To select compounds for CITP testing (Figure 1A), we began with the top 10% of candidates from a published list of computationally ranked compounds built using known drug-protein interactions (Fuentealba et al., 2019). To avoid duplicative effort and to favor novel discovery, we used the DrugAge database (Barardo et al., 2017b) to de-prioritize compounds that had already been published to extend C. elegans lifespan. We then selected 16 candidate interventions by cross-referencing the remaining compounds with the top 10% of two additional computational efforts that predicted aging effects based on comparative transcriptional responses (Janssens et al., 2019) (all-trans retinoic acid, arecoline, propranolol, thalidomide) and machine-learning models based on gene ontologies and physical structures (Barardo et al., 2017a) (aldosterone, berberine, bortezomib, dasatinib, decitabine, dexamethasone, erlotinib, gefitinib, ritonavir, temsirolimus), or listing in the DrugAge database with published lifespan extension in other systems everolimus (Spindler et al., 2012) and fisetin (Yousefzadeh et al., 2018).

The selected compounds comprise a number of common aging-related classes, including bortezomib (a proteasome inhibitor; Chen et al., 2011), fisetin (a sirtuin activator; Kim et al., 2015), temsirolimus (a PI3K/mTOR inhibitor and derivative of rapamycin; Ali et al., 2022), and dasatinib (a tyrosine kinase inhibitor; Talpaz et al., 2006), among others (Table 1). We then screened the selected compounds at 4–5 concentrations using full lifespan analysis (Figure 1—figure supplement 1). Among the 16 candidate interventions, 3 were water-soluble and 13 were DMSO-soluble. While DMSO can impact lifespan (Wang et al., 2010), we did not observe a difference between the H₂O and DMSO vehicle control treatments (median lifespan 17 days for both, p = n.s.), consistent with the published absence of DMSO effects at concentrations similar to those used in our studies (AlOkda and Van Raamsdonk, 2022). Among the 16 computationally prioritized candidate compounds, we found that aldosterone, dexamethasone, erlotinib, decitabine, dasatinib, everolimus, thalidomide, and temsirolimus did not significantly change median lifespan at any tested concentration (Figure 1B; Figure 1—figure supplement 1).

Among the eight remaining candidates, we found that three compounds shortened median lifespan (arecoline, gefitinib, and bortezomib). Tests of muscarinic/nicotinic agonist arecoline at five concentrations ranging from 50 µM to 8 mM revealed toxic effects at the highest concentration (–29.4% median lifespan, p < 0.0001). The epidermal growth factor inhibitor gefitinib also had small, but significant, negative effects at 10, 25, 50, and 80 µM (Figure 1—figure supplement 1). In contrast, the 26S proteasome inhibitor bortezomib conferred strong toxicity effects that increased with concentration through the entire concentration range we tested (5, 10, 20, and 30 µM, –10.5% to –47.3% median lifespan; p = 0.0002 at 5 µM and p < 0.0001 at all other concentrations; Figure 1—figure supplement 1). Thus, some compounds computationally predicted to enhance longevity can be found to be toxic when empirically investigated, underscoring that validation is a key element of any prediction pipeline.

We found that the remaining five compounds conferred statistically significant positive effects on median lifespan for at least one tested concentration (Figure 1B; Figure 1—figure supplement 1), representing a hit success rate of 31.25% (5/16) for our test set (Table 1). Seven additional compounds met our selection criteria, but because they had been previously tested by the CITP, we did not include these compounds in tests presented here (see Table 1). With those previously tested compounds included, we see a similar overall hit success rate of ~30% (7/23) (Table 1). The bioactivity of these compounds was as follows: ritonavir had effects at 20 and 35 µM (6.3%, p = 0.0016 and p = 0.0282, respectively). The sirtuin activation/mTOR inhibitor fisetin had positive effects at 10, 50, and 100 µM (p = 0.0081, p = 0.0025, and p = 0.0032, respectively), with an effect size up to 11.8%, and no effect detected at 20 µM. The AMPK activator berberine conferred significant effects only at 100 µM, with an 11.8% increase in median lifespan (p < 0.0001). In support of the potential translatability of the computationally predicted candidate compounds tested here, fisetin (Yousefzadeh et al., 2018) and berberine (Dang et al., 2020) have also been found to increase median lifespan in mice.

The two remaining candidate interventions induced large increases in longevity, with propranolol extending median lifespan 44.4% at 1 mM (p < 0.0001) and atRA extending median lifespan 23.5% at 150 µM (p < 0.0001) (Figure 1B). Propranolol and atRA are particularly interesting interventions because they are both FDA-approved drugs, potentially providing an easier path toward clinical use as aging interventions. For example, propranolol is a well-tolerated drug with a long history of use (Srinivasan, 2019, p. 50; Zacharias et al., 1972) treating high blood pressure (Prichard and Gillam, 1964), angina (Hamer et al., 1964), and atrial fibrillation (Rowlands et al., 1965). While we observed a large (44.4%, p < 0.0001) median lifespan extension at 1 mM propranolol, we also saw extension at 0.5 mM (16.7%, p < 0.0001) (Figure 2D). In contrast to the positive effects that we observed at 0.5 and 1 mM, we found that increasing the treatment concentration to 5 mM propranolol resulted in toxicity and a reduction in lifespan (–61.1%, p < 0.0001) (Figure 2D). These observations led us to extend our tests into the related species C. briggsae (AF16) and C. tropicalis (JU1630). In these two species, we observed similar toxicity at 5 mM (p < 0.0001), but no beneficial effects at lower concentrations (Figure 2E, F).

