genetic variants in the gene between N2 and CB4856, whether any of these were mis-/nonsense mutations, the presence of cis-e/pQTLs defined as genomic loci near the gene of interest (in cis) that explain the variation in expression levels of mRNA (eQTL) or protein (pQTL) in that gene, prior knowledge of the gene being associated with aging (in GenAge, a curated database of genes associated with age-related processes³², whether the gene was correlated with lifespan at the mRNA or protein level, and whether an RNAi clone is available in either Ahringer or Vidal library. Most of the genes had some genetic variants, many with missense or nonsense variants as well, however, we did not assess whether these variants represent a partial gain-of-function or a loss-of-function for the protein. But among these, only rict-1 met four criteria as it has been reported to be associated with both aging and lifespan. rict-1 encodes a key component of the mTORC2 complex and loss-of-function mutations have previously been shown to increase the lifespan of C. elegans in specific conditions^33,34. In addition, we were interested in gfm-1, as mitochondria play a key role in longevity regulation and gfm-1 is a known mitochondrial gene, encoding the G elongation factor mitochondrial 1.

RNAi of gfm-1 prolonged lifespan by activating the UPR^mt

To examine the causal relationship between the candidate genes unveiled hitherto and longevity modulation, we knocked down these candidate genes by feeding worms with RNAi bacteria targeting each candidate gene, starting from the maternal phase, and measured their lifespans (Figure 6A). We managed to acquire RNAi clones for seven of the eight candidate genes. RNAi clones for the candidate gene F29C12.6 were unavailable in either the Ahringer or Vidal RNAi libraries. Of note, the knockdown of gfm-1 showed the most robust lifespan extension compared to the other candidate genes (p-value < 0.0001) (Figure 6A). Besides gfm-1, our survival analysis revealed that knocking down of mltn-1 (molting cycle MLT-10-like protein³⁵) also affected lifespan (p-value < 0.001), albeit to a lower extent. Consistent with prior findings, our data also confirmed that worms fed with rict-1 RNAi exhibited an extended lifespan (Figure 6A)^33,36. Additionally, we confirmed that the observed lifespan extension in the candidate genes occurs independently of 5FU supplementation (Figure S6A).

Figure 6. RNAi of gfm-1 induced UPR^mt activation and prolonged lifespan in C. elegans.

(A) Lifespan of worms fed with ev (empty vector, the control RNAi) (black) or candidate gene RNAi (red). Except for rict-1 RNAi, all the lifespan measurements were performed at 20°C and the RNAi exposure was started from the maternal L4 stage (n = 80 worms/condition). For rict-1 RNAi, worms were fed with rict-1 RNAi from day 1 adulthood and cultured at 25°C. P-values represent a comparison with the controls calculated using the log-rank test. n.s.: not significant. (B) RNAi of gfm-1 extends worm lifespan in an RNAi dose-dependent manner. Worms fed with ev (control RNAi) or 10%-100% gfm-1 RNAi; control RNAi was used to supply to a final 100% of RNAi. (C) Age-related paralysis of worms fed with ev or 10%-100% gfm-1 RNAi (10-12 worms/plate, 8 plates/condition). Error bars denote SEM. Statistical analysis was performed by one-way ANOVA followed by Tukey post-hoc test (*p < 0.05; ***p < 0.001). (D) RNAi of gfm-1 induced the UPR^mt (hsp-6p::gfp reporter) in a dose-dependent manner. Worms fed with ev or 10%-100% gfm-1 RNAi. Scale bar: 0.5 μm. (E) mRNA levels (n = 4 biological replicates) in worms fed with ev, or 10%-100% gfm-1 RNAi. Statistical analysis was performed by one-way ANOVA followed by Tukey post-hoc test (**p < 0.01; ***p < 0.001). Values in the figure are mean ± SEM. (F) RNAi of gfm-1 reduced both basal and max. oxygen consumption rate (OCR) compared to those of controls on day 1 of adulthood (10 worms/well, 10 wells/condition). Values in the figure represent mean ± SEM. Statistical analysis of RT-qPCR results was performed by one-way ANOVA followed by Tukey post-hoc test (*p < 0.05; **p < 0.01; ***p < 0.001). Related to Figure S6.

As several lipid classes showed a moderate association with lifespan traits (Figure 4), we asked whether candidate genes within the lifespan QTL also affect major lipid storage. Using Oil Red O staining (ORO) to detect lipids, we observed increased lipid accumulation in worms with gfm-1 knockdown, whereas lipid levels remained unchanged in worms fed with RNAi against the other candidate genes (Figures S6B-S6E).

To further characterize the mechanism of gfm-1 RNAi-mediated longevity, we conducted several functional assays. We assessed the effect of gfm-1 RNAi on lifespan and healthspan with a dilution of RNAi bacteria, including 10%, 25%, 50%, 75%, and 100% (control RNAi was used to supply to a final 100% of RNAi for all conditions). Worms exposed to different amounts of gfm-1 RNAi showed a dose-dependent lifespan extension (Figure 6B) and reduction of age-related paralysis (Figure 6C). Because gfm-1 encodes a mitochondrial translation elongation factor, we considered whether the mitochondrial stress response (MSR), through components such as the mitochondrial unfolded protein response (UPR^mt), was involved in longevity changes observed with gfm-1 knockdown. Indeed, gfm-1 RNAi robustly increased the GFP expression of hsp-6p::gfp worms and upregulated the expression of the UPR^mt genes, including atfs-1, hsp-6, and gpd-2 (Figures 6D-6E). In line with this, mitochondrial respiration was also reduced upon gfm-1 knockdown in a dose-dependent manner (Figure 6F). Since gfm-1 RNAi extends worm lifespan by activating the UPR^mt, we did not expect an increase in lipid accumulation, because the UPR^mt activation typically reduces lipid levels, as seen with its other inducers like cco-1 RNAi, mrps-5 RNAi, and doxycycline, which generally reduce triglycerides levels^14,37. Interestingly, lipid biosynthesis has been linked to the mitochondrial-cytosolic stress response (MCSR), which is induced by hsp-6 RNAi and requires the activation of both UPR^mt and cytosolic stress responses³⁸. Therefore, we investigated whether gfm-1 RNAi might also trigger a cytosolic stress response (Figure S6F). Of note, gfm-1 RNAi did not activate cytosolic stress response, suggesting that lipid accumulation observed in gfm-1 RNAi is independent of the MCSR and implies a novel mechanism that requires further investigation. These results confirmed the beneficial effect of mitochondrial inhibition and UPR^mt activation on healthy aging and longevity^39,40.

Furthermore, we investigated the potential mechanism of mltn-1 RNAi-induced longevity by examining whether any of the established longevity pathways contribute to the observed lifespan extension (Figures S6G-S6K). We fed mltn-1 RNAi to worms with mutations mimicking caloric restriction (eat-2 mutant and sir-2.1 overexpression worms)^41,42, insulin/IGF-1 signaling (daf-2 mutant)⁴³, AMPK signaling (aak-2 mutant)⁴⁴ and oxidative stress response (skn-1 mutant)⁴⁵ (Figures S6G-S6K). Of note, mltn-1 RNAi prolonged the lifespan of worms overexpressing sir-2.1 overexpression and skn-1 mutants, indicating that mltn-1 RNAi regulates longevity independent of sirtuin-induced caloric restriction and oxidative stress response (Figures S6H and S6K). However, mltn-1 RNAi did not further extend the lifespan of eat-2 and daf-2 mutants (Figures S6H and S6I), and the lifespan extension induced by mltn-1 RNAi was almost completely abolished in aak-2 mutant worms (Figure S6J). These results suggest that the knockdown of mltn-1 may extend worm lifespan in an AMPK-dependent manner. These findings further reinforced the assertion that our approach enabled us to identify novel inducers of longevity.

