Insilico Medicine’s Latest Breakthrough: Its AI-Assisted Drug Shows Signs of Reversing Human “Biological Age” in Clinical Data—by 6 Years

Why Measure Aging Clocks in a Lung Disease Trial?

To understand this study, one must first recognize the unique intersection between idiopathic pulmonary fibrosis (IPF) and "aging."

IPF is a chronic, progressive, and fatal lung disease that predominantly occurs in middle-aged and older populations. Patients experience recurrent alveolar epithelial injury and aberrant repair, leading to the replacement of normal alveolar tissue with fibrotic scar tissue, causing the lungs to progressively stiffen and lose elasticity. In the fields of gerontology and pathology, IPF is often described as a model disease with a "high enrichment of aging hallmarks"—cellular senescence, mitochondrial dysfunction, abnormal extracellular matrix remodeling, and chronic sterile inflammation are precisely the core pathological mechanisms driving its fibrotic exacerbation.

Previously, the evaluation of anti-aging interventions often relied on DNA methylation-based "epigenetic clocks." However, in clinical applications, these have limitations such as poor read consistency across different models, relatively indirect mechanistic interpretations of specific biological pathways, and blood samples primarily reflecting hematopoietic lineage leukocytes. In contrast, proteins are the direct executors of biological processes and pathological phenotypes; aging clocks constructed from high-throughput serum proteomics (such as Olink technology) facilitate the observation of protein alterations in the circulation and allow direct biological pathway enrichment analysis. However, it must be clarified that changes in circulating blood protein levels reflect systemic predictive indicators and are not equivalent to a comprehensive, immediate reversal of whole-body organs or local lung tissue fibrotic lesions; furthermore, this clinical sub-study did not directly measure DNA methylation, and no parallel head-to-head comparison was conducted between the two classes of clocks.

42 Patients and 6 Algorithmic Clocks: How Were the Data Collected and Computed?

The reliability of any clinical conclusion depends on its enrollment process and sample quality. This proteomic analysis was an exploratory proteomic sub-study conducted on the basis of the original disease clinical trial; subjects underwent longitudinal serum sampling across 4 timepoints during the trial, while informed consent for this secondary analysis was obtained supplementally after the trial concluded.

Sample Screening Pathway

According to methodological data disclosed in the paper:

• The original Phase IIa clinical trial screened a total of 128 middle-aged and older IPF patients, ultimately enrolling and randomizing 71 individuals into four trial arms: placebo group (17 patients), 30mg once daily group (QD, 18 patients), 30mg twice daily group (BID, 18 patients), and 60mg once daily group (QD, 18 patients).
• During the 12-week dosing period, 16 patients discontinued/dropped out, and 55 patients completed the full trial follow-up.
• After the trial ended, the research team provided informed consent forms for serum proteomic analysis to the 55 subjects who completed the trial, of whom 43 signed consent; subsequently, during the assay stage, 1 patient was excluded due to missing serum samples at week 12.
• The final valid sample size included in the complete bioinformatic and clock analysis was 42 patients (all Asian, mean age 67.1 years), comprising:
  • Placebo group: 11 patients
  • 30mg QD group: 11 patients
  • 30mg BID group: 11 patients
  • 60mg QD group: 9 patients

Researchers collected serum samples from subjects at four timepoints: baseline (Week 0), Week 2, Week 4, and Week 12 (trial completion), which were measured using the Olink Explore 3072 high-throughput assay platform. After removing items failing quality control, longitudinal normalized protein expression (NPX) data for 2,841 serum proteins were obtained.

Nature of the 6 Clocks: They Are Not 6 Completely Independent "Replication Experiments"

The research team introduced 6 representative proteomic clock algorithms publicly available in the academic community at that time. It must be explicitly clarified to general readers that: The output value of an aging clock is merely a predictive score from a mathematical model, and by no means equals the absolute physiological age or irreversible aging rate of actual human cells.

