Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment

Decades of progress in identifying mechanisms of aging and therapeutic targets to modulate them have revealed a tremendous overlap with central regulators of disease. Consequently, recent efforts in developing therapies to treat aging-related diseases often identify regulators of aging as promising disease targets^1,2,3. Drug repurposing efforts have highlighted the potential of treatments developed for one aging-related disease to modulate other aging-related diseases or longevity more generally, due to shared underlying pathway dysregulation coalescing around the hallmarks of aging^1,4,5,6,7. The aging field now faces a critical methodological question in translating biological insights into actionable clinical results: how best to measure the effects of an intervention on biological aging from data collected in clinical trials.

The emergence of DNA methylation-based epigenetic clocks in 2013 enabled the first attempts to quantify biological aging^8,9,10. However, inconsistent clinical trial readouts¹¹, poor cross-model agreement^12,13, challenges to interpretability¹⁴ and inherently limited and indirect mechanistic insight into biological processes driving aging limit the utility of epigenetic clocks^11,15. Proteomic aging clocks have emerged recently as a potential improvement in the design of trials for aging-related disease^16,17. Unlike epigenetic marks, proteins are the immediate effectors of biological processes, which may make proteomes more useful for assessment of both biological age and mechanistic characterization through pathway analysis.

The 2023 UK Biobank proteomic dataset release kick-started the development of several proteomic clocks but, due to their novelty, most of them have received only scarce validation in clinical settings¹⁸. In a 12-week supervised trial of exercise in 26 men, the ProtAge clock detected a 10-month reduction in biological age driven by a small subset of the 204 proteins involved primarily in PI3K–Akt and MAPK signaling^16,19. In a recent 40-month simian study of metformin featuring multi-omics profiling, a dedicated proteomic clock was used to show multi-tissue aging deceleration²⁰. Some human studies of rejuvenative interventions, such as plasma exchange, featured proteomic profiling in a supporting capacity to the primary epigenetic clock assessment of aging intensity¹³.

Here we build on those studies with a head-to-head comparison of six proteomic clocks applied to serum proteome data from a controlled phase 2a clinical trial of rentosertib—an artificial intelligence (AI)-designed TRAF2- and NCK-interacting kinase (TNIK) inhibitor^21,22. We are developing rentosertib (formerly INS018_055) for the dual purposes of targeting both aging and idiopathic pulmonary fibrosis (IPF)^{21,23,24,25,26,27,28}. IPF’s high collinearity with numerous aging hallmarks^{6,29,30,31,32,33} made it a priority indication for TNIK inhibition, originally identified as a gene involved in six aging hallmarks⁴. We are attempting to structure this drug development program around aging biology from the outset, shaping target selection, preclinical validation, indication choice and trial design, while adhering to regulatory requirements (Fig. 1a). In the phase 2a trial, we incorporated aging biology by performing longitudinal proteomic screening of serum samples for subsequent proteomic clock analysis³⁴. This proteomic profiling constitutes a rare and difficult-to-obtain human dataset on an intervention with aging-modulatory potential.

Having collected proteomic data, using it effectively required innovation in proteomic clock analyses and clinical data integration. We chose six recently developed proteomic aging clocks: ProtAge¹⁶, OrganAge trained on mortality and age data³⁵, PAC³⁶, ipfP3GPT³⁷ and PAOPAC³⁸. These clocks, available at the time of trial’s completion, represent a variety of methodologies and feature models trained to predict chronological age (ProtAge, OrganAge_chrono, ipfP3GPT, PAOPAC) or mortality risk (PAC, OrganAge_mortality), trained using classical machine learning techniques (OrganAge, PAOPAC, PAC) and deep learning (ProtAge, ipfP3GPT) and are based on a varied number of features. They thus represent different perspectives on human aging and their agreement would indicate a stronger biological signal than any single clock alone. These models were applied to all phase 2a participants who had granted their consent for this ancillary analysis and had provided biospecimen at all four timepoints.

We evaluated the degree of variance among the specified proteomic models and found that chronological aging clocks are more accurate in tracking aging rate modulation than the mortality trained clocks. Despite this variance, all clocks showed a statistically significant decrease in the pace of aging in response to rentosertib treatment. The degree to which this decrease can be attributed to modulation of aging biology rather than to rentosertib’s anti-fibrotic activity is a central challenge in this setting. We addressed this partially through pathway enrichment analyses applied to proteins affected by treatment, alongside comparison with external aging-cohort data. Full disentanglement of the two signals is not achievable within an IPF cohort and requires validating the drug or its underlying mechanism in healthy volunteers; nonetheless, our analyses yielded interpretable signatures of aging-related biology that warrant further investigation. This work reveals directions for integrating proteomic clocks into clinical trial design toward the goal of ensuring that disease-focused studies can simultaneously evaluate aging-modulatory effects.

The phase 2a trial of rentosertib in IPF (NCT05938920) is a randomized, double-blind, placebo-controlled trial conducted in 2023–2024 across 21 locations in China that recruited male and female patients older than 40 years, with a confirmed IPF diagnosis and in stable condition. Out of 128 patients screened, 71 were selected to receive 30 mg rentosertib once daily (QD, N = 18), 30 mg rentosertib twice daily (BID, N = 18), 60 mg QD (N = 18) or placebo (N = 17). By the end of trial, 16 patients discontinued treatment. According to the protocol, lung forced vital capacity (FVC) was recorded at the start and end of the trial (12 weeks). Blood samples were collected from the participants at start and end of the trial, as well as at the second and fourth weeks after the start. A total of 43 participants consented to serum proteomic screening, 1 of whom was excluded due to a missing measurement at the end of trial.

The resulting cohort consisted of 42 Asian people with a mean age of 67.1 years. Full summary statistics of the cohort studied in this article can be found in Supplementary Table 1.

Aging clock performance

We evaluated six published proteomic aging clocks on baseline serum samples from 42 participants in the phase 2a IPF trial: Argentieri 2024, Kuo 2024, Han 2026, Galkin 2025, and two variants from Goeminne 2025, one trained to predict chronological age and one trained on mortality risk^{16,35,36,37,38}. The four chronological clocks (OrganAge_chrono, ProtAge, ipfP3GPT, PAOPAC) correlated well with actual age (Spearman’s r ∈ 0.70–0.84) and, after ordinary least squares correction for systematic offset, achieved root mean square errors below 4 years (Fig. 2a). The two mortality-based clocks (PAC, OrganAge_mortality) showed weaker correlation with chronological age (r ∈ 0.16 − 0.23), consistent with their training objective and the high disease burden of our cohort, which mortality trained models are expected to capture as elevated biological age regardless of calendar age^29,39. Bland–Altman analysis confirmed that chronological clocks carried a correctable constant bias, whereas mortality clocks exhibited a proportional bias that overestimated the age of younger participants, which could not be corrected by ordinary least squares (Extended Data Fig. 1). Interclock correlations further reflected this methodological divide: chronological clocks agreed closely with one another (r ∈ 0.78 − 0.89), whereas their correlation with mortality clocks was substantially lower (r ∈ 0.25 − 0.61; Fig. 2b).

Despite their differences in calibration and training objectives, all six clocks detected proteomic shifts in response to rentosertib treatment in subsequent analyses, suggesting that the drug may modulate complementary dimensions of biological aging. Anonymized predictions for all timepoints are available in Supplementary Table 2.

Rentosertib-induced aging clock predictions

Across all six clocks, treatment arms showed consistent reduction in biological age relative to baseline, whereas the placebo group showed minimal change or slight increases over the 12-week period (Fig. 3a and Supplementary Table 3). We quantified these shifts as the change in predicted biological age from baseline (ΔBioAge) and compared each treatment arm to placebo at weeks 2, 4 and 12, yielding 54 comparisons per arm (6 clocks × 3 timepoints × 3 regimens). Of these 54 comparisons, 21 reached statistical significance (Q value < 0.10), concentrated at week 4, where 11 of 18 comparisons registered significantly lower ΔBioAge in treated participants (Fig. 3b and Supplementary Tables 4 and 5). A permutation test with patient-level label shuffling confirmed that 21 significant comparisons far exceeded chance expectation (null mean = 0.15; Extended Data Fig. 2). Similar findings persisted when the six participants with higher-grade adverse events were removed from the analysis. Among the three regimens, 30 mg BID produced the most consistent signal (nine significant comparisons), followed by 60 mg QD (seven significant comparisons) and 30 mg QD (five significant comparisons).

We also tested whether baseline body mass index (BMI) correlated with the magnitude of biological age change, reasoning that fixed dosing could yield lower effective exposure in patients with higher bodyweight. Spearman correlations between BMI and ΔBioAge were computed within each treated arm for all six clocks, yielding 54 tests total. Only three tests reached nominal significance (P < 0.05), and none remained after multiple testing correction (all Q values > 0.25), confirming that the observed biological age reductions were not modulated by BMI within the studied range.