Variants in human GFM1 are associated with an increased risk of heart failure

Age-related diseases play a significant role in shaping longevity⁴⁶. To explore the human relevance of our newly identified longevity genes, we took advantage of the UK Biobank, a large-scale population-based cohort study with extensive health and medical information⁴⁷. While mltn-1 is a C. elegans-specific gene, we identified GFM1 and RICTOR as the human orthologs of worm gfm-1 and rict-1, and then used Cox proportional-hazards models^48-50 to investigate whether variants within these genes were associated with disease risk (Figure 7A). We explored the association between single nucleotide polymorphisms (SNPs) in GFM1 and RICTOR (selection based on criteria outlined in the STAR method) and the lifelong incidence of 7 age-associated diseases (with >10,000 events) as well as all-cause mortality (referred to as “Death”).

Figure 7. Exploration of association of SNPs in human GFM1 and RICTOR with an incidence of age-associated disease and death in the UK Biobank.

(A) The workflow of disease risk associations between SNPs in GFM1/RICTOR and life-long incidence of diseases and all-cause mortality with Cox proportional-hazard models. (B) Manhattan plot showing the negative log-transformed p-values for an association between SNPs in GFM1 and all outcomes (seven diseases plus death). The dashed line indicates the approximate p-value corresponding to an FDR of 0.05. Significant associations are shown as triangles while color indicates the outcome tested. (C) Forest plot of whole-gene association of GFM1 and RICTOR with each outcome from the burden-type Cox model. The dashed line indicates the approximate p-value corresponding to an FDR of 0.05. Related to Figure S7, Table S6.

While properly calibrated for common alleles, the Cox model had a high rate of type 1 error for rare variants (as assessed by permutation). We used a null distribution from permutation to correct for this effect in rare alleles (allele count ≤ 30). A set of highly correlated (cor > 0.99), common (MAF 47-48%) SNPs in GFM1 showed a significant association with the risk of heart failure (FDR = 3.9*10⁻³), with a hazard ratio of ~1.06 for each copy of the major allele (Figure 7B and Table S6). A single rare variant in each of GFM1 and RICTOR showed a possible association with acute renal failure and death, respectively; however, these would likely be found non-significant given sufficient permutations (Figures 7B and S7A). The absence of multiple signals suggesting the same gene-disease association led us to not pursue these further.

In addition to single variant analysis and boosting power for rare variants, we also performed Cox modeling using genetic burden as a predictor (Figure 7C and Table S6). The burden was calculated as a function of the variant-predicted effect and the frequency of the allele, with higher weights attributed to variants with deleterious (predicted) effects and rarer alleles. Consistent with the single variant analysis, we found that the genetic burden of GFM1 was associated with an increased risk of heart failure, with the hazard ratio between the bottom and top declines estimated at 1.068 (95% CI: 1.031-1.106). This reinforces the hypothesis that GFM1 affects the risk of heart disease through both common and rare genetic variants. For completeness, we also analyzed the human orthologs of the other candidate genes identified, namely KLHL28 and COL5A3 (orthologs of worm bath-45 and col-86, respectively), however these did not yield significant associations (Figures S7B-S7D, Table S6). These results suggest that GFM1 may affect the risk of heart failure in humans, thereby negatively impacting lifespan.

Discussion

Here we present a multi-omics atlas of the worm RIAILs, as a resource to understand the regulation of longevity. The observed difference in average lifespan between the parental strains was consistent with previous studies^14,51. The RIAIL strains exhibited extensive lifespan variation with some strains exceeding that of the parentals suggesting the presence of transgressive segregation⁵². Research across species has revealed consistent trade-offs that influence lifespan and life-history traits, with correlations observed between key phenotypic traits such as lifespan and fecundity¹⁸, development time and fecundity⁵³, development time and lifespan¹⁹, as well as body size and longevity^22,54. However, research exploring the correlations between longevity and early life history traits in wild C. elegans populations is relatively scarce^51,55. We therefore tracked various life-history phenotypes during the early life stage of the RIAILs, gathering data on developmental progression, reproductive capability, fertility, and behavioral activity. The measurements of life-history traits were conducted in liquid culture using a microfluidics device, while the lifespan measurements were performed on plates. Therefore, we cannot rule out the possibility that the different culture conditions may have influenced the observed traits. Only developmental time and egg emergence – both reflecting sexual maturity – had a weak to moderate correlation with lifespan, which is in line with the absence of any overlapping QTL between early life history and lifespan traits. These data are corroborated by a prior study in worms, which proposed that development, reproduction, and lifespan are under independent genetic regulation⁵⁶, and work in D. melanogaster, where a disconnect between life-history traits and lifespan was observed when examining variations in larval food conditions^57,58. In a similar vein, fly selection experiments have yielded inconsistent results in terms of discovering genetic correlations between development time, body size, and longevity^59-62.

Subsequently, we investigated whether multi-omic molecular characteristics, such as gene expression, protein, or lipid abundance, could be linked to lifespan in the RIAILs. We did not detect any significant correlations between individual transcript/protein/lipid and lifespan traits following multiple testing corrections possibly due to small marginal effects or more complex gene interactions. These correlations, however, strengthened and reached statistical significance when we performed GSEA on genes ranked by the mRNA-lifespan associations, supporting the presence of several biological pathways that are potentially involved in the modulation of RIAIL lifespans. The lack of associations at the individual transcript level may not negate the possibility of a functional impact and physiological relevance at the pathway level, where complex interactions and synergistic effects may come into play. For instance, neuropeptide signaling and G protein-coupled receptor (GPCR) signaling were particularly notable among the pathways that were positively associated with lifespan traits. This finding aligns with prior studies, where one demonstrated the role of the neuropeptide signaling pathway in extending C. elegans lifespan^63,64, and another highlighted the influence of the GPCR pathway on longevity across humans and various animal models including worms⁶⁵. While we identified intriguing gene sets associated with lifespan traits at the transcript level, these associations were not replicated in the analysis between protein expression and lifespan traits. The discrepancy between the transcriptomic and proteomic levels could be attributed to several factors, such as post-transcriptional regulation, protein turnover, limitations in the proteomic detection methods⁶⁶, or the differential effect of natural variation on the proteome⁶⁷. These discrepancies emphasize the importance of considering multiple omics layers to obtain a comprehensive understanding of biological processes and their role in determining phenotypic outcomes.

Alterations in circulating lipid concentrations, triggered by genetic influences, lifestyle choices, and environmental conditions, can escalate the risk of age-associated disorders²⁷. We therefore also collected full lipidomic profiles of RIAILs and investigated whether complex lipids might also be associated with specific lifespan traits. We found that TGs, CLs, and PIs were over-represented in positive lipid-lifespan associations, while PEs were enriched in negative lipid-lifespan associations. It is notable that CLs, comprised mainly of polyunsaturated fatty acid chains, were found to be among those positively associated with lifespan traits. CLs are mitochondria-specific phospholipids essential for preserving mitochondrial integrity⁶⁸. Due to their special cellular confinement, CLs are closely related to the maintenance of mitochondrial function, which connects CLs to longevity and the progression of age-related disease⁶⁹. This aligns with our findings that indicate a positive correlation between CLs and various lifespan traits. In contrast, the level of PEs consisting of less saturated fatty acids exhibited a negative correlation with lifespan traits. Although PEs are the second most prevalent glycerophospholipid in eukaryotic cells and positively regulate autophagy and lifespan in yeast and mammalian cells^70,71, decreased levels of PE were associated with lower beta-amyloid accumulation in both mammalian cells and flies⁷², suggesting a complex role of PEs in regulating age-related effects and longevity.