These 6 clocks fall into two broad categories based on design approach:

1. Chronological Age Prediction Clocks (Chronological Clocks, 4 in total):
   • Include ProtAge, OrganAge chrono, ipfP3GPT, and PAOPAC.
   • Training Objective: Train models to predict the subject's "true birth chronological age" as accurately as possible based on blood protein levels.
   • Performance at Baseline: These 4 clocks showed strong concordance in predicting the age of this cohort of IPF patients aged around 67 years (Spearman correlation coefficient r = 0.70 ～ 0.84, with root-mean-square error under 4 years after systematic bias correction).
2. Mortality Risk Prediction Clocks (Mortality-based Clocks, 2 in total):
   • Include PAC and OrganAge mortality.
   • Training Objective: Train models to predict subjects' mortality risk and convert the predictions into biological age readings.
   • Performance at Baseline: They exhibited weaker correlations with chronological age (r = 0.16 ～ 0.23) and showed systematic overestimation in younger patients. The authors explained that this aligns with the models' training objectives, and that the cohort's own high disease burden likely influenced the readings, leading models trained on mortality risk to assess higher biological ages regardless of patients' actual chronological age.

Scrutiny of Important Facts: Some media reports interpreted the convergence of these 6 clocks as "6 completely unrelated, independent AI systems simultaneously providing identical proof." This statement is inaccurate. In reality, during their original development, these 6 clocks all utilized proteomic data from the large-scale UK Biobank cohort, and the models share features such as LTBP2. They represent different perspectives of the same underlying data framework under different algorithmic paradigms (deep learning vs. classical machine learning) and different optimization objectives (chronological age vs. mortality risk), with inherent feature cross-talk among them.

Key Signals Across 54 Comparisons: Which Looked Younger? Which Did Not?

The core analysis of the study was to observe whether, over the 12-week dosing period, readings from these 6 clocks in patients underwent statistically credible downward shifts (i.e., negative ΔBioAge changes) compared to pre-treatment (baseline) and compared to the placebo group.

The overall trial design formed: 6 clocks × 3 post-treatment follow-up timepoints (Weeks 2, 4, 12) × 3 dosing groups = 54 cross-sectional comparisons.

Full Picture of the Statistical Results

1. Significance Distribution: Out of all 54 comparisons, 21 comparisons reached statistical significance after Benjamini–Hochberg multiple testing correction (false discovery rate Q < 0.10). Here, Q < 0.10 serves as an exploratory screening threshold after multiple hypothesis testing correction to control the false discovery rate across the batch of comparisons, and should not be interpreted as an individual comparison having a 90% certainty or true probability.
2. Peak Occurred at Week 4: These 21 significant signals were heavily concentrated at Week 4 post-dosing. Out of the 18 comparisons performed at Week 4, 11 comparisons observed a significant downward shift in predicted age in the treatment groups compared to placebo.
3. Observed Significant Comparisons Far Exceeded Random Expectation: To test whether the 21 significant comparisons could easily arise from random fluctuations, researchers conducted a permutation test with 100,000 random shuffles of patient labels. The results showed that under a purely random null hypothesis, an average of only 0.15 comparisons were significant false positives. The actually observed 21 significant results far exceeded this random expectation, supporting the presence of a drug-related signal; however, limited by the exploratory nature of the analysis and small sample size, potential bias or sampling error cannot be ruled out.
4. Divergent Performance Across Dose Groups:
   • 30mg BID (twice daily) group performed most comprehensively and stably: Recorded 9 significant downward shifts across all its 18 comparisons, spanning both chronological and mortality-based clock categories. At Week 4, 5 of the 6 clocks exhibited moderate to high negative effect sizes (Cohen's d).
   • 60mg QD (once daily high-dose) group: Contributed 7 significant comparisons. It must be specifically clarified that prior claims in some reports stating that "patients' physiological age decreased by 2.71 to 3.46 years" specifically referred only to the downward shift range relative to placebo in this 60mg QD group at Week 4 across the 4 chronological clocks, and cannot be generalized as a universal conclusion for all patients, all doses, or across the entire trial duration. Furthermore, in the 60mg QD group, neither of the two mortality-based clocks detected significant changes.
   • 30mg QD (once daily low-dose) group: Only 5 comparisons were significant, showing weaker signals.