In human applications, propranolol functions as a general antagonist of β1 and β2 beta-adrenergic receptors. We therefore sought to determine if a β1-specific antagonist like metoprolol could recapitulate the longevity effects in C. elegans. We assayed longevity effects for metoprolol across a concentration range of 5 µM to 1.5 mM and observed no positive effects (Figure 2—figure supplement 2A). Although there could be multiple reasons that metoprolol was not effective, we followed up by asking whether the effects of propranolol require β2 antagonism or are unrelated to β-adrenergic antagonism. A β-adrenergic-independent mechanism was suggested by the change we noted in the appearance of the bacterial lawns on propranolol-treated plates. When we tested bacterial growth, we found that propranolol reduced bacterial growth at the same concentrations at which we observed lifespan effects (Figure 2—figure supplement 2C). The propranolol impact on bacterial food source growth suggests a potential indirect food-dependent mechanism for propranolol on C. elegans lifespan. We therefore repeated the lifespan studies at 0.5 and 1 mM in the presence of paraformaldehyde-treated bacteria that are metabolically inert (Beydoun et al., 2021). Under conditions in which propranolol effects on bacterial growth were eliminated, we observed shorter lifespans in populations treated by propranolol (Figure 2—figure supplement 2B). These observations suggest that propranolol either does not exert direct beneficial effects on lifespan in C. elegans or has confounding direct and indirect effects that depend on bacterial food state. The potential food-dependent effects of propranolol require further study beyond the scope of our current screening protocols. Therefore, we elected to focus on the large lifespan extension generated by atRA treatment for the remainder of this study.

The scientific literature is rife with examples of intervention effects on longevity that vary in response to experimental differences. Indeed, previous experiments treating C. elegans with atRA have resulted in contradictory effects (Janssens et al., 2019; Statzer et al., 2021) for reasons that are not entirely clear. To determine the most efficacious concentration of atRA treatment, we tested a dosage range from 1 to 150 µM. For C. elegans N2, we observed increasing positive effects at all tested concentrations above 1 µM (which had no detectable effect; Figure 2) and a very slight but significant increase in median lifespan for C. tropicalis JU1630 at 150 µM atRA (7.7%, p = 0.0215). However, we observed no effects in C. briggsae across the tested concentration range (Figure 2B). Given these observations, we elected to use 150 µM atRA for the remainder of the experiments in this study. In contrast with propranolol, we observed no obvious effects of atRA on bacterial growth using replica plating.

Expanding this analysis across a more extensive genetic diversity set following the full CITP replication protocol, we tested the effects of atRA on three strains of C. elegans (N2, JU775, MY16), C. briggsae (AF16, ED3092, HK104), and C. tropicalis (JU1630, JU1373, QG834) (Figure 3) with replication at three distinct geographic sites (University of Oregon, Rutgers University, and the Buck Institute). Total genetic variation across C. elegans strains is comparable to that observed among humans, while the differences among species are comparable to the genetic distance between humans and mice (Teterina et al., 2022). We found that the substantial atRA-associated effects on longevity are robust to genetic variation across all three C. elegans strains, yielding lifespan extensions of 18.8–44.4% (Figure 3; Figure 3—figure supplement 1A). While the lifespan extension initially observed in C. tropicalis JU1630 failed to replicate, C. tropicalis QG834 displayed a small but significant increase in lifespan (Figure 3; Figure 3—figure supplement 1C). Clearly, given the small effect size, we are at the edge of statistical power to detect a positive effect within this species. Again, atRA did not register any significant positive effects in the C. briggsae strains (Figure 3; Figure 3—figure supplement 1B). Partitioning total variation across this large set of experimental replicates, we found only a small amount of variability attributable to site (3.9%) or differences among experimenters (7.1%), with the majority of variance being attributable to individual variation (53.2%; Supplementary file 1a), consistent with previous CITP studies (Banse et al., 2024a; Lucanic et al., 2017). Thus, atRA treatment is reproducible within and between laboratories, but subject to high levels of individual variation, as are all longevity studies.

To increase the temporal resolution of our survivorship curves, we repeated our longevity analyses using the Automated Lifespan Machine (ALM) technology (Stroustrup et al., 2013) across the same nine genetically diverse strains. We again observed a positive effect for the three C. elegans strains, demonstrating the robustness of atRA longevity effects across genetic backgrounds in these strains (Figure 3—figure supplement 2). Interestingly, we do not see any positive effects of atRA on the ALM for C. briggsae or C. tropicalis, but instead see slightly toxic effects for C. briggsae strains AF16 and HK104 and C. tropicalis strain JU1630. It is not clear what might drive this difference, although the ALM introduces some different environmental stresses and conditions as compared to manual assay conditions (e.g., repeated light exposure and distinct compound introduction; see Banse et al., 2019 for discussion).

A goal of longevity interventions is to enhance physiological health, which, like in humans, can be measured as improvement in older age locomotory capacity. We therefore determined the effect of atRA exposure on aging adult swim performance using a video analysis of swimming behavior (Ibáñez-Ventoso et al., 2016; Restif et al., 2014). In previous work, we reported that anti-aging interventions can have disparate effects on longevity and adult swimming ability and that treated strains can show positive effects in motility enhancement (Banse et al., 2024b) even in the absence of longevity enhancement. Using strain-specific models for swimming behavior to generate a composite swimming score based on eight underlying measures (Banse et al., 2024b), we observed significant improvement for two of the C. elegans strains at day 12 of adulthood (Figure 3—figure supplement 3). Similar to atRA effects on longevity, we find that atRA was generally ineffectual at promoting swimming health in C. briggsae and C. tropicalis, with improvements only seen in day 16 of AF16 (41.6%, p = 0.00191), while decreased swimming scores were observed in all three C. tropicalis strains at one or more test days. Overall, then, atRA has largely positive effects on C. elegans longevity and health while it has the potential to be detrimental to C. briggsae and C. tropicalis depending on the assay type and particular genetic background.

Given the plasticity of genetically determined longevity within C. elegans, we next sought to identify the pathways required for atRA lifespan extension. Because no retinoic acid binding transcription factors have been identified in C. elegans, we looked to the known effects of atRA in modulating human kinase activity (Albalat, 2009; Albalat and Cañestro, 2009; Fonseca et al., 2020) to identify candidate pathways. The vertebrate atRA-responsive kinases do have extensively studied orthologs in C. elegans. Akt homologs emerged as particularly relevant due to their involvement in longevity-related IIS signaling, and the fact that in mammals, phosphorylation of Akt in response to atRA occurs at a site that appears to be conserved in the C. elegans Akt homologs (García-Regalado et al., 2013). We therefore asked whether either akt-1 or akt-2 was required for atRA lifespan effects by measuring the lifespans of akt-1(ok525) and akt-2(ok393) loss of function mutants (Figure 4A, B). Consistent with previously published studies (Newell Stamper et al., 2018; Soukas et al., 2009), in control-treated animals, we observed longer median lifespans for the mutants (median 26 and 23 days vs. 17 in WT). When we performed longevity analysis of akt-1(ok525) in the presence of atRA (Figure 4A), we observed a significant decrease in longevity (–7.7% median lifespan, p = 0.00000142), demonstrating a requirement for AKT-1 in atRA-induced longevity extension. Repeating the analysis in akt-2(ok393) mutants (Figure 4B) demonstrated a complete dependence on AKT-2, with atRA having no significant effect in the akt-2 mutant background (p = 0.7170). We conclude that atRA longevity effects require akt-1 and akt-2, consistent with a known atRA signaling mechanism in mammals.