In addition to the multi-omic characterization of the RIAIL population, we performed QTL mapping and identified candidate lifespan loci on Chr. II, with strains carrying the CB4856 genotype showing longer lifespans compared to those with the N2 genotype at this locus. This finding was particularly interesting considering that the N2 parental strain displayed a longer lifespan compared to the CB4856 strain and the proportion of the N2 genome being predictive of longer survival suggests considerable pleiotropy. RNAi against the seven candidate genes in that locus found that knockdown of gfm-1 and mltn-1 led to lifespan extensions. Our analyses suggest that the dose-dependent lifespan extension and reduction of age-related paralysis through gfm-1 inhibition could be mediated by the modulation of the mitochondrial stress response. Although we did not detect an mRNA or protein QTL for gfm-1 within the same lifespan locus, the experimental findings were in line with this gene encoding the G elongation factor mitochondrial 1 and the upregulation of gene sets associated with mitochondrial ribosomes at this locus in the RIAILs population. Given that mltn-1 is specific to C. elegans, its translational relevance to human studies is uncertain.

To evaluate the potential clinical relevance of the selected candidate genes that are conserved in humans, we took advantage of the UK biobank and demonstrated that both common and rare variants in the human GFM1 were associated with the risk of heart failure, which impairs overall health and contributes to shorter life expectancy. Homozygous loss of GFM1 function is known to cause combined oxidative phosphorylation deficiency 1 (COXPD1), a multisystem disorder with pleiotropic symptoms⁷³. Among these symptoms, reduced oxidative phosphorylation of the mitochondrial OXPHOS in muscle tissue has been reported, which may explain the involvement of GFM1 in heart failure.

In summary, our study unveiled a specific genetic locus that plays a role in determining lifespan variation within the RIAIL population. Furthermore, we identified known and novel longevity modulators, including rict-1, gfm-1, and mltn-1, which we validated experimentally. For GFM1 an association with heart failure was established through human genetic analysis, underscoring that worm data can be generalizable across species. The comprehensive multi-layered characterization of the RIAIL population is now also made accessible through an open-access web resource (https://systems-genetics.org/cel_longevity), which provides a valuable tool for investigating the intricate relationships between biochemical and whole-body phenotypes and for hypothesis generation for the scientific community.

Limitations of the study

We note several limitations and future directions of our work. The relatively low sample size of worms (60 worms/strain) used for lifespan analysis restricts our ability to get an accurate estimate of late-life mortality, especially for the maximal lifespan of the strain⁷⁴. This likely undermines our statistical power in evaluating the associations between traits and late-lifespan phenotypes. The life-history trait screening was done in liquid culture using the microfluidics device, while the lifespan assays were performed on plates; we can hence not exclude a possible influence of different culture conditions on traits. The gathered molecular characteristics encompass aggregated data at the strain level and are limited to a single early time point. However, expanding the data collection to include later time points would enable the exploration of age-related dynamics associated with these traits. The challenge of false positives is inherent in high-throughput multi-omics studies due to the large number of tests performed. Additionally, replication of findings in independent cohorts and cross-validation with different and more comprehensive analytical approaches will be crucial steps in confirming the robustness of the observed associations. Another limitation of our study is that we only obtained genetic variants within coding regions using a variant calling pipeline based on transcriptome data. This approach may have contributed to the lack of a causal link between genetic variants within the QTL peak and the expression of genes within the locus. Employing whole-genome sequencing to refine the genotype map could help identify additional QTLs associated with lifespan or provide insights through multi-omics datasets.

Finally, the experimental validations of gfm-1, rict-1, and mltn-1 were conducted using RNAi knockdown in the N2 Bristol background. Moving forward, an important avenue for further investigation would involve utilizing CRISPR technology to examine the specific variant of gfm-1 in the RIAILs population. Additionally, we did not investigate the potential role of F29C12.6 in lifespan regulation due to the absence of RNAi clones in the two available RNAi libraries. Further studies may be necessary, including the development of new RNAi clones, to better understand the impact of F29C12.6 on longevity. Furthermore, our study focused solely on the downregulation of the candidate genes, without exploring the effect of the overexpression of the candidate genes.

STAR METHODS

EXPERIMENTAL MODEL AND SUBJECT PARTICIPANT DETAILS

Bacterial strains and C. elegans strains

The Bristol strain (N2) and Hawaii strain (CB4856) were used as the wild-type strains, and SJ4100 [zcIs13(hsp-6p::GFP)], CL2070 [dvIs(hsp-16.2p::GFP)], DA465 [eat-2(ad465)], VC199 [sir-2.1(ok434)], CB1370 [daf-2(e1370)], RB754 [aak-2(ok524)], and GR2245 [skn-1(mg570)] were obtained from the Caenorhabditis Genetics Center (CGC; Minneapolis, MN). E. coli OP50 and HT115 strains were also obtained from the CGC. RNAi clones against Y81G3A.4, col-86, rict-1, pqn-32, gfm-1, bath-45, and mltn-1, were obtained from the Ahringer and Vidal libraries and verified by sequencing before use (detailed in the Key Resource Table). Worms were cultured and maintained at 20°C and fed with E. coli OP50 on Nematode Growth Media (NGM) plates unless otherwise indicated.