The "Plateau" at Week 12

By Week 12 of treatment, the number of comparisons reaching significance declined somewhat. However, paired statistical tests indicated that clock readings between Week 4 and Week 12 showed no significant rebound; a more accurate scientific description is that they entered a plateau.

Was It Fibrosis Improvement, or Independent Aging Intervention?

This is the most critical scientific question confronting all researchers and readers:

Patients' pulmonary fibrosis was controlled after treatment, and as their physical condition improved, blood parameters naturally improved. Did these decreases in clock scores merely represent a "secondary byproduct of lung disease improvement," or did they signify some systemic aging modulation independent of pulmonary fibrosis?

The authors candidly acknowledge: In a patient population with organic pathology (such as IPF), it is mathematically impossible to completely disentangle antifibrotic effects from anti-aging effects relying solely on circulating proteomic algorithms. However, the study provided three parallel lines of indirect exploratory evidence suggesting that dissociation between the two may exist:

1. "Scissors Divergence" in Dose Response

In the primary efficacy report of the original Phase IIa clinical trial, the group with the most prominent improvement in lung function (forced vital capacity, FVC, the volume of air exhaled with maximum effort after a deep inhalation, reflecting a component of lung function) was the 60mg QD group; However, in the evaluation by proteomic aging clocks, the group with the highest cross-model consistency and most robust downward shift was the 30mg BID group, which received the same total daily dose administered in divided doses. If clock changes were purely a direct byproduct of lung function improvement, the group with the largest FVC improvement should theoretically have induced the most dramatic clock downward shifts, yet the empirical data demonstrated a clear dimensional separation.

2. Regression Analysis: Vital Capacity Changes Cannot Explain Clock Variance

Researchers performed univariate linear regression analyses between the change in forced vital capacity (ΔFVC) and the change in predicted biological age (ΔBioAge) across all 42 subjects. The results showed: Across all 6 clocks, the median variance in ΔBioAge explained by ΔFVC was merely R² = 0.06 (ranging between 0.01 and 0.18).

Statistically, R² = 0.06 means that "changes in this vital capacity metric can only explain around 6% of the variance in the readings of these proteomic age models." This by no means equates to "6% of this drug's efficacy comes from the lungs and 94% from anti-aging." While FVC is only an indirect surrogate marker of disease state, the extremely low R² statistically fails to support the simplistic deduction that "clock decreases are entirely driven by lung function recovery."

3. Comparison with the Aging Trajectory of 55,000 Middle-Aged and Older UK Individuals

For further validation, researchers incorporated cross-sectional proteomic data from 55,319 subjects aged 50–85 years (median age 71 years) in the UK Biobank.

• Among the 2,832 proteins overlapping between the two datasets, 758 "age-associated proteins" exhibiting clear upward or downward trajectories with age in the natural population were identified.
• Proteins altered by Rentosertib in the trial were significantly enriched among these 758 age-associated proteins (enrichment fold of 1.74, P < 0.001).
• Next, researchers computed the correlation between the "direction of protein change after patient treatment" and the "direction of protein change during natural aging":
  • 30mg BID group: Exhibited a significant negative correlation (Spearman r = -0.30, P < 0.01), indicating that many key protein changes induced by treatment in this group were directionally opposite to the natural aging process.
  • 60mg QD group: Despite excellent lung function recovery, its protein changes showed no significant correlation with the direction of natural aging (r = -0.097, P = 0.37).
  • Placebo group: Displayed a positive tracking relationship, meaning that as disease progressed over time, patients' blood protein changes moved in the same direction as natural aging.