By week 12, the number of significant comparisons decreased, indicating partial attenuation of the initial proteomic response. Direct comparison of age predictions between weeks 4 and 12 indicated that ‘plateau’ would be a more appropriate term, as no arm–clock combination showed significant (paired Wilcoxon, P < 0.05) shift in predicted age.

The two classes of clocks (chronological age and mortality trained) responded differently to dosing regimens (Fig. 3a,b). The 60 mg QD group, which showed the greatest FVC improvement in the original trial report, yielded significant biological age reduction across all four chronological clocks at week 4 (ΔBioAge ∈ −2.71 to −3.46 years, Q value < 0.10) but no significant changes in either mortality clock. In contrast, the 30 mg BID regimen was detected by both chronological and mortality clocks, making it the arm with the broadest cross-clock agreement.

Analysis of standardized effect sizes confirmed that week 4 represents the timepoint of maximal aging clock consensus, with the 30 mg BID regimen showing consistent effect sizes across five of six clocks (Extended Data Fig. 3). The twice-daily administration in the 30 mg BID group seemed to enhance aging-related effects compared to the equivalent total daily dose given once (60 mg QD), suggesting that maintaining steady drug levels throughout the day may be important for the aging process modulation. Analysis of standardized effect sizes (Cohen’s d) confirmed this pattern: at week 4, five of six clocks showed negative effect sizes for 30 mg BID, whereas ΔBioAge reductions in other regimens were detected by fewer clocks (Fig. 3c).

Organ-specific variants of OrganAge from ref. ³⁵ provided additional granularity. Although the chronological organ clocks detected no significant (Q value < 0.10) ΔBioAge shifts (Extended Data Fig. 4), several mortality-based organ clocks did. In particular, all timepoints across all treated arms showed significantly lower predicted age relative to placebo with the artery clock (ΔBioAge ∈ −6.95 to −16.57 years; Extended Data Fig. 5). The stomach, brain, pancreas and immune clocks also registered significant reductions in select arms.

These exploratory findings suggest that rentosertib treatment is associated with a broad proteomic shift toward younger predicted biological age profiles, with the 30 mg BID regimen producing the most consistent signal across methodologically diverse clocks. The temporal pattern of peak effects followed by a plateau warrants further investigation to determine whether it reflects pharmacodynamic adaptation, a new homeostatic equilibrium or a limitation of the 12-week observation window.

Rentosertib-induced proteomic trajectories

To identify proteins whose serum levels changed in response to treatment, we fitted linear mixed-effects models (LMEM) for each of the 2,841 measured proteins, with sex and BMI included as fixed-effect covariates alongside group-by-time interaction terms to capture trajectories that diverged from placebo over the 12-week trial period.

Rentosertib altered the expression trajectories of 326 proteins significantly (Q value < 0.10) across all treatment groups, compared to only 2 in the placebo group (Fig. 4a and Supplementary Table 6). Most changes were regimen-specific: 237 proteins responded only in one arm. The 30 mg BID group exhibited the broadest proteomic response, with 142 uniquely affected proteins, consistent with its strong cross-clock aging signal. The 30 mg QD regimen, by contrast, produced only one significant time-dependent change (IDO1, upregulated). A subset of 89 proteins showed concordant directional changes in two or more treatment groups, suggesting a core set of pathways affected by TNIK inhibition regardless of regimen (Fig. 4a).

Individual protein trajectories demonstrated the nature of these changes (Fig. 4b). Among the most prominently downregulated proteins were drivers of fibrosis and extracellular matrix (ECM) remodeling, including COL1A1, MMP10 and FAP. Conversely, several proteins involved in cellular metabolism and stress resistance, including NAMPT (the rate-limiting enzyme in NAD⁺ biosynthesis), the antioxidant SOD2 and the detoxification enzyme ALDH1A1. Among the decreased proteins, we also observed reduced levels of PDGFB—a growth factor involved in proliferative signaling. A heatmap of the 90 most affected proteins confirmed the cross-arm temporal patterns, with the 30 mg BID and 60 mg QD regimens producing more pronounced and sustained shifts than 30 mg QD (Extended Data Fig. 6).

Stability of rentosertib’s effects

The attenuation of the aging clock signal between weeks 4 and 12 warranted exploration of whether rentosertib’s effect genuinely fades or whether this pattern reflected a limitation of our aging clock analysis.

To address this, we recalculated the LMEM with time as a categorical variable to capture nonlinear proteomic trajectories, classifying each protein’s response as sustained, delayed or transient (Supplementary Table 7 and Methods). In the 30 mg BID and 60 mg QD arms, only 5–9% of proteomic shifts were transient (Extended Data Fig. 7a). This confirms that the proteomic response to rentosertib at effective doses continues to develop through week 12 rather than fades. But among the proteins used in four aging clocks with publicly available feature weights (OrganAge_chrono, OrganAge_mortality, PAC, ProtAge) proteins with a sustained trajectory are overrepresented with odds ratio = 4.65 in 30 mg BID, and 2.19 in 60 mg QD groups (Supplementary Table 8 and Extended Data Fig. 7b). The proteomic changes most relevant to biological aging are thus among the more durable treatment effects.

This finding leads to a paradoxical conclusion: sustained modulation of aging-relevant proteins should maintain or deepen the signal, yet all clocks register a plateau. To resolve this matter, we inspected the key contributors to biological age shifts from baseline to week 4 and week 4 to week 12. We identified a list of 35 unique proteins, whose normalized protein expression (NPX) shift contributed >3% of the total BioAge shift at either timeframe-clock combination. At both timeframes, the most important feature was LTBP2—a master regulator of fibrosis⁴⁰ in the TGF-β signaling axis, which is coincidentally the only important feature present in all six clocks (Supplementary Table 9). Other important features spanned inflammaging (AGER⁴¹), neurodegeneration (NEFL⁴²), reproductive aging (FSHB^43,44) and cellular senescence (CXCL9⁴⁵), as well as a range of fibrotic markers (SPP1, KRT19⁴⁰, ELN) and other ECM proteins (COL6A3, MMP12). In most cases, the highlighted processes cannot be separated clearly and the specified genes may be attributed to several categories at once.

This heterogeneity and the oversized influence of fibrotic proteins on the aggregate clock signal presents a challenge in distinguishing between the drug’s therapeutic and potential geroprotective effects. Resolving this entanglement requires comparison with other antifibrotics acting through distinct mechanisms, such as pirfenidone or nintedanib, for which comparable proteomic data is not currently available. However, we were able to address this important matter through several complementary analyses presented in the next section.

Isolating geroprotective effects

On their own, proteomic clocks cannot answer the question of whether the biological age reductions reflect general aging modulation per se or are secondary to rentosertib’s anti-fibrotic activity. An indirect indication, however, can be obtained from inspecting the degree of dissociation between clinical efficacy and aging clock responses. If clock changes were merely downstream of reduced disease burden, the regimen with the greatest respiratory improvement should also show the strongest aging reversal. Instead, the opposite pattern emerged: the 60 mg QD group, which produced the largest FVC gains in the original trial report, showed less consistent ΔBioAge than the 30 mg BID group. This dissociation was supported further by regression analysis: ΔFVC explained minimal variance in ΔBioAge across all six clocks (median R²= 0.06, range ∈ 0.01–0.18, N = 42). Although FVC is an incomplete proxy for overall disease status, the consistently low explanatory power across clocks argues against a simple disease improvement explanation.

To test whether rentosertib’s proteomic effects oppose normal aging trajectories independently of its anti-fibrotic action, we compared treatment-induced protein changes to age-associated protein changes in 55,319 older adults from the United Kingdom (UK) Biobank. Of 2,832 proteins shared between the two datasets, 758 showed significant age-dependent expression in UK Biobank (Q value < 0.05, |coefficient | > 0.005 NPX per year). Rentosertib-modulated proteins were enriched 1.74-fold for age-associated proteins relative to the background rate (P < 0.001). For each treatment arm, we then correlated the direction of treatment-induced changes with the direction of normal aging. The 30 mg BID regimen showed a significant negative correlation (Spearman’s r = −0.30, P < 0.01), indicating that this regimen preferentially reversed age-associated proteomic trajectories. The 60 mg QD regimen showed no such correlation (r = –0.097, P = 0.37), despite its superior FVC response. In the placebo group, proteomic changes tracked the direction of normal aging, consistent with disease progression accelerating age-related proteomic decline. This pattern supports the hypothesis that rentosertib may alter aging processes independently of fibrotic processes. However, we acknowledge the limitations of such computational and indirect evidence and intend to seek direct experimental validation.