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and virus strains
Escherichia coli: OP50	Caenorhabditis Genetics Center	RRID:WB-STRAIN:OP50
Escherichia coli: HT115 (DE3)	Caenorhabditis Genetics Center	RRID:WB-STRAIN:HT115(DE3)
Y81G3A.4 RNAi	Ahringer	II-8H23
col-86 RNAi	Vidal	11029-F11
rict-1 RNAi	Vidal	11038-B10
pqn-32 RNAi	Ahringer	II-8J03
gfm-1 RNAi	Vidal	10015-F6
bath-45 RNAi	Vidal	11032-F1
mltn-1 RNAi	Ahringer	II-10C18
Chemicals, peptides, and recombinant proteins
5-Fluorouracil (5-FU)	Sigma-Aldrich	Cat# F6627
Ampicillin sodium salt	Sigma-Aldrich	Cat# A9518
Carbenicillin disodium salt	Sigma-Aldrich	Cat# C1389
IPTG	AppliChem	Cat# A1008,0005
TriPure Isolation Reagent	Roche	Cat# 11667165001
Methanol (Optima™ LC/MS Grade)	Fisher Chemical	Cat# A454SK-4
Methyl tert-butyl ether (MTBE)	Sigma-Aldrich	Cat# 443808
18MΩ MilliQ water	Made in house
Stainless metal bead (5mm diameter)	Qiagen	Cat# 69989
Acetonitrile (Optima™ LC/MS Grade)	Fisher Chemical	Cat# A955-4
Urea	Sigma-Aldrich	Cat# U5378
Tris(2-carboxyethyl)phosphine	Sigma-Aldrich	Cat# C4706
2-chloroacetamide	Sigma-Aldrich	Cat# C4706
Formic acid, Pierce™	Thermo Scientific	Cat# PI28905
Trypsin	Promega	Cat# V5113
96 well desalting plates (10 mg/well, Strata™-X 33 μm Polymeric Reversed phase)	Phenomenex	Cat# 8E-S100-AGB
Ammonium acetate (LiChropur™)	Sigma-Aldrich	Cat# 73594
Acetic acid	Sigma-Aldrich	Cat# 695092
Ammonium hydroxide	Sigma-Aldrich	Cat# 338818
2-Propanol (Optima™ LC/MS grade)	Fisher Chemical	Cat# A461212
Acetic acid	Sigma-Aldrich	Cat# 695092
Critical commercial assays
NucleoSpin RNA, Mini kit for RNA purification	Macherey-Nagel	Cat# 740955.250
Seahorse Xfe96 Extracellular Flux Assay kit	Agilent	Cat# 102416-100
RNA using the Reverse Transcription Kit	Qiagen	Cat# 205314
LightCycler 480 SYBR Green I Master kit	Roche	Cat# 04887352001
Quantitative colorimetric peptide assay, Pierce™	Thermo Scientific	Cat# 23275
Oil Red O stain	Sigma-Aldrich	Cat# O0625
EGTA	Sigma-Aldrich	Cat# E8145
Spermidine	Sigma-Aldrich	Cat# S2626
Spermine	Sigma-Aldrich	Cat# S4264
PIPES	Sigma-Aldrich	Cat# P6757
2-mercaptoethanol	Sigma-Aldrich	Cat# M3148
16% paraformaldehyde	Alfa Asear	Cat# 43368
2-Propanol	Sigma-Aldrich	Cat# I9516
Triton X-100	Sigma-Aldrich	Cat# T8787
Deposited data
C. elegans RNA-seq data	This paper	The Shiny app; GSE252593
C. elegans proteomics data	This paper	The Shiny app; MSV000088622, MSV000089880
C. elegans lipidomics data	This paper	The Shiny app; MSV000088622, MSV000089880
Life history traits	This paper	Table S1
Transcript – lifespan associations	This paper	Table S2
GSEA of transcript – lifespan associations	This paper	Table S2
Protein – lifespan associations	This paper	Table S3
Lipid – lifespan associations	This paper	Table S4
QTL for lifespan and life history traits	This paper	Table S5
UKBB database analysis	This paper	Table S6
Summary of experimental and analytical samples	This paper	Table S7
Experimental models: Organisms/strains
C. elegans: N2 Bristol	Caenorhabditis Genetics Center (CGC); https://cbs.umn.edu/cgc/home	RRID:WB-STRAIN: N2_(ancestral)
C. elegans: CB4856 Hawaii	Caenorhabditis Genetics Center (CGC); https://cbs.umn.edu/cgc/home	RRID:WB-STRAIN:CB4856
C. elegans: 85 RIAILs, from QX1430xCB4856 RIAILs (QX240-QX598)	Andersen’s lab https://andersenlab.org/Research/Reagents/
C elegans: SJ4100 [zcIs13(hsp-6p::GFP)]	Caenorhabditis Genetics Center (CGC); https://cbs.umn.edu/cgc/home	RRID:WB-STRAIN:SJ4100
C.elegans:CL2070(dvIs[hsp-16.2p::GFP])	Caenorhabditis Genetics Center (CGC); https://cbs.umn.edu/cgc/home	RRID:WB-STRAIN:CL2070
Oligonucleotides
atfs-1 Fw: GAATAAGCCTCTATGATCCGATG	Sigma-Aldrich	N/A
atfs-1 Rv: GGTTGAAGCTGGGAAAGTGA	Sigma-Aldrich	N/A
hsp-6 Fw: AGAGCCAAGTTCGAGCAGAT	Sigma-Aldrich	N/A
hsp-6 Rv: TCTTGAACAGTGGCTTGCAC	Sigma-Aldrich	N/A
gpd-2 Fw: AAGGCCAACGCTCACTTG AA	Sigma-Aldrich	N/A
gpd-2 Rv: GGTTGACTCCGACGACGA AC	Sigma-Aldrich	N/A
pmp-3 Fw: GTTCCCGTGTTCATCACTCAT	Sigma-Aldrich	N/A
pmp-3 Rv: ACACCGTCGAGAAGCTGTAGA	Sigma-Aldrich	N/A
Software and algorithms
GraphPad Prism v8	GraphPad Software, Inc.	https://www.graphpad.com/scientificsoftware/prism/
Maxquant (version 2.0.3.1)	Max-Planck-Institute of Biochemistry	https://www.maxquant.org/
R (version 4.1.0)	The R Foundation	https://www.r-project.org/
Caenorhabditis elegans Natural Diversity Resource	Caenorhabditis elegans Natural Diversity Resource	https://elegansvariation.org Soft filtered variant file was retrieved from: http://storage.googleapis.com/elegansvariation.org/releases/20200815/variation/WI.20200815.soft-filter.vcf.gz. Hard filtered variant file was retrieved from: http://storage.googleapis.com/elegansvariation.org/releases/20200815/variation/WI.20200815.hard-filter.vcf.gz.
Adobe Illustrator 2023	Adobe	https://www.adobe.com/products/illustrator.html
survival	survival	https://cran.r-project.org/web/packages/survival/
FastQC	FastQC	https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
MultiQC	MultiQC	https://multiqc.info/
Genome Analysis Toolkit (GATK)	Genome Analysis Toolkit (GATK)	https://gatk.broadinstitute.org/
onemap	onemap	https://cran.r-project.org/web/packages/onemap/index.html
polycor	polycor	https://cran.r-project.org/web/packages/polycor/index.html
reshape2	reshape2	https://cran.r-project.org/web/packages/reshape2/index.html
DESeq2	DESeq2	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
edgeR	edgeR	https://bioconductor.org/packages/release/bioc/html/edgeR.html
qtl2	qtl2	https://cran.r-project.org/web/packages/qtl2/
limma	limma	https://bioconductor.org/packages/release/bioc/html/limma.html
FactoMineR	FactoMineR	https://cran.r-project.org/web/packages/FactoMineR/index.html
DirichletReg	DirichletReg	https://cran.r-project.org/web/packages/DirichletReg/index.html
plotly	plotly	https://cran.r-project.org/web/packages/plotly/index.html
cowplot	cowplot	https://cran.r-project.org/web/packages/cowplot/index.html
RColorBrewer	RColorBrewer	https://cran.r-project.org/web/packages/RColorBrewer/index.html
openxlsx	openxlsx	https://cran.r-project.org/web/packages/openxlsx/index.html
dplyr	dplyr	https://CRAN.R-project.org/package=dplyr
xlsx	xlsx	https://CRAN.R-project.org/package=xlsx
clusterProfiler	clusterProfiler	https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
enrichplot	enrichplot	https://bioconductor.org/packages/release/bioc/html/enrichplot.html
ggrepel	ggrepel	https://cran.r-project.org/web/packages/ggrepel/index.html
ggplot2	ggplot2	https://cran.r-project.org/web/packages/ggplot2/index.html
UpSetR	UpSetR	https://cran.r-project.org/web/packages/UpSetR/index.html
lme4	lme4	https://cran.r-project.org/web/packages/lme4/index.html
coxme	coxme	https://cran.r-project.org/web/packages/coxme/
lmerTest	lmerTest	https://cran.r-project.org/web/packages/lmerTest/index.html
GenomicFeatures	GenomicFeatures	https://bioconductor.org/packages/release/bioc/html/GenomicFeatures.html
org.Ce.eg.db	org.Ce.eg.db	https://bioconductor.org/packages/release/data/annotation/html/org.Ce.eg.db.html
BSgenome.Celegans.UCSC.ce11	BSgenome.Celegans.UCSC.ce11	https://bioconductor.org/packages/release/data/annotation/html/BSgenome.Celegans.UCSC.ce11.html
bslib	bslib	https://cran.r-project.org/web/packages/bslib/
shiny	shiny	https://cran.r-project.org/web/packages/shiny/
shinyWidgets	shinyWidgets	https://cran.r-project.org/web/packages/shinyWidgets/
shinydashboard	shinydashboard	https://cran.r-project.org/web/packages/shinydashboard/
shinydashboardPlus	shinydashboardPlus	https://cran.r-project.org/web/packages/shinydashboardPlus/
shinyalert	shinyalert	https://cran.r-project.org/web/packages/shinyalert/
shinybusy	shinybusy	https://cran.r-project.org/web/packages/shinybusy/
shinycssloaders	shinycssloaders	https://cran.r-project.org/web/packages/shinycssloaders/
biomaRt 2.58.0	biomaRt	https://bioconductor.org/packages/release/bioc/html/biomaRt.html
stringr 1.5.1	stringr	https://cran.r-project.org/package=stringr
purrr 1.0.2	purrr	https://cran.r-project.org/package=purrr
data.table 1.14.8	data.table	https://cran.r-project.org/package=data.table
lme4qtl	lme4qtl	https://github.com/variani/lme4qtl
htslib 1.14	samtools	https://www.htslib.org/
bcftools 1.14	samtools	https://www.htslib.org/
Ensembl VEP docker image (revision a9c9ffde86368006d348ec1b31163cc54 b17e187)	Ensembl	https://hub.docker.com/r/ensemblorg/ensembl-vep