Proteomic Trajectories and the Mystery of the "Week 4 Plateau"

Why did aging clock scores stop declining prominently after Week 4 and instead enter a plateau? Had the drug lost efficacy in the human body, or had another biological equilibrium emerged?

By fitting linear mixed-effects models (LMEM, adjusted for sex and BMI as covariates) across all 2,841 proteins, researchers uncovered deeper proteomic expression dynamics.

Temporal Trajectory Analysis of Proteomic Responses

Data analysis showed that Rentosertib significantly altered serum levels of 326 proteins during the trial (Q < 0.10). When classifying time by discrete timepoints, researchers found:

• In the two effective dose groups, 30mg BID and 60mg QD, only 5% to 9% of protein alterations were transient (i.e., reaching significance at Week 2 or Week 4, but not at Week 12; however, this does not mean the change completely disappeared, as lack of significant difference between two timepoints does not prove absence of change).
• The vast majority of protein alterations were sustained or delayed (i.e., reaching significance only at Week 12).
• Among the core proteins used by the 4 aging clocks with publicly released feature weights, "sustained alteration" proteins exhibited an odds ratio enrichment of 2.19 to 4.65-fold.

Core Feature Contributions and Clock Score Behavior

Feature Contribution Analysis revealed: Researchers examined 35 key proteins contributing more than 3% to changes in clock scores and found: whether from Week 0 to 4 or Week 4 to 12, LTBP2 was the most prominent contributing feature, and was a key shared protein carrying high weights across all 6 clock algorithms.

LTBP2 (latent transforming growth factor β binding protein 2) is a key factor involved in fibrotic regulation within the TGF-β signaling axis, playing an important role in processes such as extracellular matrix deposition in pulmonary fibrosis. This feature analysis indicates that the aggregate readouts of clock models were significantly influenced by fibrosis-related proteins.

Regarding the phenomenon where aging clock scores plateaued after Week 4, the specific underlying mechanisms remain undetermined. The influence of fibrosis proteins on aggregate readouts suggests that multi-clock outputs must be interpreted comprehensively alongside multi-omics and pathway analyses, but one cannot directly conclude that the plateau is merely a statistical artifact of algorithmic weighting.

What Happened to Cellular Senescence Markers and Metabolic Pathways?

Beyond composite clock scoring, researchers further applied Gene Set Enrichment Analysis (GSEA): proteins were first ranked by the direction and magnitude of drug-induced changes, and then evaluated to see whether sets of proteins associated with known aging features were concentrated toward the upregulated or downregulated side. It should be noted that this enrichment analysis at the blood protein level reflects correlative patterns in the circulatory system, rather than direct measurements of biological activity in tissues.

1. Downregulation of the Senescence-Associated Secretory Phenotype (SASP)

The study evaluated the SenMayo senescence gene set (comprising 124 senescence-associated proteins, of which 91 were measured by the assay panel in this trial):

• Placebo group: As lung disease naturally worsened, the 91 measured SenMayo proteins as a whole were significantly enriched toward the upregulated side (Normalized Enrichment Score NES = 1.48, Q < 0.01), with core contributing proteins including pro-inflammatory cytokines (such as IL1B, IL6) and matrix metalloproteinases (such as MMP1, MMP3, MMP10, MMP13); this reflects an overall enrichment trend, rather than all 91 proteins being individually significantly upregulated.
• Treatment groups: Displayed a starkly opposite downward trend. In both the 30mg QD and 30mg BID groups, the SenMayo gene set showed significant negative enrichment (Q < 0.05); in the 60mg QD group, while the direction was likewise negative, the test statistic did not meet the conventional significance threshold (Q = 0.17).
• The leading-edge subset of downregulated proteins across dosing groups included senescence-related markers such as MMP10, SPP1, and IGFBP4.