Similarly, in human cell culture (Ishijima et al., 2015) and mouse models (Zhang et al., 2019), retinoic acid activates AMPK. AMPK is a conserved sensor of intracellular energy state that regulates glucose and lipid metabolism. Previous work has shown that AMPK is involved in the transition to gluconeogenesis in the long-lived dauer stage (Penkov et al., 2020) and may also play a role in gluconeogenesis in adults (Nguyen et al., 2020). In C. elegans, AMPK is required for many longevity interventions (Jayarathne et al., 2020; Onken and Driscoll, 2010; Peng et al., 2019), while overexpression of the AMPK catalytic subunit aak-2 (Apfeld et al., 2004), or expression of a constitutively activated AAK-2 (Greer et al., 2007; Mair et al., 2011), can directly increase lifespan. Additionally, AMPK both regulates and is regulated by Akt, making AMPK an interesting candidate for involvement in atRA longevity effects. We therefore tested aak-2(ok524) mutants for atRA lifespan extension. We observed that the atRA lifespan extension was fully dependent on aak-2 (Figure 4C). Overall, our data suggest that atRA longevity may be mediated through a conserved signaling process.

Another set of atRA transducing factors signaling are hsf-1, the human heat shock factor 1 transcription factor homolog, and skn-1, the nematode homolog of the mammalian Akt-target Nrf2. Previous work has demonstrated that HSF-1 is required for several lifespan-extending genetic and pharmacological interventions in C. elegans (Lazaro-Pena et al., 2022; Steinkraus et al., 2008; Todorova et al., 2023) and that skn-1 can directly promote lifespan (Tang and Choe, 2015; Tullet et al., 2017), as well as being implicated in multiple longevity interventions (Duangjan et al., 2019; Seo et al., 2015Seo et al., 2015) including vitamin D (Mark et al., 2016) and thioflavin T (Alavez et al., 2011; Lucanic et al., 2017). In mammalian studies, Akt directly regulates HSF1 through phosphorylation (Carpenter et al., 2015; Da Costa et al., 2020; Lu et al., 2022Lu et al., 2022; Tang et al., 2020), and hsf-1 has been implicated as a downstream effector of PI3K/Akt signaling that functions in conjunction with DAF-16 to regulate lifespan in C. elegans (Chiang et al., 2012; Hsu et al., 2003). We tested for atRA lifespan extension in a hsf-1(sy441) mutant encoding a premature stop codon that removes the conserved transactivation domain (Hajdu-Cronin et al., 2004). In this hsf-1 background, atRA treatment had a greatly reduced impact on lifespan, showing only a small increase in median lifespan and no extension in maximum lifespan (Figure 4D).

Because skn-1 is an essential gene required developmentally to specify mesodermal fates (Bowerman et al., 1992; Maduro et al., 2001), we tested for longevity effects of atRA in animals fed HT115 E. coli carrying an RNAi vector targeting skn-1 (Kamath and Ahringer, 2003) starting from the L3/L4 developmental stage (skn-1 is an essential gene for early development, and thus assaying a knockout mutation or beginning with earlier interventions is not possible). It should be noted that we observed that the HT115 strain of E. coli itself extends lifespan relative to strain OP50, consistent with previous findings (Stuhr and Curran, 2020; Figure 4—figure supplement 1B). When we compared skn-1 RNAi-treated animals exposed to atRA versus vehicle control, in contrast to the lifespan extension for animals under control RNAi conditions, we observed a 23.8% decrease in median lifespan (p < 0.001) (Figure 4E and Figure 4—figure supplement 1B), suggesting that atRA is toxic in the absence of skn-1 function. Our data are consistent with hsf-1 and skn-1 being necessary for the lifespan-extending transcriptional response to atRA and/or for addressing potential toxic side effects of atRA.

Given the dependence on akt-1/2, we sought to determine if atRA lifespan extension requires the canonical C. elegans AKT-target daf-16/FOXO, a known regulator of aging (Murphy and Hu, 2013) for which activation is a common feature of chemical interventions that extend C. elegans lifespan (Kim et al., 2019; Wang et al., 2015; Zhao et al., 2017) (although daf-16 independent lifespan extension is also possible; Onken and Driscoll, 2010). We therefore measured longevity in daf-16(mu86) null mutants treated with atRA. We found that daf-16(mu86) animals still exhibited a lifespan extension (12%, p < 0.0001) compared to vehicle control animals (Figure 4F). The fact that the atRA longevity effect size is larger in wildtype animals (24% vs. 12%) reveals that although daf-16 contributes in part to the atRA effect, DAF-16 is not absolutely required, and therefore additional or alternative outputs must be operative. Inputs to longevity pathways are well documented to be complex and inter-related (Narasimhan et al., 2009; Nikoletopoulou et al., 2014; Parkhitko et al., 2020). For example, akt-1 and akt-2 are primary upstream modulators of daf-16 in the IIS pathway regulation of the long-lived alternative dauer larval state (Paradis and Ruvkun, 1998), but have little effect on IIS modulation of adult longevity, when sgk-1 becomes the primary regulator of DAF-16 (Hertweck et al., 2004). We conclude that atRA acts in part via DAF-16 but infer that atRA either acts independently of the IIS pathway, or primarily through the PI3K/Akt portion of the IIS pathway, which would be consistent with Akt-dependent atRA signaling in mammals (Bastien et al., 2006; García-Regalado et al., 2013; Qiao et al., 2012).