Protein–protein interaction (PPI) analysis provided further insight into how the two regimens diverge at the pathway level. The 30 mg QD group was excluded from this analysis due to insufficient significantly affected proteins (Fig. 4a). In the remaining arms, the 60 mg QD group affected 184 proteins (95 unique) and the 30 mg BID group affected 231 proteins (142 unique). Both regimens modulated ECM remodeling and immune function clusters (Supplementary File 1), but their unique protein sets differed in character. Proteins affected exclusively by 30 mg BID were enriched for metabolic pathways, including pentose phosphate, glutathione and cholesterol metabolism (Extended Data Fig. 8a), whereas those unique to 60 mg QD were enriched for Wnt signaling and broader immune activities (Extended Data Fig. 8b). This pathway-level divergence offers a potential mechanistic explanation for why the two regimens differ in their aging clock profiles despite comparable total daily doses.

Together, these analyses suggest that rentosertib’s effects on biological age cannot be attributed fully to its anti-fibrotic activity. Definitive confirmation, however, will require studies in non-IPF populations where severe fibrosis cannot confound interpretation, such as in other patients with aging-related diseases or generally healthy long-living people.

Rentosertib’s senomorphic effect

As an independent line of evidence for rentosertib’s potential to suppress cellular senescence at the proteomic level, we performed geneset enrichment analysis (GSEA) using three established senescence signatures: SenMayo (91 proteins measured), CellAge-upregulated (93 proteins known to promote senescence) and CellAge-downregulated (116 proteins known to inhibit senescence). Proteins were ranked by the strength and direction of their treatment response in each arm, derived from LMEM time coefficients (Supplementary Tables 10–13 and Methods).

Placebo and treated groups showed opposite senescence trajectories (Fig. 5a,b). In the placebo group, SenMayo proteins were enriched significantly among upregulated proteins (normalized enrichment score (NES) = 1.48; Q value < 0.01), with the leading edge (biggest contributors to NES) comprising ECM remodelers (MMP1, MMP3, MMP9, MMP10, MMP13), interleukins (ILs) (IL1B, IL6, IL10, IL13), growth factors (FGF7, IGFBP4, EGF, GDF15) and chemokines (CCL20, CXCL1, CXCL3). All treated groups showed the reverse pattern: significantly negative SenMayo scores and significantly positive CellAge-downregulated scores (Q value < 0.05), except for 60 mg QD in SenMayo (Q value = 0.17; Supplementary Table 14). Seven proteins appeared in the SenMayo leading edge of every treated arm: EREG, ESM1, IGFBP4, ITGA2, MMP10, MMP13 and SPP1 (Fig. 5c). In the CellAge-downregulated set, FOS and DPY30 showed discordant dynamics between placebo and all treated groups (Fig. 5d). These circulating protein patterns are consistent with a senomorphic effect of rentosertib, previously shown in vitro⁴⁶.

We next examined whether rentosertib modulates broader signaling pathways implicated in aging. GSEA against the Kyoto Encyclopedia of Genes and Genomes (KEGG) Medicus collection (658 genesets, filtered to those with >25 proteins measurable by Olink) identified five significantly enriched pathways, all involving growth factor signaling: RTK–PLCG–ITPR, and pathways involving MAPKs, RAS, ERK and PI3K–Akt (Fig. 5e,f). These pathways were upregulated in placebo but downregulated in all treated arms (Q value < 0.25). Growth factor and nutrient-sensing pathways are well-established drivers of aging, targeted by caloric restriction and rapamycin, so their coordinated suppression by rentosertib provides a plausible mechanistic link to the observed reduction of biological age. At a more stringent threshold (Q value < 0.10), the 30 mg BID arm was distinctive in that it did not upregulate the cytokine–Janus kinase-signal transducer and activator of transcription (JAK–STAT) axis significantly, whereas both 30 mg QD (Q value < 0.05) and 60 mg QD (Q value < 0.10) did, with IL2, IL3, IL4 and IL12 subunits forming the leading edge. As JAK–STAT-mediated cytokine signaling promotes fibrosis through STAT activation⁴⁷, the absence of this signal in the 30 mg BID arm may partly explain its more consistent aging clock response.

GSEA against the larger Reactome collection (1,787 genesets, 145 with >25 proteins in our data) identified 64 significantly enriched pathways across all arms at Q value < 0.25, with 36 remaining at Q value < 0.10 (Fig. 6 and Supplementary Table 14). All pathways showed concordant regulation: no pathway had opposite enrichment signs between two treated arms, and no pathway shared a sign between any treated arm and placebo. The 30 mg BID group again showed the broadest response, with 40 enriched pathways. ECM remodeling pathways were downregulated in the 30 mg BID and 60 mg QD groups, as expected from rentosertib’s anti-fibrotic activity. Several additional pathways with no direct ECM connection were also affected in both arms: MET signaling (Q value < 0.10, NES ∈ −1.64 to −1.63), Gαs signaling (Q value < 0.10, NES ∈ −1.65 to −1.56), death receptor signaling (Q value < 0.25, NES ∈ −1.64 to −1.39) and regulation of insulin-like growth factor (IGF) transport and uptake by IGF binding proteins (IGFBPs) (Q value < 0.10, NES ∈ −1.70 to −1.66). In the 30 mg BID group specifically, transcriptional regulation by TP53 (Q value = 0.17, NES = 1.40) and fatty acid metabolism (Q value = 0.09, NES = 1.45) were upregulated (Fig. 6).

The IGFBP pathway warrants particular attention given the appearance of IGFBP4 in the SenMayo leading edge and the established role of IGF signaling in aging. The shared leading edge for the 30 mg BID and 60 mg QD arms contained 15 proteins, including the senescence- and fibrosis-associated proteins CCN1, FN1, IGFBP3, IGFBP4 and SPP1 (Extended Data Fig. 9a and Supplementary Table 15). Among these, SPP1 and FN1 showed significant downward trends in both arms (Q value < 0.10), whereas CCN1 reached significance only in the 60 mg QD group (Q value < 0.05). To capture nonlinear shifts that time interaction coefficients may miss, we assessed baseline-to-timepoint changes directly for all available IGFBPs, IGF1R and IGF2R (ten proteins total; Extended Data Figs. 9b and 10). Six proteins were downregulated significantly relative to baseline across all treated groups: IGFBP-1, IGFBP-4, IGFBP-6 and IGFBP-7, IGFBPL1 and CCN1, with CCN1, IGFBP4 and IGFBPL1 showing the most consistent reductions across all three regimens. These coordinated shifts in IGF-axis proteins suggest a potential mechanism linking TNIK inhibition to the metabolic and senescence-related components of biological aging.

This study advances the integration of proteomic clocks into clinical studies of targeted disease-modulating drugs with potential geroprotector activity. The rentosertib phase 2a trial was designed to accommodate several aging biology assessments based on serum proteome profiling: longitudinal multi-clock aging rate monitoring, cross-referencing with UK Biobank normal aging trajectories, PPI analysis and GSEA with the emphasis on aging hallmark processes. Previous geroprotector trials studied long-approved interventions and measured biomarkers of aging as endpoints^48,49. In contrast, we propose to accelerate geroprotector discovery by applying aging research methods to new targets and agents from the beginning of the drug discovery and development pipeline. The evidence described here provides an early indication that proteomic aging clock analyses can detect aging signatures in trial participants.

Proteomic models are the newest development in the field of aging clocks. They offer more direct and interpretable mechanistic insights than epigenetic models, as proteins are the immediate effectors of biological change. This emerging technology is, however, not as widely adopted and refined as epigenetic models. To our knowledge, no previous study has systematically compared several proteomic aging clocks in a clinical intervention. The six recently released models we selected were trained originally with UK Biobank data composed of mostly healthy people, yet all offer relatively consistent results in the out-of-scope case of the rentosertib trial featuring exclusively IPF patients. All chronological clocks achieved strong correlations with actual age (Spearman’s r > 0.77) and low error after linear correction (root mean square error < 4 years), confirming robust aging signal detection despite the disease state (Fig. 3). Meanwhile, mortality-based clocks showed lower age correlations (r ∈ 0.16–0.23) but higher predicted ages (Fig. 2b), consistent with their training objective. Although clocks trained to predict chronological age were more accurate than mortality trained clocks (Fig. 3a), all models detected treatment-associated age reduction. The convergence of signals from different models argues against clock-specific artifacts, but whether this shared signal reflects modulation of aging biology, the drug’s anti-fibrotic activity, or both remains to be determined.

