Measure thousands of proteins floating in blood serum, feed their relative proportions into an algorithm, and out comes a number: a predicted "biological age" that estimates how worn down the body is, distinct from chronological age. On September 7, 2026, a paper published in the journal Nature Biotechnology (Zhavoronkov, Galkin, Chen et al., DOI: 10.1038/s41587-026-03286-y) reported that, based on serum data from a clinical trial, this number rolled back by roughly three years.
The molecule in question is rentosertib (development code INS018_055), an experimental small-molecule drug being developed by Insilico Medicine as a candidate treatment for idiopathic pulmonary fibrosis (IPF). Its target identification and molecular scaffold optimization were both carried out using an artificial intelligence platform. When researchers ran blood data from patients who received the drug through six independent proteomic aging-clock models, every single model showed a shift toward a younger biological age.
But it would be premature to interpret this finding as "humans biologically got younger" or "lifespan was extended." What was actually observed is simply a change in model predictions calculated by an algorithm from shifts in blood protein concentrations. When the composition of biomolecules—severely disrupted by disease progression—moves back toward healthy levels thanks to the drug's antifibrotic action, the clock algorithm interprets that shift as "slower aging." Rentosertib remains in clinical trials and has not received any regulatory approval. How much overlap exists between healing damaged tissue and genuinely suppressing aging is a question that still demands rigorous scientific scrutiny.
A trial design tracking 42 patients' serum proteomes over 12 weeks
This aging analysis was conducted as an exploratory secondary analysis of a randomized, double-blind, placebo-controlled Phase 2a clinical trial (ClinicalTrials.gov registration number: NCT05938920) conducted at 21 sites across China between 2023 and 2024. The parent trial enrolled 71 patients aged 40 or older with confirmed, stable IPF, who were assigned to one of four groups: rentosertib 30 mg once daily, 30 mg twice daily, 60 mg once daily, or placebo. Sixteen patients discontinued treatment during the 12-week dosing period.
The aging-clock analysis proceeded with 42 patients who provided additional consent and had complete serum samples collected at all four timepoints—baseline, and 2, 4, and 12 weeks post-dosing. Of the 43 patients who gave consent, one was excluded because their final blood draw was missing. The 42 analyzed patients had an average age of 67.1 years, and all were of Asian ethnicity. This is an elderly, single-ethnicity population with a specific serious respiratory disease, and findings from this data cannot be extrapolated to the general healthy population.
The research team ran the collected serum samples through Olink Explore 3072, a highly sensitive Swedish proteomics platform, comprehensively profiling the expression of 2,841 proteins. The raw data has been deposited in the archive of the China National Center for Bioinformation (CNCB) (accession number: OMIX008341), and the source code for the analysis pipeline has also been made public.
To estimate biological age from serum proteins, the research group applied six independently developed proteomic aging clocks, created by research groups at Harvard University, Oxford University, Peking University, and Insilico Medicine, among others. These included models trained to predict chronological age (ProtAge, OrganAge chronological, ipfP3GPT, PAOPAC) as well as models trained to predict mortality risk or remaining lifespan (OrganAge mortality, PAC). Rather than relying on a single proprietary algorithm, the team adopted a design that tested reproducibility using multiple external models built on different assumptions and training methods.
Age-prediction declines concentrated at week 4, with signals diverging by dose
Comparisons between each dosing group and the placebo group were carried out across 54 combinations total: six aging clocks, three measurement timepoints (weeks 2, 4, and 12), and three dosage levels. After statistical correction for multiple comparisons (false discovery rate $Q < 0.10$), 21 of these combinations passed the significance threshold. In permutation testing, the expected number of significant differences arising purely by chance averaged only 0.15 combinations, meaning that 21 detected combinations clearly exceeds statistical noise.
However, the signal was not evenly distributed. The majority of significant differences clustered at the 4-week dosing timepoint, with 11 of 18 comparisons at that point showing significant declines.