METHOD DETAILS

Lifespan and paralysis measurements

In general, 5-10 L4 worms of each worm strain were transferred onto RNAi plates (containing 2 mM IPTG and 25 mg/mL carbenicillin) seeded with E. coli HT115 bacteria or RNAi bacteria⁷⁵. After the F1 progenies reached the last larval stage L4, worms were then transferred onto RNAi plates containing 10 μM 5FU, adhering to the standard concentration, to prevent egg hatching. Lifespan measurements of 85 RIAILs were performed in 10 batches of experiments (Batch 1-8: 8 RIAIL strains/batch; Batch 9: 10 RIAILs; Batch 10: 11 RIAILs. N2 and CB4856 worms were replicated in each batch), and 60 RIAIL worms were used and scored every other day. Individual RIAIL strains do not have biological replicates.

For the validation experiments of the candidate genes, 80 worms were used for each condition. Paralysis was manually assessed through the poking method⁷⁶, with a minimum of 80 worms analyzed per condition. All the validation experiments were performed at least twice. For lifespans with non-5FU supplemented conditions, worms were manually transferred daily to fresh RNAi plates during the first 12 days.

Phenotyping by microfluidics

The phenotypic readouts reflecting development, growth dynamics, fertility, and reproduction of different RIAIL strains were collected with the SydLab robotic microfluidic-based platform developed by Nagi Bioscience SA, which allows high-throughput and high-content C. elegans screenings. A synchronized population of C. elegans was injected into microfluidic chips at the L1 larval stage. Worms were confined within dedicated microfluidic chambers and were continuously fed with freeze-dried E. coli OP50 solution. The experiments were performed at 23°C.

Worm growth monitoring.

Pictures of each microfluidic chamber were taken at every hour, from the time of the first on-chip feeding (synchronized worms at the L1 stage) till the end of the experiment (t100h). Image sets were analyzed using software algorithms developed by Nagi Bioscience, allowing to extract data about the size (area) of each worm. From the model used to describe the growth, the software algorithms extracted two parameters: K, which is the maximum area the worm is aiming at, and r, which corresponds to the time when the worm reaches half of its maximal size (area) and determines the dynamic of the growth. This model is fitted to data with the Markov chain Monte Carlo (MCMC) method. For egg-laying worms, the progeny was computationally removed (selected through their size) so fits are not altered. For all the fits, a threshold was applied to Pearson’s chi-squared test, which measures the goodness of fit, to only keep the sensible ones.

Eggs and progeny emergence monitoring.

The eggs and progeny emergence, which correspond to the time point when the 1^st egg or the 1^st progeny appeared in each micro-chamber, was determined automatically by the software algorithm developed by Nagi Bioscience. For the egg emergence, the machine learning algorithm detected the exact step (in hours) at which the new object appeared for each micro-chamber. For the progeny emergence, a two-step processing was applied. First, the initial number of worms injected into each micro-chamber was computed by calculating the average number of worms counted over the first 50 steps of the experiments. Then, the algorithm calculated a rolling average of the number of worms in each micro-chamber on a 10 steps window: if a threshold of 1.5 is passed at step (x), it meant that a new entity (i.e., a new progeny) emerged at step (x).

Eggs and progeny accumulation monitoring.

The mean number of eggs and progeny produced by the worms’ population was automatically computed by the software algorithm developed by Nagi Bioscience, by detecting and counting the number of new eggs and new progeny produced at each step. The model used to describe the evolution of the eggs and progeny accumulation was a linear model. The details on the number of replicates of each RIAIL stain are provided in Table S1.

Sample collection for RNA-seq, proteomics, and lipidomics analyses

Worms of each RIAIL strain were cultured on plates seeded with E. coli OP50, and then worm eggs were obtained by alkaline hypochlorite treatment of gravid adults. A synchronized L1 population was obtained by culturing the egg suspension in sterile M9 butter overnight at room temperature. Approximately 2000 L1 worms of each RIAIL strain were transferred onto plates seeded with E. coli HT115. L4 worms were harvested after 2.5 days with M9 buffer and washed three times. Worm pellets were immediately submerged in liquid nitrogen for snap-freezing and stored at −80°C until use. 2 tubes of worm pellets were collected for each RIAIL stain, 1 pellet was used for RNA extraction to perform RNA sequencing, the other pellet was used for protein and lipid extractions.

RNA extraction and RNA-seq data analysis

On the day of the RNA extraction, 1 mL of TriPure Isolation Reagent was added to each sample tube. The samples were then frozen and thawed quickly eight times with liquid nitrogen and a 37 °C water bath to rupture worm cell membranes. Total worm RNA was extracted by using a column-based kit from Macherey-Nagel. RNA-seq was performed by BGI with the BGISEQ-500 platform. FastQC (version 0.11.9) was used to verify the quality of the reads⁷⁷. Low-quality reads were removed and no trimming was needed. Alignment was performed against worm genome (WBcel235 sm-toplevel) following the STAR (version 2.73a) manual guidelines⁷⁸. The STAR gene-counts for each alignment were analyzed for differentially expressed genes using the R package DESeq2 (version1.32.0)⁷⁹ using a generalized-linear model. Count data were normalized to counts per million (CPM) using edgeR (version 3.36.0) for visualizations of expression data. Biological process (BP) overrepresentation analysis and Gene Set Enrichment Analysis (GSEA) were performed using Clusterprofiler (version 4.2.2) and org.Ce.eg.db (version 3.14.0). A principal component analysis was also generated to explore the primary variation in the data^80,81.

For RT-qPCR, worms were collected and total RNA was extracted for the RNA-seq sample preparation. cDNA synthesis was conducted from total RNA by the Reverse Transcription Kit (Qiagen, Cat# 205314). qPCR was performed using the Light Cycler 480 SYBR green I Master kit (Roche, Cat# 04887352001). The primers used for RT-qPCR are listed in the Key Resource Table. pmp-3 was used as housekeeping controls.

Lipid extraction

The extraction procedures have been detailed previously³⁷. All reagents were chilled on ice and samples were maintained at ≤ 4°C during the extraction procedure. A metal bead was added to each sample. Next, 500 μL M1 (tert-Butyl methyl ether:Methanol = 3:1, v:v) was added to each tube and vortexed for 2 minutes. 325 μL M2 (H₂O: Methanol = 3:1, v:v) was added to each tube. Samples were vortexed briefly. Then, samples were flash-freezed in liquid nitrogen and thawed on ice. This step was done three times to facilitate cell breakage. Samples were transferred to a bead-beater and shaken at 1/25 s frequency for 5 min, and this process was done three times. The samples were then centrifuged for 10 min at 12,500 g at 4°C. For downstream lipid analysis, 200 μL of the organic layer (upper phase) was transferred to a glass autosampler vial and dried by vacuum centrifugation. Remaining protein pellets on the bottom were kept on ice until further digestion.