We approached this question by examining whether the proteins driving predicted biological age shifts and effects on fibrosis can be dissociated. The regimen showing the most consistent aging clock responses, 30 mg BID, uniquely modulated metabolic pathways including cholesterol metabolism, pentose phosphate pathway and glutathione metabolism (Fig. 6b and Supplementary File 1). These redox homeostasis and metabolic remodeling pathways represent established aging mechanisms and are not primary features of the anti-fibrotic response. In contrast, the 60 mg QD regimen showed greater enrichment for Wnt signaling and immune pathways linked more directly to fibrotic processes and target engagement. Reactome GSEA revealed notable downregulation of proteins involved in IGF transport and uptake, such as various IGFBPs, which are known to regulate IGF1 and IGF2 interactions with their receptors^50,51, supporting our hypothesis that rentosertib affects IGF signaling, consistent with evidence that TNIK inhibition prevents diet-induced obesity in mice²⁸. This coordinated regulation provides a mechanistic link between TNIK inhibition, senescence attenuation and aging clock changes, which is distinct from Wnt-mediated fibrosis suppression. Intra-clock comparisons across several timepoints can serve as additional indications of mechanistic dissociation. In our case, rentosertib-induced peak biological age reduction at week 4 across several clocks followed by a plateau, even though individual fibrosis-associated genes exhibited more progressive expression changes²².

We conclude that, although aging clocks are a valuable tool for detecting broad shifts in biological age, their outputs are best interpreted in conjunction with additional analyses to resolve potential artifacts (for example, other multi-omic aging-cohort studies and pathway-level GSEA analyses). Conclusively establishing the functional differences between the 30 mg BID and the 60 mg QD regimens will require more direct experimental procedures, such as longitudinal tissue sampling where ethically feasible, followed by single-cell and/or spatial transcriptomic, proteomic, epigenomic and histopathologic profiling; direct quantification of senescent-cell burden and senescence-associated secretory phenotype markers in affected and peripheral tissues; orthogonal multi-omic clocks and prospective studies with prespecified pharmacodynamic biomarkers linked to functional outcomes. Proteomic aging clocks still carry the caveats common among other omics-based clocks. Establishing causality remains a challenge, as the statistical trends they capture may be associated with both compensatory and harmful aging processes. Thus, the rebound or a plateau in biological age at later timepoints does not necessarily imply the arrest of beneficial anti-aging changes, but may reveal a statistical artifact of context-specific protein dynamics. Future trials featuring multi-modal clocks may provide more robust conclusions. Yet, even within the purely proteomic setting of this study, complementary analyses let us probe how rentosertib affects aging and senescence drivers in humans.

Our senescence marker analysis builds on previous in vitro demonstrations of the ability of rentosertib to prevent senescent-cell formation in IMR-90 cells⁴⁶. GSEA revealed opposing trajectories between placebo and treated groups: placebo patients showed strong SenMayo signature upregulation, aligning with the known role of senescence in IPF progression⁵², whereas all treatment arms demonstrated major downregulation of this signature. Likewise, the CellAge_down gene signature, indicative of senescence reversal, was upregulated across treatment groups. At the individual protein level, senescence markers such as EREG, IGFBP4, MMP10, MMP13 and SPP1 were consistently downregulated across all treated groups. Although the GSEA significance cutoff (Q value < 0.25) reflects the standard protocol for geneset permutation analysis, many findings in individual timepoints and arms were detectable at more stringent statistical levels (Fig. 5a,b).

Although the results obtained could not be compared to epigenetic clocks, we see several factors presenting proteomic clocks as a superior option in our setting. First, the features of proteomic clocks are interpretable immediately and allow for a critical inspection of predictions, unlike CpG methylation marks whose impacts on biological processes are much less direct. Second, methylation profiles derived from blood samples reflect the responses to stimuli by hematopoietic cell lineages, whereas circulating proteins may better capture systemic response. However, a counterpoint may be presented in the scope of the features each technology produces: modern epigenetic arrays report methylation levels at nearly a million genomic sites, whereas proteomic arrays report only under 3,000 proteins. This technological constraint makes some statistical tests incomplete or impossible, as was the case for our GSEA when enrichment scores could be obtained only for 145 out of 1,787 available Reactome pathways due to low intersection between array-pathway features. TNIK itself is not part of the Olink 3072 platform used in this work, allowing only for indirect (PPI, GSEA) validations of target engagement.

Beyond their interpretive scope, proteomic clocks may allow optimization of therapeutic regimens beyond traditional efficacy metrics. Higher drug exposures may engage different sets of downstream effectors, especially as targets become saturated and more free drug is available⁵³. Subsaturation and supersaturation drug concentrations could therefore induce drastically different responses across tissues based on the expression patterns of targets, transporters and metabolizers^{54,55,56,57,58,59,60}. With a relatively short half-life of 7–11 h, individual doses of rentosertib result in large peak-to-trough fluctuations²⁵. As shown in our previously published rentosertib clinical trial report, 60 mg QD dosing achieves 2- to 3-fold higher maximum plasma concentrations (C_max) than either 30 mg regimen²². In light of our aging clock analysis identifying much stronger anti-aging signatures in the lower but more frequent 30 mg BID regimen compared to 60 mg QD, we hypothesize that more frequent, low dose and less-frequent, high-dose surges may interact with downstream kinase cascades differently, eliciting distinct proteomic responses despite equal exposure, ultimately leading to differences not just in circulating proteomes but in the clinical responses observed²².

The observed week 4 plateau in the aging clock signal provides additional hypotheses for future testing. This temporal pattern may emerge from compensatory signaling pathways counteracting sustained TNIK inhibition analogous to cancer resistance mechanisms⁶¹, senescent-cell heterogeneity exhausting readily accessible rejuvenation capacity or cellular stress partially offsetting rejuvenative benefits at later timepoints. These possibilities all suggest that optimal geroprotective effects may require intermittent dosing schedules^62,63,64. The patterns we observe in several clock- and aging-associated expression trajectories with regimen and time dependence may thus inform therapy optimization for different stages of IPF, in other indications, or even in general geroprotection trials.

Within current Unites States Food and Drug Administration (FDA) frameworks, such studies would be conducted most realistically not as trials for an ‘aging’ indication, but either as exploratory pharmacodynamic analyses embedded in trials for recognized age-related diseases or risk-defined conditions, or as early-phase studies in non-IPF older populations when the safety margin and risk–benefit profile are appropriate. Aging clock or senescence endpoints could be prespecified as exploratory biomarkers, with clinical claims requiring conventional clinical endpoints or a qualified biomarker under the FDA Biomarker Qualification Program and FDA–National Institutes of Health (NIH) BEST framework⁶⁵. A relevant precedent is the TAME trial concept, which proposed testing metformin against a composite of incident age-related diseases rather than aging per se. Another notable example is the evolution of obesity and/or overweight drug development, including FDA approval of semaglutide and tirzepatide for cardiovascular risk reduction⁶⁶. Both these precedents show how a biology-linked risk state can become a tractable therapeutic setting. We therefore propose a stepwise framework: first, collect aging and senescence biomarkers prospectively as exploratory endpoints in disease trials and explore the different dose and administration regimens; second, replicate effects in non-IPF age-related or risk-enriched populations where IPF disease modification is not the primary driver and third, pursue biomarker qualification or composite clinical endpoints to establish evidence of geroprotection.

In this light, our work embodies the first stage of this framework. Its progression and scaling to other clinical programs, however, is associated with several logistical and institutional hurdles. First, identifying a new drug target such as TNIK involved probing disease–gene relationships with very limited data. Multi-model AI-powered target discovery tools such as PandaOmics incorporate models effective in low-data regimes, such as random-walk on heterogeneous graph representations of data derived from scientific literature⁶⁷, and enable rapid, iterative exploration of potential targets. Second, new drug targets require substantial preclinical validation to define the mechanisms of action before human administration, which is complicated in the cases of complex regulatory hub proteins such as TNIK⁶⁸. Third, incorporating exploratory aging clock and proteomic analyses into clinical trial protocols required careful consideration of informed consent and regulatory oversight practices for these nascent technologies. Regulators rightfully expect that every analysis, or usage of patient data, gains knowledge about the disease or drug mechanisms. It is therefore even more important that the aging biology under study overlaps the disease pathology.

Meeting these challenges enabled a rare look, however limited, into the aging-associated biological mechanisms underlying the effect of rentosertib and, therefore, implications for its activity in the context of human aging. We used the trial design both to test whether proteomic aging clocks can offer stable, reliable results in clinical trials and to gain mechanistic insights into TNIK inhibition inaccessible by standard clinical endpoints. Although combined proteomic aging clock and senescence marker analysis revealed potential geroprotective effects, the modest sample size, relatively short duration, dominance of computational approaches and lack of complementary omics modalities precluded clear deconvolution of anti-fibrotic and anti-aging effects. However, our findings of treatment-associated senomorphic signatures, predicted biological age reductions and metabolic pathway modulation warrant further investigation. More generally, the results presented here support our proposed strategy for dual-purpose clinical trial designs that include simultaneous evaluation of geroprotective effects in studies of new agents for specific aging-related indications. As exemplified by rapamycin and metformin, clinically approved drugs that are now actively studied as candidate geroprotectors, the transition from use in conventional disease indications to rigorous evaluation for geroprotective activity can take decades^69,70. In contrast, our strategy could enable early detection of anti-aging activity and reveal repurposing opportunities that would otherwise be missed.