While our observation that atRA requires AKT-1/2 and SKN-1 is consistent with a simple signaling cascade in which atRA modulates Akt regulation of SKN-1, more complicated responses are possible. Previous research has established that there is crosstalk between Akt and p38 MAPK signaling in humans (Gonzalez et al., 2004) and in C. elegans (Tullet et al., 2008). Genetic and biochemical analysis of SKN-1 has shown that in addition to Akt regulation, SKN-1 is also post-translationally regulated by the pmk-1/p38 MAPK pathway (Figure 4—figure supplement 1A). Importantly, work in mammals has implicated p38/MAPK signaling in atRA responses (Alsayed et al., 2001; De Genaro et al., 2013; Hormi-Carver et al., 2007; Lee et al., 2008; Roe et al., 2020; Shinozaki et al., 2007), suggesting that atRA may affect two different signaling cascades that can regulate SKN-1.

In light of these considerations, we addressed p38 MAPK signaling as a potential effector pathway for atRA. C. elegans has three known p38 mitogen-activated protein kinase homologs, pmk-1, pmk-2, and pmk-3. SKN-1 is regulated by the MAPK cascade that culminates with p38/PMK-1 phosphorylation of SKN-1 at serines 164 and 430 (Figure 4—figure supplement 1A). The phosphorylation of S164 and A430 sites results in increased nuclear SKN-1 levels, resulting in transcription of innate immunity and oxidative stress genes (Inoue et al., 2005). When we tested pmk-1(km25) mutants for atRA longevity effects, we found that pmk-1(km25) mutants exhibit an atRA-induced extension in median lifespan (26.3%, P<0.0001), but did not exert an effect on maximum lifespan, suggesting enhanced importance of pmk-1 later in life (Figure 4G). Although our data identify AKT-1 and AKT-2 as more impactful than p38/PMK-1 in atRA-mediated longevity, additive and more complex atRA regulation of SKN-1 by PI3K/Akt and p38 MAPK pathways may be possible. For example, Akt regulates SKN-1 through phosphorylation of serine 12 (Blackwell et al., 2015), while pmk-1 regulates SKN-1 through serines 164 and 430 (Figure 4—figure supplement 1A).

To probe the candidate signaling pathways further, we considered potential pathway receptors. C. elegans PMK-1 functions downstream of TIR-1 (Liberati et al., 2004; Peterson et al., 2022), one of two Toll/interleukin-1 receptor homology (TIR) domain-containing genes (Paysan-Lafosse et al., 2023). The second TIR-domain containing protein is the membrane-associated TOL-1, which signals through a p38 MAPK cascade including mom-3 and pmk-3, and ultimately IKB-1. We measured the lifespan of tol-1(nr2033) mutants treated with atRA to find that tol-1(nr2033) animals exhibit an enhanced response to atRA, with an 82.4% increase in median lifespan in the mutant background relative to the 23.5% increase observed in the N2 wildtype background (Figure 4H).

One widely conserved mechanism for lifespan extension is caloric restriction. In C. elegans longevity research, one frequently used caloric restriction model is genetic mutation of eat-2. EAT-2 is a nicotinic acetylcholine receptor expressed in the pharyngeal muscle that facilitates normal, fast feeding behavior (Avery, 1993; McKay et al., 2004; Raizen et al., 1995). In an eat-2 mutant background, feeding behavior is slowed, inducing a caloric restriction state that extends life (Lakowski and Hekimi, 1998), either via dietary restriction itself or via a combination of dietary restriction and an innate immunity response to altered bacterial processing (Kumar et al., 2019). Previous work has shown that some compound interventions are incapable of further prolonging eat-2 lifespan (e.g., metformin; Onken and Driscoll, 2010), while other interventions appear independent/additive (e.g., Sonneradon A; Jiang et al., 2022) with eat-2 effects. We were particularly interested in the possibility that atRA might act as an eat-2-like dietary restriction mimetic because previous characterization demonstrated that eat-2 longevity was independent of daf-16 (Lakowski and Hekimi, 1998), but dependent on skn-1 (Park et al., 2010), mimicking our observations for atRA. We therefore treated eat-2(ad1113) mutants with vehicle and atRA. Consistent with atRA acting through a mechanism distinct from eat-2, we observed a significant atRA-induced extension in lifespan in the eat-2(ad1113) animals (36.8% increase in median survival, p < 2e−16) (Figure 4I). In C. elegans studies, caloric restriction can be induced through several different experimental regimes, each of which requires a different set of genetic pathways to exert longevity effects (Greer and Brunet, 2009). We therefore conclude that atRA longevity effects are either unrelated to and/or are mechanistically distinct from eat-2 effects on longevity.

Given the dependency of lifespan extension under atRA treatment, we used RNA-seq to assess transcriptional changes under atRA treatment in wildtype N2 animals. We performed RNA-seq on day 4 adult animals treated with atRA compared to carrier control. We were able to detect the expression of 12,746 C. elegans genes in our dataset (Figure 5A). Among the detected genes, 17% (2169) were differentially expressed with atRA treatment (FDR <0.05). We defined the subset of differentially expressed genes with an absolute log₂ fold change (LFC) greater than one as wildtype atRA response genes (WaRGs). The WaRGs represent ~5.1% of all detected genes (653 total) and were more heavily weighted toward downregulated genes, with 487 (3.8% of total) downregulated versus 166 (1.3% of total) upregulated. Analysis of the expression pattern of the WaRGs shows a skewed distribution among the upregulated genes, with 86% (138/160; q = 5.9e−38) of the genes with characterized expression being produced in the intestine. The downregulated genes are enriched for genes expressed in the excretory duct (32/470; q = 4.1e−9), excretory socket cell (29/470; q = 5.1e−08), and the epithelial system (281/470; q = 2.7e−8). Potentially relevant to metabolic regulation of aging, the intestine and the hypodermal cells are the primary energy storage tissues in C. elegans (Mak, 2012; Mullaney and Ashrafi, 2009).