Once dried, organic extracts intended for lipid analysis were resuspended in 100 μL 65:30:5 Isopropanol:Acetonitrile:Water and vortexed for 20 s prior to analysis by Liquid chromatography–mass spectrometry (LC-MS). Aqueous extracts intended for metabolomic analysis were resuspended in 50 μL 1:1 Acetonitrile (ACN):Water and also vortexed for 20 s prior to analysis by LC-MS.

Protein digestion

Remaining protein pellets on the bottom were washed with 1 mL ACN and centrifuged at 10 kg for 3 min at 4°C. Supernatant ACN was aspirated and the protein pellets sit for 10-15 min at room temperature, or vacuum dried briefly to dry up the liquid in the bottom of the tube. 300 μL lysis buffer (8M urea with 100 mM tris(2-carboxyethyl)phosphine, 40 mM chloroacetamide and 100 mM tris (pH = 8.0) was added to each sample and vortex till the protein pellets were fully dissolved. 5 μg LysC was added to each sample with protein:enzyme ratio 70:1 (digestion lasted overnight at room temperature). Trypsin at 70:1 protein:enzyme was added to each sample after diluting the lysis buffer to 2 M urea and digestion lasted for six hours at room temperature. Desalting was carried out with 96 well desalting plates. A blank well between any two samples was reserved to avoid cross-contamination. Desalting started with equilibrating the desalting wells with 1 mL 100% ACN, followed by 1 mL 0.2% FA. Acidified peptide mixture was loaded onto the 96-well desalting plate, followed by 2 mL 0.2% FA wash. Peptides were eluted into a 96-well collection plate with 600 μL 80% ACN with 0.2% FA. Peptides were vacuum dried down and stored in −80°C freezer until resuspension with 0.2% FA. After resuspension, peptide concentration was measured using a quantitative colorimetric peptide assay.

LC-MS setup

Proteomics:

Peptides were separated on an in-house prepared high pressure reversed phase C18 column⁸². Briefly, a 75-360 μm inner-outer diameter bare-fused silica capillary was packed with 1.7 μm diameter, 130 Å pore size, Bridged Ethylene Hybrid C18 particles (Waters) under high pressure of 25K psi to a final length of ~40 cm. The column was installed onto a Thermo Ultimate 3000 nano LC and heated to 50 °C for all runs. Mobile phase buffer A was composed of water with 0.2% FA. Mobile phase B was composed of 70% ACN with 0.2% FA. Samples were separated with a 120 min LC method: peptides were loaded onto the column for 13 min at 0.37 μL/min. Mobile phase B increased from 0 to 6% in 13 min, then to 53% B at 104 min, 100% B at 105 min and held for 4 min at 100% B, decreased to 0% B at 110 min, and a 10 min re-equilibration at 0% B.

Eluting peptide fragments were ionized by electrospray ionization and analyzed on a Thermo Orbitrap Eclipse. Survey scans of precursors were taken from 300 to 1350 m/z at 240 000 resolution. Maximum injection time was 50 ms and automatic gain control (AGC) target was 1E6 ions. Tandem MS was performed using an isolation window of 0.5 Th with a dynamic exclusion time of 10 s. Selected precursors were fragmented using a normalized collision energy level of 25%. MS2 AGC target was set at 2E4 ions with a maximum injection time of 14 ms. Scan range was 150-1350 m/z. Scans were taken at the Turbo speed setting and only peptides with a charge state of +2 or greater were selected for fragmentation.

Lipidomics:

Extracted lipids were separated on an Acquity CSH C18 column (100 mm x 2.1 mm x 1.7 μm particle size; Waters) at 50°C using the following gradient: 2% mobile phase B from 0-2 min, increased to 30% B over next 1 min, increased to 50% B over next 1 min, increased to 85% over next 14 min, increased to 99% B over next 1 min, then held at 99% B for next 7 min (400 μL/min flow rate). Column re-equilibration of 2% B for 1.75 min occurred between samples. For each analysis 10 μL/sample was injected by autosampler. Mobile phase A consisted of 10 mM ammonium acetate in 70:30 (v/v) acetonitrile:milliQ H2O with 250 μL/L acetic acid. Mobile phase B consisted of 10 mM ammonium acetate in 90:10 (v/v) isopropanol:ACN with 250 μL/L acetic acid.

The LC system (Vanquish Binary Pump, Thermo Scientific) was coupled to a Q Exactive Orbitrap mass spectrometer through a heated electrospray ionization (HESI II) source (Thermo Scientific). Source and capillary temperatures were 300°C, sheath gas flow rate was 25 units, aux gas flow rate was 15 units, sweep gas flow rate was 5 units, spray voltage was ∣3.5 kV∣ for both positive and negative modes, and S-lens RF was 90.0 units. The MS was operated in a polarity switching mode; with alternating positive and negative full scan MS and MS2 (Top 2). Full scan MS were acquired at 17,500 resolution with 1 x 10⁶ AGC target, max ion accumulation time of 100 ms, and a scan range of 200-1600 m/z. MS2 scans were acquired at 17,500 resolution with 1 x 10⁵ AGC target, max ion accumulation time of 50 ms, 1.0 m/z isolation window, stepped normalized collision energy (NCE) at 20, 30, 40, and a 10.0 s dynamic exclusion.

The LC system (Vanquish Binary Pump, Thermo Scientific) was coupled to a Q Exactive HF Orbitrap mass spectrometer through a heated electrospray ionization (HESI II) source (Thermo Scientific). Source and capillary temperatures were 350°C, sheath gas flow rate was 45 units, aux gas flow rate was 15 units, sweep gas flow rate was 1 unit, spray voltage was 3.0 kV for both positive and negative modes, and S-lens RF was 50.0 units. The MS was operated in a polarity switching mode; with alternating positive and negative full scan MS and MS2 (Top 10). Full scan MS were acquired at 60K resolution with 1 x 10⁶ AGC target, max ion accumulation time of 100 ms, and a scan range of 70-900 m/z. MS2 scans were acquired at 45K resolution with 1 x 10⁵ AGC target, max ion accumulation time of 100 ms, 1.0 m/z isolation window, stepped NCE at 20, 30, 40, and a 30.0 s dynamic exclusion.

For mass spectrometry-based analysis, we included technical replicates and sample extraction replicates (the same sample extracted on each day of sample prep). For proteomics, technical replicates had a median CV of 12.5%, and extraction replicates had a median CV of 15.63%. For lipidomics, technical replicates had a median CV of 19.2%, and extraction replicates had a median CV of 15.6%.

Data analysis for proteomics and lipidomics

Proteomics:

LC-MS files for proteomics were searched in Maxquant (version 2.0.3.1). Original outputs from Maxquant were inspected and potential contaminant proteins, protein groups that contain proteins identified with decoy peptide sequence, and those identified only with modification site were removed. LFQ intensities were used as the quantification metric.