Trial design and participants

Trial design

This analysis used data from a phase 2a, multicenter, double-blind, placebo-controlled, randomized, multidose trial of rentosertib in adults with IPF (NCT05938920). The trial was conducted at 22 sites across China from 19 July 2023, through 11 June 2024. The following ethics committees and review boards at recruiting institutions approved this trial: The Drug Clinical Trials Ethic Review Committee of Peking Union Medical College Hospital; The Medical Ethic Committee of Sichuan University; The Dug Research Ethic Committee of The Second Hospital of Anhui Medical University; The Ethic Committee of Shanghai Chest Hospital; The Ethic Committee of Shengjing Hospital of China Medical University; The Drug, Device and New Medical Technology Ethic Committee of The First Affiliated Hospital, Sun Yat-sen University; The Ethic Committee of Tianjin Medical University General Hospital; The Ethic Committee of Shanghai Pulmonary Hospital; The Medical Ethic Committee of Jiangxi Provincial People’s Hospital; The Life Ethic Committee of Beijing Friendship Hospital, Capital Medical University; The Medical Ethic Committee of Nanfang Hospital of Southern Medical University; The Scientific Research and Clinical Trials Ethic Committee of First Affiliated Hospital of Zhengzhou University; The Medical Ethic Committee of Xiangya Hospital, Central South University; The Registration Popurse Clinical Trials Ethic Committee of The First Affiliated Hospital - Zhejiang University School of Medicine; The Medical Ethic Committee of Zhongshan Hospital Fudan University; The Medical Ethic Committee of Hainan General Hospital; The Medical Ethic Committee of Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School; The Medical Ethic Committee of Peking University Shougang Hospital; The Clinical Trials Ethic Committee of Anhui Medical University - Anhui Chest Hospital; The Clinical Trials Medical Ethic Committee of Shanxi Academy of Medical Sciences - Shanxi Bethune Hospital (Shanxi Dayi Hospital); The Drug Clinical Trials Medical Ethic Committee of Qilu Hospital of Shandong University; and The Medical Ethic Committee of The Second Xiangya Hospital of Central South University.

Adults with IPF were assigned randomly in a 1:1:1:1 ratio using interactive response technology to receive oral rentosertib at doses of 30 mg QD, 30 mg BID, 60 mg QD or placebo (QD) for 12 weeks, along with continued standard-of-care medications.

The trial was conducted following the Declaration of Helsinki and International Council for Harmonization Good Clinical Practice guidelines. Institutional review boards at participating centers approved the protocol, and all patients provided written informed consent. Trial management and data processing were conducted by Fortrea Clinical Pharmacology Services (Leeds, UK). All participants, investigators and analysts remained blinded to treatment assignments until data freeze. See full details on trial design in the original trial report²².

Study participants

An informed consent form for the proteomic assessment was offered to all 55 participants who completed the trial after its end. Among them, 43 signed the form in person and agreed to have their data included in the proteomic substudy. One participant was excluded from this study due to missing proteomic measurements at week 12 (end of trial). Full summary statistics (including sex, age, BMI and adverse events) of the cohort are available in Supplementary Table 1.

Data preparation

Sample collection and processing

Serum samples were collected from all participants at baseline, week 2, week 4 and week 12 (end of treatment). Samples were processed according to standardized protocols and stored until analysis. Protein expression levels were measured using the Olink Explore 3072 panel following the manufacturer’s standard procedures.

Olink data preprocessing

The Olink Explore system generates raw count data, which are converted to NPX values through a two-step process featuring (1) extension normalization and (2) intensity normalization. In (1), for each assay i and sample j, assay counts are divided by the extension control value from the same sample, followed by a log₂ transformation:

In (2), for each assay on each plate, intensity normalization is performed by subtracting the median extension NPX value of all samples (excluding control strips) on that plate:

No bridging or anchoring controls were applied. After removing assays annotated as ‘EXCLUDED’ according to Olink quality control, 2,841 proteins remained for further analysis.

Aging clock and GSEA details

Proteomic aging clocks

Six published proteomic aging clocks were applied: ProtAge¹⁶; ipfP3GPT³⁷; OrganAge_chrono and OrganAge_mortality³⁵; PAOPAC³⁸ and PAC³⁶. Clock implementations were standardized using a unified Python library (https://github.com/Insilico-org/proteoclock). For OrganAge, we used two conventional variants trained to assess either mortality risk or chronological age, referred to as OrganAge_mortality and OrganAge_chrono in the main text. For ProtAge, protein expression was renormalized using the original normalization procedure, with minimum–maximum scaling (0–1) using trained scikit-learn MinMaxScaler objects from UK Biobank reference data, then centered on population medians. ProtAge’s interface was accessed at https://github.com/miargentieri/proteomic-age-ukb with model weights accessed through direct collaboration. PAOPAC was accessed at https://github.com/JackieHanLab/PAOPAC. For the organ-specific aging clocks in ref. ³⁵, proteins were standardized by subtracting the mean and dividing by the s.d., before multiplying with the s.d. values given in Supplementary Table 3 in ref. ³⁵. The output from the mortality-based models was converted to years following the instructions on https://github.com/ludgergoeminne/organAging.

Time-dependent treatment effects

LMEMs were fitted for each protein using the statsmodels Python package to analyze temporal changes across treatment groups. The model specification was:

where treatment groups are indicator variables (placebo as reference), Time represents weeks (0, 2, 4 and 12) and interaction terms capture differential protein expression changes relative to placebo. Patient-specific random intercepts account for individual baseline differences. Restricted maximum likelihood estimation was used for model fitting. We used the Group × Time interaction coefficients as the primary measure of the effect of rentosertib on protein expression. Coefficients passing the Q value < 0.10 threshold after Benjamin–Hochberg correction were deemed significant.

For part of the results a modified LMEM model was used, in which the Time and Group × Time terms were replaced with categorical terms, treating each observation timepoint as a separate variable. For these models, the following classification of protein trajectories applies: sustained—significant effect at week 12 and earlier timepoints; transient—significant at week 2 and/or week 4 but not at week 12; delayed—significant only at week 12.

Geneset enrichment analysis

GSEA was performed using treatment effect coefficients (Group × Time interactions) multiplied by −log₁₀ (P value) as preranked gene lists. All weights were also multiplied by 100 for visualization purposes.

The SenMayo geneset (124 senescence-associated proteins in total, 91 measured in trial) from MSigDB was used to assess senescence pathway modulation^71,72,73. Furthermore, we used the CellAge database of genes affecting senescence^74,75. All genes annotated as inducing senescence (N = 370, of which 93 measured) formed the CELLAGE_UP geneset. All genes annotated as inhibiting senescence (N = 475, of which 116 measured) formed the CELLAGE_DOWN geneset.

For KEGG MEDICUS⁷⁶ GSEA we used the c2.cp.kegg_medicus.v2025.1.Hs.symbols.gmt obtained from MSigDB^77,78. For Reactome⁷⁹ GSEA we used c2.cp.reactome.v2025.1.Hs.symbols.gmt obtained from MSigDB.

GSEA tests and statistical measures were obtained with gseapy for Python⁸⁰. Genesets containing <25 proteins shared with the preranked lists were excluded, Q values were calculated using 1,000 permutations for each geneset. Custom visualizations were prepared in Plotly based on gseapy’s output.

Statistics and reproducibility

Statistical significance was set at Q value < 0.10 for most statistical tests, with Q value < 0.25 used as a threshold for GSEA, as suggested by the standard GSEA protocol⁸¹ unless stated otherwise.

Treatment effects on biological age were analyzed using Mann–Whitney U tests for pairwise comparisons (placebo versus treatment) at each timepoint with Benjamin–Hochberg multiple correction procedure.

Differential protein expression of IGFBP-related proteins was analyzed with paired t-tests using the ttest_rel from the scipy.stats module for Python v3.11. Benjamin–Hochberg multiple correction procedure was applied to obtain Q values.

For senescence marker trajectories, proteins showing significant (P < 0.10) decreases in both ‘60 mg QD × Time’ and ‘30 mg BID × Time’ coefficients relative to placebo were identified as characteristic anti-aging trajectories.

All statistical tests were two-sided unless otherwise specified. No data were excluded from the analyses, except for the tests in which we tested the robustness of the key results in the subcohort without higher-grade adverse events.