A WormCat 2.0 (Higgins et al., 2022; Holdorf et al., 2020) enrichment analysis of the WaRGs (Figure 5B) showed overlapping and differing enrichments between up- and downregulated genes. For example, we noted stress-related gene enrichments in both classes of WaRGs, consistent with known PI3K/Akt and p38 MAPK functions in C. elegans and with the strong correlation between stress response and longevity (Zhou et al., 2011). Among the non-overlapping enrichments, we found that the upregulated set was enriched for metabolism and transmembrane transport genes (Figure 5B). We found that collagen and neuropeptide-related genes (Figure 5B) were disproportionally represented among the non-overlapping but downregulated set. We were particularly surprised by the former because atRA was identified as a longevity modulator through induction of a collagen (col-144::gfp) reporter (Statzer et al., 2021) and some observations correlate longevity with collagen expression (Ewald et al., 2015; Goyala and Ewald, 2023). Separation of the collagen and collagen-related genes by type Teuscher et al., 2019 demonstrated a general trend of atRA either not changing or downregulating collagen genes (Figure 5—figure supplement 1). For example, the cuticular collagens, of which col-144 is a predicted member, and other core genes associated with the extracellular matrix (matrisome genes) were either unchanged in expression or downregulated (Figure 5—figure supplement 1). In contrast with the core-matrisome genes, the matrisome-associated category did include a number of upregulated genes in the ECM-regulator and ECM-affiliated subclasses.

We also analyzed the WaRGs from a metabolic perspective using the WormFlux Pathway enrichment tool (Walker et al., 2021). Among the 166 upregulated WaRGs, we documented an enrichment of sphingolipid metabolism (8/45 genes, p_enrichment = 2.3e−07) (Figure 5—figure supplement 2) and iron metabolism (2/15 genes, p_enrichment = 0.025) pathway genes, while the 487 downregulated WaRGs were enriched for fatty acid biosynthesis (6/24 genes, p_enrichment = 0.00051), fatty acid degradation (2/8, p = 0.048), folate biosynthesis (2/8 genes, p_enrichment = 0.048), and UGT enzyme (8/67, p_enrichment = 0.0094) pathway genes. Interestingly, WormFlux analysis also suggests that genes related to the electron transport chain may be under-represented (0/88 genes, p_depletion = 0.013) among the downregulated WaRGs.

Because of the potential overlap of enriched gene categories with the functions of the IIS-PI3K/Akt and Nrf2-p38 MAPK pathways in C. elegans, we wanted to determine if the genes with the largest fold change in expression were among the known IIS and Nrf2 regulons. Focusing on genes with a significant absolute LFC >3, we observed that among the 24 most upregulated genes, 83% (20) have previously been observed to be regulated by the IIS pathway and 71% (17) have been observed to be regulated by the Nrf2 pathway (Supplementary file 1b). Interestingly, the four genes without a known connection to the IIS pathway appear non-random. For example, the most upregulated gene (E02C12.10, LFC = 9.9) is a member of a family of 31 paralogs in C. elegans predicted to have kinase-like activity (Davis et al., 2022; Vilella et al., 2009). Interestingly, the E02C12.10 gene, which clearly merits further investigation, was also identified as a significant contributor to survival of AMPK-deficient dauer larvae using a genome-wide RNAi screen (Xie and Roy, 2012). Two additional members of the ‘most upregulated’ gene set (E02C12.12 and E02C12.6) were also members of this gene family, in addition to 10 additional genes in the upregulated WaRGS, representing 32% of all family members and ~48% (10/21) of the family members detected in our dataset, a significant enrichment over the observed rate (1.3%, p < 0.0001). In contrast with the upregulated WaRGs, none of the family members were classified as downregulated WaRGs. The function of these genes is unknown, but the family is defined by a putative protein kinase domain and a nuclear hormone receptor-like structure (Davis et al., 2022) that suggests a potential for transduction of an atRA regulatory response. There were fewer downregulated genes with an LFC <−3, with only 12 genes reaching the threshold. Among those 12 genes, 100% have previously been shown to be regulated by both the IIS and Nrf2 pathways (Supplementary file 1c).

Overall, consistent with our genetic results, there is a clear footprint of atRA activity across a broad set of stress-response and longevity-related pathways, with some indication of novel activity as well.

To further dissect the transcriptional response to atRA in detail, we repeated our transcriptional analysis in several mutant backgrounds. Using hsf-1(sy441) mutants, we were able to identify mRNA from 13,737 genes (compared to 12,746 in N2). A comparison of transcriptional responses to atRA for all genes shows that there is a strong correlation between the N2 and hsf-1(sy441) expression changes (R² = 0.306, p < 0.0001) (Figure 6A). Using the same cutoffs that we used for our wildtype dataset to define WaRGs, we determined that the atRA regulon for hsf-1(sy441) animals (298/13737) is ~42% the size of wildtype (653/12,746), with half of the response unique to hsf-1(sy441) (Figure 6B). Comparing the 653 WaRGs from the general analysis with the subset identified in hsf-1 mutants, ~96% (629/653) were detected in both datasets. Among the 470 downregulated WaRGs, only 14% (64/470) still meet the WaRG thresholds in the hsf-1 background. Among the 158 upregulated WaRGs that were detectible in our hsf-1(sy441) dataset, half (79/158) of the genes still met the threshold for classification as a WaRG. The loss of differential expression could result from fewer genes changing expression or by a decrease in magnitude of the response that drops genes below our current threshold for defining WaRG genes. This potential for ‘lost’ regulation would be particularly skewed for genes whose expression change was near the absolute LFC = 1 threshold, where a negligible change could alter the categorization of the response. We therefore categorized the WaRG response as being maintained (0.5–2x WT response), weakened (<0.5X WT), lost (FDR >0.05), or flipped in the hsf-1 background (Figure 6C). We found that a greater portion of the downregulated WaRG response was dependent on hsf-1, with only 29% of the downregulated response being maintained, compared to 61% of the upregulated response (Figure 6D). We then sought to determine if the lost and maintained WaRGs represented unique functions by performing an enrichment analysis using the WormCat analysis tool (Figure 6—figure supplement 1). We found that the maintained response was enriched for sphingolipid (6/44, p = 6.1e−06), sterol (6/59, p = 6.9e−04), and short chain dehydrogenase (6/42, p = 0.00011) metabolism genes. Additionally, the maintained response was also enriched for C-type lectin (10/256, p = 0.00029), CYP detoxification (10/82, p = 1.7e−07), and CUB pathogen (4/25, p = 0.004) stress-response genes. The same analysis of the lost WaRG response genes suggests an hsf-1 dependence for atRA regulation of solute carrier (12/197, p = 0.000103), neuropeptide (10/139, p = 0.000196), and stress-response genes. Thus, hsf-1 plays an important, but hardly absolute, role in mediating the atRA longevity response.