Lipidomics:

LC-MS files for lipidomics were processed using Compound Discoverer 3.1 (Thermo Scientific) and LipiDex⁸³. All peaks with a 1.4-23 min retention time and 100 Da to 5000 Da MS1 precursor mass were aggregated into compound groups using a 10 ppm mass tolerance and 0.4 retention time tolerance. Peaks were excluded if peak intensity was less than 2 x 106, peak width was greater than 0.75 min, signal-to-noise ratio was less than 1.5, or intensity was < 3-fold greater than blank. MS2 spectra were searched against an in-silico generated spectral library containing 35,000 unique molecular compositions of 48 distinct lipid classes⁸⁴. Spectra matches with a dot product score > 500 and reverse dot product score > 700 were retained for further analysis. Lipid MS/MS spectra that contained < 75% interference from co-eluting isobaric lipids, eluted within a 3.5 median absolute retention time deviation (M.A.D. RT) of each other, and were found within at least 4 processed files were used for identification at the individual fatty acid substituent levels of structural resolution. If individual fatty acid substituents were unresolved, then identifications were made with the sum of the fatty acid substituents. Peak intensities were normalized with the peptide amount to correct for different amounts of starting materials across the RIAIL panel.

Survival analysis and lifespan traits extraction

The survfit function of the survival (version 3.5-0) R package was used to analyze survival data. The following formula was used “survival::Surv(Age_of_death, status) ~ Strain” with default parameters. Parental strains (N2 and CB4856) lifespan from each batch was compared to check for possible batch effects. No batch correction was performed. The quantile function was used to obtain the average lifespan as well as the 25%, 50% and 75% mortality.

To estimate the association between the proportion of N2 genome in the strain and lifespan (Figure 5B), the hazard ratio of a Cox regression model was estimated with R package survival using the following formula: survival::coxph(survival::Surv(Age_of_death, status) ~ N2 genome %.

Life-history trait batch correction

The lmer function of the lmerTest (version 3.1-3) R package was used to adjust for batch effects in data collection. The following formula was used “value ~ (1∣batch/channel)” with default parameters.

Statistical analyses and correlation analysis

In the analysis of continuous variables across groups, we computed p-values using two-sided Student's t-tests to ascertain statistical significance (Figures 5D, 5E, S3C, S5A, S5B, and S5E). To explore relationships among variables we used Pearson correlations (Figures 2F, 2G, S2C, and S4A). The resulting p-values (where applicable) were corrected for multiple testing using the Benjamini–Hochberg false discovery rate (FDR).

Dirichlet regression analysis

The associations between worm shapes and life-history traits were analyzed by Dirichlet Regression using DirichletReg (version 0.7-1) R package (Figures 2H and S1E-S1F). Univariate analysis among variables was performed with default parameters following package documentation.

Variant calling and genetic map construction

Using the RNA-seq data, we performed variant calling employing the Genome Analysis Toolkit (GATK) (version 4.2.4.0)⁸⁵ following their best practices workflow (Figures S3E-S3G)⁸⁶ to genotype the 85 RIAILs strains. In brief, RNA-seq reads were mapped to the reference genome using STAR and prepared for variant calling (Mark Duplicates, SplitNCigarReads, Base Quality Recalibration). Then short variants (SNPs and Indels) were called using GATK’s HaplotypeCaller. Next, we exploited the design of the study to obtain a high-confidence set of germline variants. Comparison of identified variants with publicly available variant information for C. elegans (https://www.elegansvariation.org/) and previous genetic work on the RIAILs allowed us to perform quality control checks on the obtained variants (Figures S3E-S3G and S4). We then used the onemap (version 2.8.2)⁸⁷ software to construct a genetic map for subsequent QTL mapping. For comparison to known variants (Figure S3F), variants for the CB4856 strain were retrieved from the “Caenorhabditis elegans Natural Diversity Resource” (https://elegansvariation.org/). Soft filtered variant file was retrieved from http://storage.googleapis.com/elegansvariation.org/releases/20210121/variation/WI.20210121.soft-filter.vcf.gz. The Hard filtered variant file was retrieved from http://storage.googleapis.com/elegansvariation.org/releases/20210121/variation/WI.20210121.hard-filter.vcf.gz. These variants were then filtered for the CB4856 strain keeping only 1/1 variants with high impact consequence.

Association mapping and gene-set enrichment analysis (GSEA)

For data shown in Tables S2, S3, and S4, we used the lmekin function of the coxme R package as it allows one to model the correlation structure of the random effects. Analysis of deviance for lmekin from https://aeolister.wordpress.com/2016/07/07/likelihood-ratio-test-for-lmekin/ was used to calculate the Likelihood ratio between the null model “value ~ 1 + (1∣kinshipCov)” and the model of interest “value ~ 1 + predictorV alue + (1∣kinshipCov)”. An INT transformation was to transform the data before mapping. The Benjamini-Hochberg procedure was selected for multiple-testing correction. As the traits mapped are not independent, such a correction would be over-correcting. For Figures 3B and 3E, Tables S2 and S3, GSEA analysis was performed using Clusterprofiler (version 4.2.2) and org.Ce.eg.db (version 3.14.0). The used gene list was ranked by the signed LOD-value obtained from the association mapping analysis. The background set of genes/proteins in every analysis was always the genes and proteins detected in each assay.

Lipid class over-representation analysis

All measured lipid species were used to define lipid class sets. These were used along with the enricher function from the clusterProfiler (version 4.2.2) R package to conduct lipid class enrichment analysis (Figure S3A), which is designed to accept customized annotations through the TERM2GENE parameter. The Benjamini-Hochberg procedure was selected for multiple-testing correction. Enrichment was tested for each lifespan trait (average lifespan, 25%, 50% and 75% mortality) for two groups of lipids: positively associated (non-adjusted p-value < 0.05 & association coefficient > 0) and negatively associated (non-adjusted p-value < 0.05 & association coefficient < 0).

Quantitative trait locus (QTL) mapping

The qtl2 (version 0.34) R package⁸⁸ was used to perform QTL mapping of all phenotypic and molecular traits. An INT transformation was used to transform the data before mapping. Gene codes were encoded as N2 = 1, CB4856 = 2 and heterozygotic (N2/CB4856) = 3. Crosstype was specified as “risib”. Pseudomarkers were inserted into the genetic map with a step of 1 and default values for other parameters. Conditional genotype probabilities, kinship and genome scans were performed using qtl2 package functions with default parameters. Significance thresholds for each trait were obtained through permutation testing using the scan1perm qtl2 function with 1000 permutations. Finally, the find_peaks function was used to identify significant QTLs with a threshold of 0.05 and a drop of 0.5.

Heritability analysis

Heritability (Figure S1G) was estimated using a linear mixed model from the lme4qtl package⁸⁹, with the kinship matrix as a covariate.

QTL effects

QTL effects (Figures S5C-S5D) were plotted with the 'plot.coeff' function in the qtl2 package.

UK Biobank SNP-disease time-to-event analysis

The time-to-event analysis was performed in the UK Biobank, a population cohort of ~500,000 participants from the United Kingdom⁹⁰ (project 48020). The sample analyzed was restricted to participants of European ancestry (as determined in Pan-UKBB, https://pan.ukbb.broadinstitute.org) who were unrelated, as determined by their inclusion in the original calculation of the genetic principal components (field 22020). Time-to-event was measured from birth to the first occurrence of the event. We selected seven with at least 10,000 events as rarer diseases tended to have inconsistent and unreliable results. These data are summarized in the table below. Variants were selected from whole-exome sequencing, where at least 5 minor alleles were detected. These selection criteria resulted in 339’967 individuals and 577 and 821 SNPs for GFM1 and RICTOR, respectively. For completeness, we also analyzed KLHL28 (513 SNPs) and COL5A3 (1550 SNPs).

The time-to-event analysis was done with Cox proportional hazards in R using the Coxph function from the survival packages^49,50. The top 40 genetic principal components, sex, and the batch (specifically the initial 50k released, field 32050) were included as covariates. Close inspection of the p-value distribution for rare alleles (minor allele count ≤ 30) suggested a higher-than-expected rate of associations, so we performed permutations in each of the diseases for allele counts 5 to 30, which confirmed a high false positive rate. We used these permutations to generate a null distribution of p-values, which could then be used to estimate the true probability of observing a given p-value under the null hypothesis. These p-values were then corrected using the Benjamini-Hochberg method⁹¹.