Cohen’s d

Standardized effect sizes were calculated using Cohen’s d to quantify the magnitude of biological age changes between treatment and placebo groups. For each aging clock at weeks 4 and 12, Cohen’s d was computed as:

where μ represents the mean change in biological age from baseline (ΔBioAge) and S_p is the pooled s.d. calculated as:

The 95% confidence intervals (CIs) for Cohen’s d were estimated using the variance approximation method⁸²:

Negative values Cohen’s d indicates biological age reduction relative to placebo.

UKB analytics

This study features validation using 55,319 proteomic samples from the UK Biobank. This cohort consists of 54% female and 46% male participants with a median age of 71 years (range ∈ 50–85 years).

We defined normal aging protein trajectories in linear models as:

The list of 758 significant proteins associated with aging was aggregated based on the significance (Q value < 0.05) of the age coefficient and its absolute value filter (|coef| > 0.005 NPX per year). The magnitude threshold excluded proteins with trivial age effects, ensuring biological relevance while maintaining statistical significance.

To assess whether rentosertib-affected proteins are associated preferentially with aging versus disease improvement, we performed hypergeometric tests comparing protein set overlaps. The background set consisted of 2,832 proteins shared by both UK Biobank and phase 2a trial datasets, representing all the proteins where both age-association and treatment effects could be assessed. For each comparison, the overlap size (k), rentosertib-affected set size (n = 325), reference set size (N) and background size (M = 2,832) were used to calculate P values and odds ratios:

P value = 1 − hypergeometric CDF(k − 1; M, n, N)

Separate enrichment analyses were performed for age-associated proteins (N = 758, Q value < 0.05, |coef | ≥ 0.005) at both aggregate and arm-specific levels (60 mg QD: n = 183; 30 mg BID: n = 231). Analysis for 30 mg QD was not carried out due to a low number of n = 1.

To test whether rentosertib treatment reverses age-associated protein trajectories, we calculated Spearman rank correlations between treatment-induced protein changes (Group × Time coefficients from LMEM) and age-associated changes (age coefficients from UK Biobank). For each trial arm, proteins affected significantly by treatment (Q value < 0. 05) were matched to the list of age-associated proteins (Q value < 0.05, |coef | > 0.005), resulting in arm-specific datasets (30 mg BID: n = 72, 60 mg QD: n = 88). For the placebo group, the Q value filter was not applied, due to a low number of n = 2. In this setting, negative correlations indicate reversal of aging trajectories, whereas positive correlations indicate concordance with aging.

Explainable variance

To quantify the contribution of dFVC to biological age shifts, we fitted OLS regression models (ΔBioAge ~ dFVC) for each aging clock using all participants data (N = 42) with paired baseline and week 12 measurements. The variance explained (R²) was calculated to assess how much ΔBioAge could be attributed to respiratory function improvement.

PPI network analysis

To characterize pathway specificity of rentosertib-affected proteins, we performed Markov network clustering using the StringDB (v12.0) web interface⁸³.

Proteins affected significantly by rentosertib treatment (Q value < 0.05 in LMEM analysis) were submitted to the web service of StringDB for full interaction network construction with a confidence threshold set to 0.4. The 30 mg QD group was excluded from detailed pathway analysis due to insufficient significant proteins (N = 1).

Markov network clustering analysis was carried out with the inflation parameter set to 3 to produce arm-specific networks: 60 mg QD (93 unique proteins forming 18 clusters, 8 with ≥3 proteins) and 30 mg BID (141 unique proteins forming 34 clusters, 13 with ≥3 proteins). The cluster descriptions and components are available in Supplementary File 1. Clusters were characterized by their predominant biological processes based on the integrated annotations of StringDB. Cluste visualizations were prepared using networkx for Python.

Software and reproducibility

All analyses were performed using Python v3.11. The aging clock predictions for ipfP3GPT, OrganAge and PAC clocks may be obtained with the proteoclock v1.0.0 library (https://github.com/Insilico-org/proteoclock). LMEM were built with the statsmodels v0.14.2 package⁸⁴. Mann–Whitney U tests and paired t-tests were calculated with scipy.stats v1.12.0⁸⁵. Multiple corrections were calculated with the statsmodels.stats.multitest module for Python. GSEA tests and statistical measures were obtained with gseapy v1.1.9 for Python v3.11. Visualizations were created using Plotly v5.23.0 for Python v3.11, with UpSet diagrams prepared in plotly_upset extension of Plotly⁸⁶. Images were saved as SVG files and later composed into panels using Adobe Illustrator v29.6.1 for Windows 10. Venn diagrams were prepared using InteractiVenn web-based tool⁸⁷.

Materials availability

This study did not generate new unique reagents. Rentosertib was provided by Insilico Medicine under Good Manufacturing Practice guidelines as part of the original phase 2a trial (NCT05938920).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

The proteomic aging clock analysis code has been made available on GitHub for ipfP3GPT, OrganAgem_chrono, OrganAge_mortality and PAC clocks: https://github.com/Insilico-org/proteoclock. ProtAge clock is available at https://github.com/miargentieri/proteomic-age-ukb. Original code for OrganAge clocks is available at https://github.com/ludgergoeminne/organAging. Original code for the PAC clock is available at https://github.com/kuo-lab-uchc/PAC. Original code for ipfP3GPT clock is available at an Opens Science Framework: https://osf.io/457w8/. Original code for PAOPAC clock is available at https://github.com/JackieHanLab/PAOPAC.

1. Li, Z. et al. Aging and age-related diseases: from mechanisms to therapeutic strategies. Biogerontology 22, 165–187 (2021).

   Article  PubMed  PubMed Central  Google Scholar 

2. Guo, J. et al. Aging and aging-related diseases: from molecular mechanisms to interventions and treatments. Signal Transduct. Target. Ther. 7, 391 (2022).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

3. Dönertaş, H. M., Fabian, D. K., Fuentealba, M., Partridge, L. & Thornton, J. M. Common genetic associations between age-related diseases. Nat. Aging 1, 400–412 (2021).

   Article  PubMed  PubMed Central  Google Scholar 

4. Zhavoronkov, A., Wilczok, D., Ren, F. & Galkin, F. Age-related diseases as a testbed for anti-aging therapeutics: the case of idiopathic pulmonary fibrosis. Aging 17, 1911–1928 (2025).

   CAS  PubMed  PubMed Central  Google Scholar 

5. Kulkarni, A. S. et al. Geroscience-guided repurposing of FDA-approved drugs to target aging: a proposed process and prioritization. Aging Cell 21, e13596 (2022).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

6. López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243–278 (2023).

   Article  PubMed  Google Scholar 

7. Blagosklonny, M. V. Validation of anti-aging drugs by treating age-related diseases. Aging 1, 281–288 (2009).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

8. Galkin, F., Zhang, B., Dmitriev, S. E. & Gladyshev, V. N. Reversibility of irreversible aging. Ageing Res. Rev. 49, 104–114 (2019).

   Article  PubMed  Google Scholar 

9. Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol 14, R115–R115 (2013).

   Article  PubMed  PubMed Central  Google Scholar 

10. Min, M., Egli, C., Dulai, A. S. & Sivamani, R. K. Critical review of aging clocks and factors that may influence the pace of aging. Front. Aging 5, 1487260 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

11. Apsley, A. T., Etzel, L., Ye, Q. & Shalev, I. From population science to the clinic? Limits of epigenetic clocks as personal biomarkers. Epigenomics 17, 1447–1461 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

12. Crimmins, E. M., Klopack, E. T. & Kim, J. K. Generations of epigenetic clocks and their links to socioeconomic status in the Health and Retirement Study. Epigenomics 16, 1031–1042 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

13. Fuentealba, M. et al. Multi-omics analysis reveals biomarkers that contribute to biological age rejuvenation in response to single-blinded randomized placebo-controlled therapeutic plasma exchange. Aging Cell 24, e70103 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

14. Teschendorff, A. E. & Horvath, S. Epigenetic ageing clocks: statistical methods and emerging computational challenges. Nat. Rev. Genet. 26, 350–368 (2025).

    Article  CAS  PubMed  Google Scholar 

15. Mitteldorf, J. Methylation clocks for evaluation of anti-aging interventions. Aging 17, 1082–1090 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Argentieri, M. A. et al. Proteomic aging clock predicts mortality and risk of common age-related diseases in diverse populations. Nat. Med. 30, 2450–2460 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. Johnson, A. A., Shokhirev, M. N., Wyss-Coray, T. & Lehallier, B. Systematic review and analysis of human proteomics aging studies unveils a novel proteomic aging clock and identifies key processes that change with age. Ageing Res. Rev. 60, 101070 (2020).