To determine the role of AMPK in regulating the transcriptional response to atRA, we repeated our RNA-seq analysis in aak-2(ok524) mutants. A comparison of the transcriptional response to atRA for all genes shows that there is a strong correlation between the N2 and aak-2(ok524) data sets for expression changes (R² = 0.589, p < 0.0001) (Figure 6E). Using the same cutoffs that we used for our wildtype dataset to define WaRGs, we determined that the atRA regulon for aak-2(ok524) mutants (802/12,611) is actually larger than wildtype (653/12,746), with more than half (456/802) of the response unique to aak-2(524) (Figure 6F). We next sought to determine what portion of the wildtype atRA transcriptional response was lost in aak-2 mutants. We therefore analyzed the 96% of the WaRGs (627/653) that were detected in our aak-2 dataset and found that 76% (122/160) of the upregulated and 79% (367/467) of the downregulated WaRGs were similarly regulated in aak-2 animals (Figure 6G, H). A WormCat 2.0 analysis (Figure 6—figure supplement 1) of the maintained atRA response demonstrated enrichment for secreted extracellular proteins (7/54, p = 0.004), collagen (12/184, p = 0.006), and solute carrier (18/197, p = 8.67e−7) genes. The maintained response was also enriched for sphingolipid (6/44, p = 1.5e−06) metabolism genes, suggesting that this metabolism category of changed expression observed in wildtype animals is upstream of aak-2. Additionally, there is an enrichment for caenacin (4/11, p = 0.007), and CUB pathogen (5/25, p = 0.009) stress-response genes. There was also an enrichment for ABC transmembrane transport (9/50, p = 82.9e−5) and cathepsin (5/22, p = 0.005) genes. We conclude that while aak-2 is absolutely required for the longevity effects of atRA, aak-2 is required for only a small proportion (~1/4) of the transcriptional response.

Given that HSF-1 and AAK-2 are both required for atRA lifespan extension, we sought to determine if HSF-1 and AAK-2 act in series or in parallel. Because a typical genetic analysis of longevity would not enable such a determination, we turned to the atRA transcriptional response (653 WaRGs) identified in wildtype animals. Compared to our datasets from aak-2(ok524) and hsf-1(sy441), 610/653 WaRGs were detectible in both mutants. We therefore analyzed those 610 genes for patterns of transcriptional response. We observed that aak-2(ok524) mutants retained a larger portion of the response, with 84.4% (515/610) of the WaRGs still being differentially expressed (FDR <0.05) in the aak-2(ok524) mutants, while only 52.6% (315/610) were in hsf-1(sy441) animals.

We next sought to categorize the atRA response overlap between mutants. We first used our normalized LFCΔ-based classification of WaRGs (maintained, enhanced, lost, weakened, or flipped) to determine the relationship between regulation in hsf-1 and aak-2 backgrounds. We find that nearly 87% (85/98) of the lost response in aak-2 was also lost in hsf-1 animals, while nearly 90% of the response retained in hsf-1(sy441) animals was retained in aak-2(ok524) animals (208/232; Supplementary file 1d). These results are inconsistent with two parallel responses where we would expect (mostly) non-overlapping classes of regulated (lost) genes. In fact, we see a significant enrichment of overlap beyond the expected overlap for random regulation between hsf-1 and aak-2. This suggests that HSF-1 and AAK-2 regulators act in series, with hsf-1 upstream of aak-2, in the atRA pathway.

We then reanalyzed the WaRGs after subsetting based on response in the mutant backgrounds. We found no enrichments at the levels used in our previous analyses for the response lost in both genetic backgrounds. In contrast, we observed enrichments for iron, amino acid, and sphingolipid metabolism among those genes whose response was maintained in both hsf-1 and aak-2 backgrounds. We interpret these changes to be either independent of the atRA longevity pathway, or upstream of hsf-1 in the atRA longevity response. A similar analysis of the 212 WaRGs that were lost in hsf-1(sy441), but retained in aak-2, showed an enrichment for fatty acid biosynthesis and UGT pathway genes. Interestingly, HSF1 has been implicated in regulating fatty acid biosynthesis in mammals (Jin et al., 2011), suggesting a potentially conserved hsf-1 function that lies upstream of aak-2 in the atRA longevity response.

Analysis of mutants in a number of key regulatory and stress-response systems treated with atRA suggests that atRA functions through the AKT-1 and AKT-2 kinases to affect conserved AMPK, Nrf2, and HSF1 pathways. Using a comprehensive RNA-seq approach with and without atRA treatment in both wildtype and mutant backgrounds suggests extensive remodeling of sphingolipid and fatty acid metabolic networks, both of which are known to modulate lifespan. While these data support a model for atRA affecting longevity through Akt and its downstream longevity transcription factors hsf-1 and skn-1, the mechanism of initiation upstream of Akt remains unknown. One explanatory model of upstream initiation is suggested by our observation that atRA transcriptionally alters sphingolipid metabolism in C. elegans, which has also been seen in mammals (Camdzic et al., 2023; Clarke et al., 2011; Kalén et al., 1992; Sun and Wang, 2021). Sphingolipids are known to regulate developmental rate and lifespan in C. elegans (Cutler et al., 2014), and potentially in mammals as well (Cutler and Mattson, 2001). For example, remodeling the sphingolipid metabolism network through RNAi-induced reductions in ttm-5 (dihydroceramide desaturase homologue), W02F12.2 (neutral/acidic ceramidase homologue), cgt-2 (glucosylceramide synthase homologue) or K06A9.1 (neutral sphingomyelinase homolog), all result in lifespan extension. Additionally, genetic disruption of the ceramide synthases alters lifespan, with loss of hyl-2 shortening and simultaneous loss of hyl-1 and lagr-1 extending C. elegans lifespan through a skn-1-dependent process (Mosbech et al., 2013). Additionally, the control of the relative ceramide and sphingomyelin levels by sphingomyelin synthases mediates crosstalk between DAF-16 and CREBH (He et al., 2021), which would have a significant impact on glucose and lipid metabolism, and therefore longevity. As such, the atRA-altered sphingolipid network observed in our RNA-seq data (Figure 5B; Figure 5—figure supplement 2) could have significant impacts on longevity.