The downregulation of SenMayo proteins in serum provides clues consistent with senomorphic modulation (i.e., presenting a trend consistent with the inhibition of senescence-associated secretory phenotypes at the blood level), but this is not equivalent to senolytic clearance of senescent cells in vivo, nor does it represent comprehensive blockade of senescence pathways.

2. Growth Factors, Nutrient Sensing, and Metabolic Remodeling

In classical aging theories, hyperactivation of growth factor receptor and nutrient-sensing pathways (such as PI3K-Akt, MAPK, RAS-ERK, etc.) represents a key mechanism accelerating aging and exhausting tissue stem cells.

• GSEA analysis revealed that these major receptor tyrosine kinase (RTK) downstream pro-growth/pro-proliferative signaling pathways were enriched toward upregulation in the placebo group, whereas all treatment groups showed coordinated negative enrichment (Q < 0.25); this reflects statistical enrichment trends at the pathway level rather than direct local tissue activation or inhibition.
• Insulin-like Growth Factor Axis (IGF Axis): IGF is a critical signaling molecule regulating cell growth and metabolism, whose binding proteins (IGFBPs) regulate IGF transport, tissue uptake, and receptor interactions. Pathway analysis and baseline comparisons indicated that regulatory proteins in serum including IGFBP-4, CCN1, and IGFBPL1 were significantly downregulated compared to baseline, with CCN1, IGFBP4, and IGFBPL1 showing the most consistent downregulation across all dosing regimens.

3. Why Did 30mg BID and 60mg QD Behave Differently? The Pharmacokinetic Hypothesis

Why did two regimens with identical total daily doses (30mg BID twice daily vs. 60mg QD once daily) yield such noticeable divergence across molecular pathways? According to prior clinical pharmacokinetic (PK) data:

• The elimination half-life of Rentosertib is relatively short, approximately 7 to 11 hours.
• Following a single dose of the 60mg QD regimen, peak concentration (Cmax) reached 2 to 3 times that of the 30mg regimen.
• The authors hypothesize that the 30mg BID regimen may provide relatively steadier blood concentration exposure, though this explanation remains untested.

Protein-protein interaction (PPI) network analysis revealed:

• Protein sets unique to 60mg QD: Primarily enriched in Wnt signaling and immune-related pathways; the authors hypothesize that pulsatile peak concentrations may correlate with the intensity of local target inhibition, potentially influencing clinical endpoints like forced vital capacity (FVC), though this is not a proven direct translational relationship. Concurrently, this group and the 30mg QD group exhibited significant upregulation of the fibrosis-associated JAK-STAT cytokine axis (Q < 0.10).
• Protein sets unique to 30mg BID: Primarily enriched in basal metabolic pathways such as the pentose phosphate pathway, glutathione metabolism, and cholesterol metabolism; this group did not exhibit significant upregulation of the JAK-STAT axis (absence of significant upregulation does not equal complete lack of aberrant activation in this pathway).
• Hypothetical Deductions: Different dosing frequencies of a short half-life kinase inhibitor may induce the cellular kinase cascade network into different response states. The authors hypothesize that smoother plasma concentration coverage may be more conducive to metabolic homeostasis and improvement in aging clock signals, but this remains a mechanistic hypothesis based on current data, cannot be generalized as a universal prescribing principle, and still requires subsequent experimental verification.

From Case Study to Paradigm: What Is the True Value of This Research?

Divorced from a rigorous medical context, readers could easily misinterpret this study as "AI has produced an elixir of youth that makes people younger." However, from a rigorous regulatory and clinical science perspective, the greatest contribution of this study is not "proving an anti-aging target," but rather demonstrating a highly instructive methodology for clinical R&D.