- 60mg once daily
- 30mg twice daily
データを表で見る
| 60mg once daily (years) | 30mg twice daily (years) | |
|---|---|---|
| ProtAge | 2.71 | 3.2 |
| OrganAge (chronological) | 3.46 | 3.85 |
| ipfP3GPT | 3.12 | 3.6 |
| PAOPAC | 2.95 | 3.4 |
| PAC (mortality risk) | — | 2.1 |
| OrganAge (mortality risk) | — | 1.85 |
At week 4, the 60 mg once-daily group showed predicted-age reductions of $-2.71$ to $-3.46$ years across all four chronological-age-trained clocks, reaching statistical significance. However, the two clocks trained on mortality risk showed no significant change in this group. By contrast, the 30 mg twice-daily group showed rejuvenation signals across both chronological-type and mortality-risk-type clocks, with significant differences detected in 9 of the comparisons. Insilico describes the peak change in this group as a rejuvenation effect of three to four years, with a maximum reduction of roughly six years.
Notably, the aging-clock signal peaked at week 4 and showed a tendency to weaken by week 12. Even though the underlying changes in individual protein concentrations persisted through week 12, the algorithm's aggregate score decayed. Reanalysis after excluding six patients who experienced high-grade adverse events during the trial did not change these statistical trends. Across the whole serum proteome, only two proteins showed significantly altered expression trajectories in the placebo group, compared with 326 proteins showing expression changes across the rentosertib dosing groups. Of these, 142 were changes unique to the 30 mg twice-daily group, which showed the broadest response.
| Dosing regimen (sample size) | Significant aging-clock comparisons (out of 18 per group) | Predicted age change at week 4 | FVC change in parent Phase 2a trial | Number of uniquely altered serum proteins |
|---|---|---|---|---|
| Placebo group ($n=11$) | Not applicable (control group) | Baseline (slight increase or no change) | 2 | |
| 30 mg once daily ($n=10$) | 5 | Minor change | 45 | |
| 30 mg twice daily ($n=11$) | 9 | Up to $-3.85$ years (detected in both chronological and mortality types) | 142 | |
| 60 mg once daily ($n=10$) | 7 | $-2.71$ to $-3.46$ years (significant only in the 4 chronological-type clocks) | 88 |
As the table shows, there is a mismatch between the aging-marker response and the clinical measure of lung function. The largest improvement in vital capacity occurred in the 60 mg group, but the strongest consistency across aging clocks was seen in the 30 mg twice-daily group.
AI target discovery on the fibrosis gene TNIK, and a computational hypothesis about suppressing senescent cells
The compound rentosertib was born as a demonstration case for AI in drug discovery pipelines. Using a target-discovery engine, Insilico Medicine identified TNIK (TRAF2 and NCK interacting kinase) as a molecule implicated in six of the hallmark features of cellular senescence and deeply connected to the fibrotic cascade in lung tissue. The company then used its molecular design AI platform, Chemistry42, to synthesize a small-molecule compound that selectively inhibits TNIK. The process from target identification to preclinical candidate selection reportedly took about 18 months. The details of this preclinical development were reported in Nature Biotechnology in 2024.
Unlike efforts to repurpose existing approved drugs for anti-aging use—as with rapamycin or metformin—the company positions this drug as "the first case in which AI carried out both target identification and molecular design from scratch, advancing to clinical trials that verified aging biomarkers."
Against the backdrop of serum proteomic changes, the research team cross-referenced proteomic dynamics data from 55,319 UK Biobank participants and conducted pathway analysis. The results confirmed that, following rentosertib treatment, the expression profile of inflammation- and fibrosis-related factors that normally increase with aging was downregulated. Specifically, components of the senescence-associated secretory phenotype (SASP)—harmful bioactive substances secreted by senescent cells, including EREG, ESM1, IGFBP4, ITGA2, MMP10, MMP13, and SPP1—were uniformly suppressed.
Based on this data, the authors proposed that rentosertib may be behaving not as a senolytic, which induces death in senescent cells, but as a senomorphic agent, which suppresses the harmful secretory activity of senescent cells without killing them. However, this mechanistic interpretation is derived purely from computational pathway enrichment analysis. The researchers did not directly measure senescence markers in patients' lung tissue, and causality has not been experimentally established. The paper's authors themselves cautiously describe these findings as "molecular signatures warranting further investigation."
LTBP2: the confounder blocking a clean distinction between disease relief and true anti-aging
At the core of this paper lies a fundamental limitation of the measurement system itself: the proteomic aging clock. As the authors themselves state explicitly in the paper, it is inherently impossible to separate "a change in aging itself" from "disease-specific pathological improvement" using the aging-clock score alone. They conclude that this separation "cannot be achieved within an IPF patient cohort alone."
The biggest confounding factor is the behavior of LTBP2 (latent transforming growth factor beta binding protein 2). LTBP2 was the only protein given consistently high weighting across all six aging-clock models, and it was the single largest contributor to the predicted decline in biological age observed in this study.
However, LTBP2 is an extracellular matrix protein tightly linked to the TGF-β signaling pathway, the primary driver of pulmonary fibrosis. It floods into the blood as lung tissue destruction and fibrosis progress, and it decreases as disease activity subsides. In other words, one cannot rule out the scenario in which rentosertib's primary antifibrotic action eased local lung inflammation and tissue remodeling, lowering serum LTBP2 concentrations—and the algorithm then misread this as "the whole body got younger." It is possible that the biochemical normalization that accompanies disease recovery is simply making the aging clock's hands appear to spin backward.
Furthermore, the correlation between respiratory function improvement and aging-clock scores is extremely weak. When researchers examined the relationship between the degree of improvement in forced vital capacity (FVC)—a measure of lung ventilation capacity—and the reduction in predicted age produced by each aging clock, the median R² (a measure of explained variance) was only 0.06. There were cases where patients with dramatic recovery of respiratory function showed almost no movement in the aging model's predicted age, and vice versa.
As noted above, there is also a mismatch between dosage and clinical endpoints. In the parent trial, the greatest improvement in FVC occurred in the 60 mg group ( versus in the placebo group), yet the most consistent rejuvenation signal across all six aging-clock models appeared in the 30 mg twice-daily group. These discrepancies suggest that the degree of disease improvement and the aging clock's computational results are not capturing the same phenomenon. The parent trial was a small Phase 2a study designed primarily to evaluate 12-week safety and tolerability, and the significance of exploratory biomarkers extracted from it does not mean that patients' actual healthy lifespan was extended.
Not a lifespan extension—a methodological attempt to embed aging markers into a disease trial
How should this result be understood? Commentators like American entrepreneur Peter Diamandis, who has publicly claimed that "if you survive until the 2030s when AI extends lifespan, you'll reach immortality," or enthusiasts like Bryan Johnson, who tests every conceivable anti-aging protocol on his own body, might welcome this paper as "proof of AI-driven drug rejuvenation." But such narratives are not scientific fact—they are observations mixed with commercial expectation.
The essential claim the research team is making in this paper is not lifespan extension or rejuvenation of healthy people. Rather, following the U.S. Food and Drug Administration's (FDA) BEST (Biomarker, Endpoint, and other Tools) framework—which defines and organizes drug-development tools such as biomarkers and clinical endpoints—the paper presents a "dual-purpose clinical trial approach": embedding multiple aging biomarkers within a trial designed around a single disease, in order to simultaneously measure how a drug's target pathway intersects with systemic aging processes. This is, fundamentally, a proposal about trial methodology.
The stage of scientific verification has already moved forward. Insilico Medicine has announced a 52-week Phase 3 clinical trial enrolling a total of 320 IPF patients across 47 sites in China. The primary endpoint of this large-scale trial is the annual rate of decline in forced vital capacity over 52 weeks—not aging-clock scores. As a trial aimed at practical approval, the sole pass/fail criterion will be whether the drug can halt the progression of respiratory failure. According to records in the U.S. clinical trial registry (ClinicalTrials.gov), as of the update on July 7, 2026, this trial has a status of not yet recruiting participants.
Whether healthy individuals who take rentosertib would experience suppressed aging, or instead develop serious side effects, remains completely unknown at this point. Improvement in the condition of elderly patients with a serious, potentially fatal disease is not the same concept as suppressing aging in a healthy human body. To move from the computational fact that a serum protein profile temporarily changed to claims about a genuine extension of healthy lifespan, there remain multiple experimental and epidemiological hurdles still to be cleared.