The importance of sphingolipid regulation of metabolism, and ultimately lifespan, is not a unique feature of nematodes. The network has been proposed to function in mammals as a metabolic rheostat that uses the ratios of ceramide, ceramide-1P, sphingosine, and sphingosine-1P to determine metabolic regulatory response (Summers et al., 2019). This may be a mechanistic contributor to atRA longevity effects as there is known cross-talk between longevity pathways and ceramide/sphingolipids (Jęśko et al., 2019). Interestingly, ceramide and sphingolipid metabolism may provide a conserved functional connection to the observed relationship between atRA and protein kinase B/Akt function. In cell culture, treatment with atRA increases ceramide levels (Kalén et al., 1992), and cell-permeable ceramide inhibits Akt kinase activity (Zhou et al., 1998). Additionally, exogenous ceramide induces dephosphorylation and inhibition of Akt (Zinda et al., 2001). This functionality is known to be biologically relevant, as ceramide is a known negative regulator of insulin activity via regulation of Akt (Hsieh et al., 2014).

Considering these findings, a simple model consistent with our observations is that the application of atRA changes sphingolipid metabolism, which in turn induces a change in the functional state of akt-1 and akt-2. Our observations that HSF-1, AAK-2, and SKN-1 are necessary for atRA longevity extension are easily understood within this model, as all three have been identified as potential direct targets of Akt regulation. How (and if) atRA directly regulates sphingolipid metabolism remains an open question. In cell culture, atRA induces growth arrest in many cell types, and that arrest is mediated through nSMase2 induction, which increases ceramide levels (Clarke et al., 2011). Additionally, the involvement of sphingosine kinases in atRA signaling has been demonstrated in K562 chronic myeloid leukemia cells (Sun and Wang, 2021), but the identity of the transcriptional effector remains unclear.

The retinoids – atRA in particular – are broadly conserved regulators of transcription (Amann et al., 2011). In vertebrates, atRA functions in a broad range of biological activities, from development (Duester, 2008; Niederreither and Dollé, 2008), immune function (Huang et al., 2018), and memory and learning, to energy metabolism (Zhang et al., 2015). In mammals, some research suggests a potential role for atRA signaling in modulating aging. Among clinical aging studies of both natural and synthetic retinoids, the bulk of the research has been for aging and/or UV photoaging of skin. Among those studies, atRA is the most widely investigated retinoid and potentially the most potent (Mukherjee et al., 2006). Beyond skin phenotypes, studies in mouse models have also shown that age-dependent decreases in atRA signaling result in poor performance on spatial learning and memory tasks, and dietary supplementation with atRA can ameliorate the age-related decreases in hippocampal long-term-memory potentiation and other brain functions (Etchamendy et al., 2001). The potential use of atRA as an intervention in age-related diseases of neurophysiology has not been ignored and is receiving significant attention as a therapeutic for Alzheimer’s disease and related dementias (Das et al., 2019; Lee et al., 2009; Szutowicz et al., 2015).

While atRA may function as an anti-aging agent due to the phenotypic outcomes of application, the molecular mechanisms responsible for these activities are not fully understood. One possibility is that atRA functions as a high-affinity ligand for PPARβ/δ peroxisome proliferation-activated receptor, which is a master regulator of lipid metabolism and glucose homeostasis. Activation of PPARβ/δ increases lipid catabolism in adipose tissue and skeletal muscle to prevent obesity (Kuri-Harcuch, 1982; Pairault et al., 1988; Sato et al., 1980; Schwarz et al., 1997). Additionally, in an obese mouse model, treatment with atRA-induced PPARβ/δ and RAR regulated genes, correlating with weight loss and improved insulin responsiveness (Berry and Noy, 2009). Alternatively, atRA could be affecting longevity through effects on Akt proteins, as we observed and has been shown for other atRA phenotypes in mammals. Indeed, multiple pathways are likely to be engaged.

We tested the hypothesis that using intersecting computational predictions can identify aging interventions at a high frequency in Caenorhabditis species. We found that using cross-validated computational predictions resulted in a high discovery rate (30%), which is compatible with screening across a dosage range using full lifespan analysis. The future success of computational prediction approaches should increase as AI methodologies are brought to bear on an ever-increasing body of research. In the example described above, computational predictions led to the identification of an endogenous signaling ligand that regulates metabolism and can be co-opted to extend life. Our study demonstrates the potential of metabolic manipulation for aging interventions and the benefits of computational predictions in prioritizing a compound screening.

All Caenorhabditis strains were obtained from the Caenorhabditis Genetics Center: N2-PD1073 Teterina et al., 2022; Yoshimura et al., 2019; CF1038 daf-16(mu86) (Lin et al., 1997); DA1113 eat-2(ad1113) Raizen et al., 1995; IG10 tol-1(nr2033) Pujol et al., 2001; RB754 aak-2(ok524) C. elegans Deletion Mutant Consortium, 2012; PS3551 hsf-1(sy441) Hajdu-Cronin et al., 2004; KU25 pmk-1(km25) Mizuno et al., 2004; RB759 akt-1(ok525); and VC204 akt-2(ok393). Wild isolates in this study include C. elegans JU775 and MY16, C. briggsae AF16, ED3092, and HK104, and C. tropicalis JU1630, JU1373, and QG834. All strains were maintained on nematode growth medium plates seeded with Escherichia coli OP50-1 at 20°C. For experimental synchronization, cohorts were generated by timed egg lays (Lucanic et al., 2017).

Compound treatment was conducted as previously published (Banse et al., 2019; Lucanic et al., 2017). Compounds were obtained as a solid and dissolved in DMSO (dimethyl sulfoxide) or H₂O to obtain a stock solution. The following compounds were used: temsirolimus (Cayman 11590), ritonavir (Sigma-Aldrich SML0491), thalidomide (Calbiochem 585970), arecoline (Cayman 13662), everolimus (Cayman 11597), temsirolimus (Cayman 11590), erlotinib (Cayman 10483), berberine (Cayman 10006427), dasatinib (Cayman 11498), propranolol (Sigma-Aldrich P0884), aldosterone (Sigma-Aldrich A9477), dexamethasone (Sigma-Aldrich D1756), gefitinib (Sigma-Aldrich SML1657), tretinoin (all-trans retinoic acid) (Sigma-Aldrich PHR1187), bortezomib (Sigma-Aldrich 5043140001), decitabine (Selleck Chemical S1200), fisetin (Tocris 5016), and metoprolol (Sigma-Aldrich M5391). DMSO stock solutions were mixed with water to form a working solution before being added to plates. In-plate concentrations were calculated by presuming the final volume to be equal to that of the volume of agar. For 50, 100, and 150 µM atRA plates, stock solutions formed precipitates at working solution concentrations, requiring working solutions to be prepared individually for each plate.