"Dual-Purpose Design" in Clinical Research

Traditional anti-aging drug development faces severe regulatory and ethical bottlenecks:

• Under current regulatory frameworks (such as the US FDA), the most realistic path for such research is not conducting clinical trials directly with "aging" as the indication, but embedding them into specific disease trials as exploratory pharmacodynamic biomarkers.
• Historically, most drugs with potential aging-intervention activity (such as metformin, rapamycin, etc.) only had their anti-aging mechanisms slowly explored through retrospective analyses or supplemental trials by academia decades after being approved for diabetes or the prevention of organ-transplant rejection—a cycle spanning decades.

The study proposed a progressive three-step exploratory framework for "Dual-Purpose Clinical Trial Design":

1. Step One (Prospectively embed exploratory endpoints): In clinical trials targeting established age-related diseases (such as IPF), prospectively collect aging and cellular senescence markers as exploratory endpoints, and explore response differences across dosing regimens.
2. Step Two (Replication in non-IPF populations): Under the premise of adequate safety margins and favorable risk-benefit profiles, replicate relevant biological effects in non-IPF age-related diseases or risk-enriched populations to move beyond the confounding background of severe fibrotic tissue remodeling.
3. Step Three (Pursuing functional clinical endpoints or qualified biomarker support): Establish evidence of geroprotection or anti-aging by pursuing qualified biomarkers supported by the FDA Biomarker Qualification Program and FDA-NIH BEST resource framework, or via conventional functional clinical endpoints such as the composite incidence of multiple chronic age-related conditions.

Limitations That Must Be Understood to Interpret the Conclusions

When evaluating this cutting-edge research, the scientific community must remain sober and measured. This study has several objective limitations that restrict the generalizability of its conclusions:

1. Limited Sample Size and Short Observation Window: Only 42 patients were fully included in the proteomic analysis, and the observation period was only 12 weeks. While this small sample withstood rigorous permutation tests and FDR multiple corrections, it is insufficient to support claims regarding long-term alterations in whole-body biological age across the general population.
2. Lack of Healthy Controls and Active Comparators: All trial participants were IPF patients; the independent effect of this drug on proteomic clocks could not be assessed in a healthy elderly population. Furthermore, due to the lack of parallel longitudinal proteomic data from approved antifibrotic agents of the same class (such as pirfenidone or nintedanib), it is currently impossible to determine directly within the trial whether this remodeling of aging clocks and metabolic pathways is unique to Rentosertib (a TNIK inhibitor).
3. Coverage Constraints of the Technology Platform: Although the Olink Explore 3072 panel employed covers nearly 3,000 important serum proteins, the assay panel itself does not contain the target protein TNIK itself. The study could only make indirect inferences through downstream pathways and protein networks, lacking direct omic validation of target binding and local tissue target occupancy. In addition, this study relied solely on serum proteomic data, lacking orthogonal multi-omic validation from tissue biopsies, single-cell transcriptomics, or epigenomics.
4. Disclosure of Conflicts of Interest: Among the authors of this paper, Alex Zhavoronkov is the founder and CEO of Insilico Medicine, and several co-authors are employees of the company. As a commercial entity centered around AI drug discovery, its published research findings require further parallel replication and objective verification by independent third-party academic institutions in follow-up trials.

Future Questions for Verification and Literature Sources

While this study observed modulation signals in aging clocks and metabolic pathways at the serum proteome level, it leaves several critical scientific questions awaiting subsequent validation:

1. Biological Mechanism of the Clock Signal Plateau: Regarding the aging clock signal plateau observed at Week 4, is the underlying cause compensatory feedback following sustained kinase inhibition, exhaustion of susceptible senescent cell populations, or late-stage cellular stress? Would achieving superior geroprotective efficacy necessitate exploring intermittent dosing regimens?
2. Further Decoupling of Antifibrotic and Anti-Aging Effects: Future studies will need larger sample sizes, longer follow-up durations, and incorporation of tissue biopsies or multi-modal clocks (such as epigenomics) to further verify whether changes in blood protein markers translate into systemic physiological functional benefits.