    Article  CAS  PubMed  Google Scholar 

18. Sun, B. B. et al. Plasma proteomic associations with genetics and health in the UK Biobank. Nature 622, 329–338 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

19. Lee-Ødegård, S., Austin Argentieri, M., Norheim, F., Drevon, C. A. & Birkeland, K. I. Reversal of proteomic aging with exercise—results from the UK biobank and a 12-week intervention study. NPJ Aging 12, 19 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

20. Yang, Y. et al. Metformin decelerates aging clock in male monkeys. Cell 187, 6358–6378 (2024).

    Article  CAS  PubMed  Google Scholar 

21. Pun, F. W. et al. Hallmarks of aging-based dual-purpose disease and age-associated targets predicted using PandaOmics AI-powered discovery engine. Aging 14, 2475–2506 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

22. Xu, Z. et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat. Med. 31, 2602–2610 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

23. Olsen, A. et al. Identification of dual-purpose therapeutic targets implicated in aging and glioblastoma multiforme using PandaOmics—an AI-enabled biological target discovery platform. Aging 15, 2863 (2023).

    CAS  PubMed  PubMed Central  Google Scholar 

24. Ewald, C. Y. et al. TNIK’s emerging role in cancer, metabolism, and age-related diseases. Trends Pharmacol. Sci. 45, 478–489 (2024).

    Article  CAS  PubMed  Google Scholar 

25. Ren, F. et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nat. Biotechnol. 43, 63–75 https://doi.org/10.1038/s41587-024-02143-0 (2025).

26. Chen, J. et al. From clock to clock: therapeutic target discovery for aging and age-related diseases. Ageing Res. Rev. 112, 102871 (2025).

    Article  CAS  PubMed  Google Scholar 

27. Aladinskiy, V. et al. Discovery of bis-imidazolecarboxamide derivatives as novel, potent, and selective TNIK inhibitors for the treatment of idiopathic pulmonary fibrosis. J. Med. Chem. 67, 19121–19142 (2024).

    Article  CAS  PubMed  Google Scholar 

28. Pham, T. C. P. et al. TNIK is a conserved regulator of glucose and lipid metabolism in obesity. Sci. Adv. 9, eadf7119 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

29. Gross, T. J. & Hunninghake, G. W. Idiopathic pulmonary fibrosis. N. Engl. J. Med. 345, 517–525 (2001).

    Article  CAS  PubMed  Google Scholar 

30. Chilosi, M. et al. Aberrant Wnt/β-catenin pathway activation in idiopathic pulmonary fibrosis. Am. J. Pathol. 162, 1495–1502 (2003).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

31. King, T. E., Pardo, A. & Selman, M. Idiopathic pulmonary fibrosis. Lancet 378, 1949–1961 (2011).

    Article  PubMed  Google Scholar 

32. Palombella, L., Amrein, M., Bührer, E., Riether, C. & Ochsenbein, A. P1343: TNIK signalling impairs self-renewal and long-term reconstitution capacity of HSCS. HemaSphere 7, e78156dc (2023).

    Article  PubMed Central  Google Scholar 

33. Kiessling, P. et al. Polyploid cardiomyocytes define disease-specific transcriptional states in the mammalian heart. Preprint at bioRxiv https://doi.org/10.64898/2026.01.31.701472 (2026).

34. InSilico Medicine Hong Kong Limited. Study Evaluating INS018_055 Administered Orally to Subjects With Idiopathic Pulmonary Fibrosis. ClinicalTrials.gov https://clinicaltrials.gov/study/NCT05975983 (2024).

35. Goeminne, L. J. E. et al. Plasma protein-based organ-specific aging and mortality models unveil diseases as accelerated aging of organismal systems. Cell Metab. 37, 205–222 (2025).

    Article  CAS  PubMed  Google Scholar 

36. Kuo, C.-L. et al. Proteomic aging clock (PAC) predicts age-related outcomes in middle-aged and older adults. Aging Cell 23, e14195 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

37. Galkin, F., Chen, S., Aliper, A., Zhavoronkov, A. & Ren, F. AI-driven toolset for IPF and aging research associates lung fibrosis with accelerated aging. Aging 17, 1999–2014 (2025).

    CAS  PubMed  Google Scholar 

38. Xu, H., Chen, J., Chen, D., Mao, K. & Han, J.-D. J. Proteome-aware organ proxy aging clocks. Preprint at bioRxiv https://doi.org/10.64898/2026.04.24.720503 (2026).

39. Schwartz, D. A. et al. Determinants of survival in idiopathic pulmonary fibrosis. Am. J. Respir. Crit. Care Med. 149, 450–454 (1994).

    Article  CAS  PubMed  Google Scholar 

40. Oldham, J. M. et al. Proteomic biomarkers of survival in idiopathic pulmonary fibrosis. Am. J. Respir. Crit. Care Med. 209, 1111–1120 (2023).

    Article  Google Scholar 

41. Ramasamy, R., Shekhtman, A. & Schmidt, A. M. The multiple faces of RAGE—opportunities for therapeutic intervention in aging and chronic disease. Expert Opin. Ther. Targets 20, 431–446 (2016).

    Article  CAS  PubMed  Google Scholar 

42. Yuan, A. & Nixon, R. A. Neurofilament proteins as biomarkers to monitor neurological diseases and the efficacy of therapies. Front. Neurosci. 15, 689938 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

43. Gunes, S., Hekim, G. N. T., Arslan, M. A. & Asci, R. Effects of aging on the male reproductive system. J. Assist. Reprod. Genet. 33, 441–454 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

44. Zaidi, M. et al. FSH, bone mass, body fat, and biological aging. Endocrinology 159, 3503–3514 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

45. Seo, D. H. et al. Chemokine CXCL9, a marker of inflammaging, is associated with changes of muscle strength and mortality in older men. Osteoporos. Int. 35, 1789–1796 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

46. Tang, Q. et al. AI-driven robotics laboratory identifies pharmacological TNIK inhibition as a potent senomorphic agent. Aging Dis 17, 432–451 (2025).

    PubMed  PubMed Central  Google Scholar 

47. Liu, J., Wang, F. & Luo, F. The role of JAK/STAT pathway in fibrotic diseases: molecular and cellular mechanisms. Biomolecules 13, 119 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

48. Rolland, Y. et al. Challenges in developing Geroscience trials. Nat. Commun. 14, 5038 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

49. Nielsen, J. L., Bakula, D. & Scheibye-Knudsen, M. Clinical trials targeting aging. Front. Aging 3, 820215 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

50. Baxter, R. C. Signaling pathways of the insulin-like growth factor binding proteins. Endocr. Rev. 44, 753–778 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

51. Alessio, N. et al. Increase of circulating IGFBP-4 following genotoxic stress and its implication for senescence. eLife 9, e54523 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

52. Schafer, M. J. et al. Cellular senescence mediates fibrotic pulmonary disease. Nat. Commun. 8, 14532 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

53. de Witte, W. E. A., Danhof, M., van der Graaf, P. H. & de Lange, E. C. M. The implications of target saturation for the use of drug–target residence time. Nat. Rev. Drug Discov. 18, 84–84 (2019).

    Article  Google Scholar 

54. Redfern, W. S. et al. Predicting clinical outcomes from off-target receptor interactions using Secondary Intelligence^TM. J. Pharmacol. Toxicol. Methods 131, 107570 (2025).

    Article  CAS  PubMed  Google Scholar 

55. Kanai, Y. & Endou, H. Functional properties of multispecific amino acid transporters and their implications to transporter-mediated toxicity. J. Toxicol. Sci. 28, 1–17 (2003).

    Article  CAS  PubMed  Google Scholar 

56. Gayvert, K. M., Madhukar, N. S. & Elemento, O. A data-driven approach to predicting successes and failures of clinical trials. Cell Chem. Biol. 23, 1294–1301 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

57. Evers, R. et al. Disease-associated changes in drug transporters may impact the pharmacokinetics and/or toxicity of drugs: a white paper from the International Transporter Consortium. Clin. Pharmacol. Ther. 104, 900–915 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

58. Rodrigues, A. D. & Rowland, A. Profiling of drug-metabolizing enzymes and transporters in human tissue biopsy samples: a review of the literature. J. Pharmacol. Exp. Ther. 372, 308–319 (2020).

    Article  CAS  PubMed  Google Scholar 

59. Pang, Z. et al. Mapping cellular targets of covalent cancer drugs in the entire mammalian body. Cell 189, 725–738 (2026).

    Article  CAS  PubMed  Google Scholar 

60. Kim, D., Lee, J., Lee, S., Park, J. & Lee, D. Predicting unintended effects of drugs based on off-target tissue effects. Biochem. Biophys. Res. Commun. 469, 399–404 (2016).