In order to have a threshold to display in the Figures (Figures 7B and S7), we calculated an approximate threshold where the p-value would have an FDR of 0.05. This was done by successively testing different values to find the one which, if added to the existing p-values for that gene, would obtain an FDR of approximately 0.05.

Event	ICD10 code(s)	Number of events
Acute ischaemic heart disease	I21, I22, I24	18,016
Heart failure	I50	14,097
Chronic ischaemic heart disease	I25	37,124
Cerebrovascular disorders	G45, I60-I64, I67, I68	23,710
Diabetes mellitus	E11, E13, E14	31,663
Acute renal failure	N17	17,241
Chronic renal failure	N18	20,369
Death	-	25,915

UK Biobank burden-disease time-to-event analysis

The burden test addresses the low statistical power of testing for associations with rare genetic variants by combining all variants over a gene into a single test. For each variant, the genetic burden weight was defined as a product of two weights, the first based on variant effect prediction (VEP), and the second on the allele frequency.

The effect-based weight used were those proposed by Curtis⁹², combining VEP, SIFT, and PolyPhen effect predictions into weights ranging from 1 to 40. The VEP was performed using the Ensembl VEP docker image (with Singularity, for technical reasons) using the human genome assembly GRCh38, version 111. The frequency-based weight was chosen to be

$$w_f=100\ast {(0.25-q\ast (1-q))}^{1.5}+0.5$$

modified from Curtis⁹³. The burden weight was multiplied by the number of minor alleles and summed over the entire gene, providing a genetic burden for each individual, ranging from 78.0 to 994.8, which was used as a predictor in a Cox regression model with the same covariates and outcomes as the SNP-based model. P-values were corrected using the Benjamini-Hochberg method⁹¹.

Figures and visualizations

Data visualization was performed using ggplot2 (version 3.4.2). The resulting p-values (where applicable) were corrected for multiple testing using the Benjamini–Hochberg false discovery rate. Clusterprofiler (version 4.2.2) was used to generate graph representations of enrichment results (Figures 3B and 3E). The R package enrichplot (version 1.14.2) was used to generate running GSEA plots (Figure 3C). UpSetR (version 1.4.0) was used to generate upset plots (Figures S2B, S2D, and S2F).

Data availability

The RNA-Seq data generated in this study have been deposited in the GEO database (GSE252593). The remaining data generated in this study are provided in the Source Data files. Scripts for analysis and figure generation have been deposited in a GitHub (https://github.com/auwerxlab/Project_RIAILs) repository along with additional data used in this work.

Fluorescent image for assessing the UPR^mt activation and cytosolic stress response

RNAi bacteria were cultured overnight in lysogeny broth (LB) medium containing 100 mg/mL ampicillin at 37°C. Then the bacteria were five times concentrated and seeded onto RNAi plates. Random L4/young adult worms were picked onto the RNAi bacteria-seeded plates and cultured at 20°C until their progenies reached the young adult stage. 6 - 10 worms were then randomly picked in a drop of 20 mM tetramisole (Cat. T1512, Sigma) and then aligned on an empty NGM plate. Fluorescent images, with the same exposure time for each condition, were captured using a Nikon SMZ1000 microscope. Positive control worms of cytosolic stress response were prepared by culturing L4 hsp-16.2p::gfp fed on E. coli HT115 at 30°C for 8 h and recovered overnight at 20°C. The experiment was repeated three times.

Real-time quantitative PCR (RT-qPCR)

For qRT-PCR, worms were cultured and total RNA was extracted for the RNA-seq sample preparation. cDNA synthesis was performed using the Qiagen Reverse Transcription Kit (205314) from the extracted RNA samples. The qPCR was then conducted with the Roche Light Cycler 480 SYBR Green I Master kit (Cat. 04887352001). The specific primers utilized are detailed in the key resources table, with the pmp-3 primer serving as housekeeping control. Measurements were repeated twice.

OCR measurements by Seahorse

Oxygen consumption rate (OCR) was assessed using the Seahorse XF96 (Seahorse Bioscience), following the protocol outlined in⁹⁴. Briefly, a synchronized culture of ~100 worms was harvested on day 1 of adulthood with sterile M9 buffer. After three washes in the M9 buffer, the worms were transferred to a 96-well Seahorse plate, where their OCR was measured six times to determine mitochondrial activity for each condition at basal level and another six times measurement after adding 10 μM FCCP as the final concentration. Measurements were repeated twice.

Oil red O staining

Worms were cultured at 20°C and fed with E. coli HT115 or RNAi on RNAi plates. C. elegans synchronization by bleaching, transfer synchronized worms to new plates every 2 days until they reach the desired stages. At day 1 of adulthood, worms were collected, washed twice with 1 x PBS and subsequently resuspended in 120 uL of PBS. An equal volume of 2x MRWB buffer (containing 160 mM KCl, 40 mM NaCl, 4 mM EGTA, 30 mM PIPES at pH 7.4, 1 mM spermidine, 0.4 mM spermine, 2% paraformaldehyde, and 0.2% beta-mercaptoethanol) was added for 1 h. After fixation, the worms underwent three freeze-thawing cycles using dry ice and a 37°C water bath, followed by centrifugation for 1 min at 14,000 x g and then washed once in PBS buffer. Prior to Oil Red O (ORO) staining, worms were re-suspended and dehydrated in 60% isopropanol for 15 minutes. Each sample was treated with 250 uL of 60% ORO (pre-heated in an 85°C water bath for 2 h, and ddH₂O at a ratio of 3:2) stain and incubated overnight at 4°C. Following ORO staining, the worms were washed twice with a 60% isopropanol solution⁹⁵. Worm images were acquired with the Leica DM5500 Upright Microscope. The lipid staining was quantified with the open-source image analysis software package FIJI/ImageJ from https://imagej.net/Fiji. The experiment was repeated three times.

QUANTIFICATION AND STATISTICAL ANALYSIS

No statistical methods were applied to pre-determine worm sample size. Comparison between more than two groups was assessed by using a One-way ANOVA test. Prism 8 (GraphPad Software) was used for statistical analysis of all lifespan, qRT-PCR, OCR, paralysis and ORO staining experiments. Variability in panels is given as the s.e.m. All p<0.05 were considered to be significant (****p<0.0001; ***p<0.001; **p<0.01; *p<0.05; n.s., not significant). For lifespan measurement, a log-rank test was used to determine the significant difference. For lifespan, and OCR measurement in worms, sample size was determined based on the known variability of the experiments. All experiments were done non-blinded.

Supplementary Material

Table S1

Table S1: Details of the phenotypic traits collected using the microfluidics device. Related to Figures 2 and S1.

Table S2

Table S2: Correlations of lifespan traits and mRNAs (both single mRNA and GSEA). Related to Figures 3 and S2.

Table S4

Table S4: Associations between lifespan traits and lipids. Related to Figures 4 and S3.

Table S5

Table S5: Lifespan and phenotypic QTL identified above the suggestive threshold (p<0.1), and genetic markers and their LOD score under selected QTL peaks. Related to Figures 5, S3-S5.

Table S7

Table S7: Details of replicates, sample size, and analyses applied in the study. Related to Figure 2-6.

Table S6

Table S6: Association of SNPs in human GFM1 with heart failure in the UK Biobank database, and burden test associations of genes with disease outcomes. Related to Figures 7 and S7.

Table S3

Table S3: Associations between lifespan traits and proteins (both single proteins and GSEA). Related to Figures 3 and S2.