Lifespan assays were performed as previously published (Banse et al., 2019; Lucanic et al., 2017). Briefly, worms were age-synchronized via timed egg lays and transferred to control or compound treated plates on days 1, 2, and 5 of adulthood (or day 4 for C. tropicalis strains), and once weekly thereafter until dead. All lifespans were conducted using 51 µM FUdR (5-fluoro-2′-deoxyuridine) to prevent progeny production (Hosono, 1978; Lucanic et al., 2017; Mitchell et al., 1979). For automated lifespan assays, worms were transferred to the Automated Lifespan Machines (Epson Perfection V800s) (Abbott et al., 2020) on day 5 (C. elegans and C. briggsae) or day 4 (C. tropicalis) of adulthood, at which point survival data was collected and analyzed using the Lifespan Machine software (https://github.com/nstroustrup/lifespan; Stroustrup, 2022; Stroustrup et al., 2013). For lifespans using RNAi feeding, RNAi plates (25 mg/l carbenicillin and 1 mM IPTG) were seeded using an RNase II deficient E. coli strain (HT115) harboring either the skn-1 (T19E7.2) targeting or control L4440 plasmid from the Ahringer RNAi library (Kamath and Ahringer, 2003). Worms were transferred to RNAi plates at the L3/L4 stage before being transferred to compound-treated RNAi plates containing 51 µM FUdR on day 1 of adulthood. An additional transfer on day 3 of adulthood was also added for lifespans using RNAi. All lifespan assays were conducted at 20°C and 80% relative humidity with 50 animals per Petri plate. Analysis of propranolol effects in the presence of paraformaldehyde-treated bacteria was performed as published (Beydoun et al., 2023).

CeleST health assays were performed as previously published (Banse et al., 2024b). In brief, animals were exposed to compound intervention during adulthood as described above until CeleST measurements were collected at two time points (adult days 6 and 12 for C. elegans and C. tropicalis, and days 8 and 16 for C. briggsae). For two biological replicates at each of the three CITP sites, 40 animals were tested per condition (age and compound or control) per strain. For full experimental protocols, see our online protocol (Caenorhabditis Intervention Testing Program, 2022). Eight different parameters (wave initiation rate, body wave number, asymmetry, stretch, curling, travel speed, brush stroke, and activity index; Ibáñez-Ventoso et al., 2016; Restif et al., 2014) were measured using the CeleST software and used to create a composite swimming score (Banse et al., 2024b).

Statistical analyses for lifespan experiments were performed as previously described (Lucanic et al., 2017). In summary, we used a mixed-model approach where compound treatment was considered a fixed effect, and other potential variables were treated as random effects. Survival was analyzed both with generalized linear models using the lme4 (version 1.1.32) package (Bates et al., 2015), and a mixed-model Cox-Proportional Hazards (CPH) model using the coxme package (version 2.2-18.1) (Therneau, 2020) in the R statistical language (R Development Core Team, 2021). The effect of compound treatment was tested using CPH analysis within each strain to allow for each compound treatment replicate to be compared to its specific control in the randomized blocks design. Compound effects were analyzed as a planned comparison between the responses of individuals on the treatment in question and those on the appropriate treatment control. Hits were classified based on a significant p-value from the CPH model coupled with an increase in median lifespan. It should be noted that one compound, aldosterone at 50 µM, showed a significant decrease in the hazard estimate without an increase in median lifespan, and thus was not considered a hit.

Swimming behavior was analyzed using the composite score (described above) as the variable of interest in mixed effects general linear models built for each strain in R using the lme4 package (version 1.1.32) (Bates et al., 2015) as previously described (Banse et al., 2024b). Determination of significant age by compound interactions was made using the R car package (version 3.1-2) (Fox and Weisberg, 2019).

For RNA-sequencing, worms were synchronized and compound treated as described above. Four biological replicates of both atRA-treated (150 µM) and vehicle control worms were aged to day 4 of adulthood and collected in tandem (approximately 50 worms total per replicate). We selected this timepoint because it corresponded to the timepoint at which the Ewald study detected increased col-144p::GFP that predicted longevity (Statzer et al., 2021). Worms were picked into 0.2 ml tubes each containing 50 µl of lysis buffer (45 µl elution buffer plus 5 µl proteinase K) and flash frozen with liquid nitrogen, then stored at –80°C until library prep. Libraries were prepared using the KAPA mRNA HyperPrep kit (KK8580 from Kapa Biosystems) as per the manufacturer’s protocol except that the total volume was adjusted to ¼ per reaction. Final libraries were normalized by concentration and sequenced on an Illumina Novaseq 6000 with the SP 100 cycle (GC3F, University of Oregon).

Paired-end FASTQ files for all atRA-treated and DMSO control samples were aligned to the C. elegans WBcel235 (build 104) reference genome using the Subread package (version 2.0.2) (Liao et al., 2013). Uniquely mapped reads were assigned to C. elegans genes with Subread’s featureCounts program (Liao et al., 2014) using reversely stranded read counting. Subsequent filtering, normalization, and differential expression analysis were performed on each strain-specific dataset with the edgeR package (version 3.28.1) (Robinson et al., 2010), using R (version 3.6.2) (R Development Core Team, 2021). Lowly expressed genes were removed from each dataset; only genes that had at least 10 reads in at least four samples and a minimum total count of 15 reads across samples were retained. To remove composition biases between libraries, the library sizes were normalized using a trimmed mean of M-values (Robinson and Oshlack, 2010) between each pair of samples. A pairwise expression analysis was performed on the transcriptomes of the treatment and control samples from each strain. Quasi-likelihood F-tests for treatment vs. control sample effect were carried out on fitted gene-wise negative binomial generalized log-linear models. p-values were corrected for false discovery using the Benjamini–Hochberg method.