    Article  CAS  PubMed  Google Scholar 

61. Niederst, M. J. & Engelman, J. A. Bypass mechanisms of resistance to receptor tyrosine kinase inhibition in lung cancer. Sci. Signal. 6, re6 (2013).

    Article  PubMed  PubMed Central  Google Scholar 

62. Ilan, Y. Overcoming compensatory mechanisms toward chronic drug administration to ensure long-term, sustainable beneficial effects. Mol. Ther. Methods Clin. Dev. 18, 335–344 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

63. Romano, S. et al. Drug holiday of interferon beta 1b in multiple sclerosis: a pilot, randomized, single blind study of non-inferiority. Front. Neurol. 10, 695 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

64. Arriola Apelo, S. I., Pumper, C. P., Baar, E. L., Cummings, N. E. & Lamming, D. W. Intermittent administration of rapamycin extends the life span of female C57BL/6J mice. J. Gerontol. Ser. A 71, 876–881 (2016).

    Article  Google Scholar 

65. FDA–NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and Other Tools) Resource (United States Food and Drug Administration, 2016).

66. Do, D. et al. FDA cardiovascular. indication expansion and dispensing of semaglutide/tirzepatide in CVD patients with overweight/obesity. JACC Adv. 5, 102575 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

67. Kamya, P. et al. PandaOmics: an AI-driven platform for therapeutic target and biomarker discovery. J. Chem. Inf. Model. 64, 3961–3969 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

68. Santos, R. et al. A comprehensive map of molecular drug targets. Nat. Rev. Drug Discov. 16, 19–34 (2017).

    Article  CAS  PubMed  Google Scholar 

69. Bailey, C. J. Metformin: historical overview. Diabetologia 60, 1566–1576 (2017).

    Article  CAS  PubMed  Google Scholar 

70. Powers, T. The origin story of rapamycin: systemic bias in biomedical research and cold war politics. Mol. Biol. Cell 33, pe7 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

71. Mootha, V. K. et al. PGC-1α-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat. Genet. 34, 267–273 (2003).

    Article  CAS  PubMed  Google Scholar 

72. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA 102, 15545–15550 (2005).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

73. Saul, D. et al. A new gene set identifies senescent cells and predicts senescence-associated pathways across tissues. Nat. Commun. 13, 4827 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

74. Avelar, R. A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

75. Chatsirisupachai, K., Palmer, D., Ferreira, S. & de Magalhães, J. P. A human tissue-specific transcriptomic analysis reveals a complex relationship between aging, cancer, and cellular senescence. Aging Cell 18, e13041 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

76. Kanehisa, M., Goto, S., Furumichi, M., Tanabe, M. & Hirakawa, M. KEGG for representation and analysis of molecular networks involving diseases and drugs. Nucleic Acids Res. 38, D355–D360 (2010).

    Article  CAS  PubMed  Google Scholar 

77. Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst. 1, 417–425 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

78. Liberzon, A. et al. Molecular signatures database (MSigDB) 3.0. Bioinformatics 27, 1739–1740 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

79. Fabregat, A. et al. Reactome pathway analysis: a high-performance in-memory approach. BMC Bioinf. 18, 142 (2017).

    Article  Google Scholar 

80. Fang, Z., Liu, X. & Peltz, G. GSEApy: a comprehensive package for performing gene set enrichment analysis in Python. Bioinformatics 39, btac757 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

81. Weidner, C. et al. A protocol for using gene set enrichment analysis to identify the appropriate animal model for translational research. J. Vis. Exp. 16, 55768 (2017).

    Google Scholar 

82. Borenstein, M., Hedges, L. V., Higgins, J. P. T. & Rothstein, H. R. Effect sizes based on means. in Introduction to Meta-Analysis (eds Borenstein, M., Hedges, L. V., Higgins, J. P. T. & Rothstein, H. R.) Ch. 4, 21–32 https://doi.org/10.1002/9780470743386.ch4 (Wiley, 2009).

83. von Mering, C. et al. STRING: known and predicted protein–protein associations, integrated and transferred across organisms. Nucleic Acids Res. 33, D433–D437 (2005).

    Article  Google Scholar 

84. Seabold, S. & Perktold, J. Statsmodels: econometric and statistical modeling with Python. scipy https://doi.org/10.25080/Majora-92bf1922-011 (2010).

85. Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat. Methods 17, 261–272 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

86. Lex, A., Gehlenborg, N., Strobelt, H., Vuillemot, R. & Pfister, H. UpSet: visualization of intersecting sets. IEEE Trans. Vis. Comput. Graph. InfoVis 20, 1983–1992 (2014).

    Article  Google Scholar 

87. Heberle, H., Meirelles, G. V., da Silva, F. R., Telles, G. P. & Minghim, R. InteractiVenn: a web-based tool for the analysis of sets through Venn diagrams. BMC Bioinf. 16, 169 (2015).

    Article  Google Scholar 

Download references

Authors and Affiliations

1. Insilico Medicine AI Limited, Abu Dhabi, United Arab Emirates

   Alex Zhavoronkov, Fedor Galkin, Alex Aliper, Mikhail Durymanov & Denis Sidorenko

2. Insilico Medicine Shanghai Ltd., Shanghai, China

   Alex Zhavoronkov, Shan Chen, Feng Ren & Hui Cui

3. Insilico Medicine US Inc., Cambridge, MA, USA

   Alex Zhavoronkov

4. Peking-Tsinghua Center for Life Sciences, Academy for Advanced Interdisciplinary Studies, Center for Quantitative Biology (CQB), Peking University, Beijing, China

   Jing-Dong J. Han & Hao Xu

5. Peking University Chengdu Academy for Advanced Interdisciplinary Biotechnologies, Chengdu, China

   Jing-Dong J. Han

6. Westlake Laboratory of Life Sciences and Biomedicine, Hangzhou, China

   Xiaodong Liu

7. School of Life Sciences, Westlake University, Hangzhou, China

   Xiaodong Liu

8. Research Center for Industries of the Future, Westlake University, Hangzhou, China

   Xiaodong Liu

9. Department of Pulmonary and Critical Care Medicine, Peking Union Medical College Hospital, Peking Union Medical College and Chinese Academy of Medical Sciences, Beijing, China

   Zuojun Xu

10. Nephrology Department, RWTH Aachen University, Aachen, Germany

    Christoph Kuppe

11. Analytic and Translational Genetics Unit, Massachusetts General Hospital, Boston, MA, USA

    M. Austin Argentieri

12. Program in Medical and Population Genetics, Broad Institute of MIT and Harvard, Cambridge, MA, USA

    M. Austin Argentieri

13. The Phil and Penny Knight Initiative for Brain Resilience, Stanford University, Stanford, CA, USA

    Kejun Ying

14. Wu Tsai Neurosciences Institute, Stanford University, Stanford, CA, USA

    Kejun Ying

15. Institute for Protein Design, University of Washington, Seattle, WA, USA

    Kejun Ying

16. Department of Medicine, Division of Genetics, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA

    Ludger J. E. Goeminne, Mahdi Moqri, Alexander Tyshkovskiy & Vadim N. Gladyshev

17. Broad Institute of MIT and Harvard, Cambridge, MA, USA

    Ludger J. E. Goeminne, Mahdi Moqri, Alexander Tyshkovskiy & Vadim N. Gladyshev

Contributions

Conceptualization: A.A., A.Z., F.R. Data curation: H.C., S.C. Formal analysis: D.S., F.G., H.C., J.-D.J.H., K.Y., M.A.A. Investigation: A.A., A.Z., F.G., M.D. Methodology: A.A., A.Z., F.G., H.X., M.A.A., M.M., Z.X. Project administration: A.A., A.Z., S.C., Z.X. Resources: A.Z., F.R., S.C. Software: D.S., F.G., H.X., K.Y., L.J.E.G., M.A.A., M.M., V.N.G. Supervision: A.A., A.Z. Validation: A.A., A.T., F.G., J.-D.J.H., K.Y., L.J.E.G., M.A.A., V.N.G. Visualization: F.G., M.D. Writing—original draft: A.Z., A.T., C.K., F.G., M.M., M.D., V.N.G., X.L.

Corresponding author

Correspondence to Alex Zhavoronkov.

Competing interests

A.Z. is the founder and chief executive officer of Insilico Medicine, a publicly traded (3696.HK) company developing and using generative AI for drug discovery and aging research. F.G., S.C., F.R., A.A, M.D., D.S. and H.C. are employees at Insilico Medicine. Insilico Medicine has developed a portfolio of therapeutic programs targeting fibrotic diseases, including rentosertib for IPF. The other authors declare no competing interests.

Peer review information

Nature Biotechnology thanks Tamir Chandra, Samuel Crofts and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Reporting Summary (download PDF )

Peer Review File (download PDF )

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions