Did an AI-Designed Drug Really Make Patients Biologically Younger?

The headline sounds ridiculous

An AI-designed drug made people:

“3–6 years younger.”

That sounds like the sort of longevity headline we would normally approach very cautiously.

And we should.

But the underlying study is genuinely interesting.

The drug is called rentosertib. It was not developed as an anti-ageing drug. It was developed for idiopathic pulmonary fibrosis, or IPF — a progressive disease in which lung tissue becomes scarred, making breathing increasingly difficult.

Researchers had already tested rentosertib in a 12-week Phase 2a clinical trial. They then went back to stored blood samples and asked a different question:

Did the protein patterns in these patients also look biologically younger?

That is where the story gets interesting — and where the wording needs care.

First: what is rentosertib?

Rentosertib is an experimental small-molecule drug that targets TNIK, short for TRAF2- and NCK-interacting kinase. Insilico Medicine used artificial-intelligence systems in the drug-discovery process, but the drug was developed through the usual human sequence of experiments, synthesis, preclinical work and clinical trials.

The original Nature Medicine trial was multicentre, randomised, double-blind, placebo-controlled and 12 weeks long. The 71 randomised participants received:

• 30 mg rentosertib once daily: 18 participants;
• 30 mg rentosertib twice daily: 18 participants;
• 60 mg rentosertib once daily: 18 participants; or
• placebo: 17 participants.

The trial was designed primarily to examine safety, tolerability, biological activity and lung-related outcomes. It was not designed as an anti-ageing trial.

So where did the “age reversal” claim come from?

The Nature Biotechnology paper analysed stored serum samples from the clinical trial. The proteomic dataset contained measurements for 2,841 proteins. The ageing-clock analysis used baseline serum samples from 42 trial participants with suitable data.

Researchers then applied six published proteomic ageing clocks. Their labels are not important for remembering the story, but they included:

• ProtAge;
• OrganAge chronological;
• OrganAge mortality;
• the proteomic ageing clock, or PAC;
• ipfP3GPT; and
• PAOPAC.

These clocks were developed differently and use different combinations or interpretations of protein data. Four were chronological-age clocks and two were trained around mortality or organ-specific risk. That distinction matters: a model trained to reflect mortality risk is not simply measuring the same thing as a model trained to estimate calendar age.

What is a proteomic ageing clock?

Your blood contains thousands of proteins. Their levels change with age, disease, immune activity, metabolism, organ function and many other biological processes.

Researchers can train statistical or machine-learning models to recognise patterns associated with chronological age, organ function or mortality risk. The resulting model can produce a predicted biological age.

But this is not like measuring blood pressure or body temperature. It is a model output: a calculated estimate based on patterns in biological data.

A biological-age clock does not measure age directly. It estimates age from patterns associated with ageing.

There is no universally agreed single “true biological age”. A person might simultaneously have a DNA-methylation age, a proteomic age, an organ-specific age, an immune age and a functional age that do not perfectly agree. That does not automatically make every clock useless. It means ageing is multidimensional.

Your body probably does not have one hidden biological birthday waiting to be discovered.

What did the six clocks show?

Across the six clocks, treatment arms generally showed reductions in predicted biological age relative to baseline, while the placebo group showed little change or slight increases over the 12-week period.

The researchers compared each treatment regimen with placebo at weeks 2, 4 and 12 across all six clocks. That produced 54 treatment-versus-placebo comparisons per regimen. Of those, 21 reached the paper's statistical significance threshold of Q < 0.10 after false-discovery correction.

The strongest cluster was at week 4: 11 of 18 comparisons showed significantly lower predicted biological age in treated participants. The 30 mg twice-daily regimen produced the most consistent signal, with nine significant comparisons, compared with seven for 60 mg once daily and five for 30 mg once daily.

A patient-level permutation test suggested that 21 significant comparisons exceeded what would be expected under the shuffled-label null distribution. That strengthens the case that the pattern was not simply random noise. It still does not turn a post-hoc biomarker analysis into proof of age reversal.

Where does the 3–4 years — or six years — come from?

Some of the strongest week-4 comparisons were summarised as roughly 3–4 years of lower predicted biological age. One clock produced a shift of around six years in some comparisons.

Those are model outputs. They mean that the measured blood-protein pattern moved in a direction that the particular model associates with a younger age.

They do not mean that a 70-year-old became physiologically equivalent to a typical 64-year-old in every meaningful way.

Six years younger on a clock is not the same thing as six extra years of life.

The artery-clock result looks even more dramatic

The organ-specific mortality models produced some larger numerical shifts. The artery-related clock showed significantly younger predicted ages across treated groups and timepoints, with reported shifts of roughly 6.95 to 16.57 years relative to placebo in the relevant comparisons.

That number sounds spectacular. It is also exactly where biomarker headlines can become misleading.

Organ-specific ageing clocks are experimental models. Large numerical shifts can reflect how a model is constructed and calibrated. A “16-year younger artery clock” is not equivalent to a validated reduction in heart attacks, stroke or vascular disease.

The safe wording is younger predicted age on an experimental artery-related model, not “16 years younger arteries”.

Six clocks agreeing is still interesting

It would be wrong to dismiss the finding. One ageing clock moving could be noise, model-specific behaviour or a statistical artefact. Six differently constructed clocks shifting in the same general direction is more interesting.

It suggests that treatment caused a broad change in circulating proteins associated with age-related patterns. That consistency is potentially important.

But agreement between models is not the same as clinical validation. The clocks partly measure overlapping biology, and none can independently determine whether the change reflects disease improvement, ageing biology, or both.

The biggest problem: was this ageing or was the lung disease improving?

This is the central scientific limitation.

Idiopathic pulmonary fibrosis itself affects inflammation, fibrotic signalling, metabolism and circulating proteins. If rentosertib improves disease biology, those proteins may begin looking “younger” to an ageing clock. That does not necessarily mean that the underlying ageing process changed.

The Nature Biotechnology authors are explicit about this limitation: proteomic clocks alone cannot deconvolute ageing-specific effects from disease-specific effects. That sentence deserves more attention than the “six years younger” headline.

The researchers did try to address that

The team did not stop at “the clock score went down”. They also performed pathway analyses and found treatment-related changes involving biological processes associated with cellular senescence, metabolism, growth-factor signalling, antioxidant processes and cholesterol metabolism, alongside expected anti-fibrotic effects.

The 30 mg twice-daily regimen also showed a significant negative correlation with normal age-related protein trajectories derived from 55,319 older adults in UK Biobank. In simpler terms, some proteins changed in the opposite direction to the way they normally change with age.

That provides supportive biological evidence. It still does not prove human age reversal. Pathway-level consistency can make a result more compelling while leaving the core disease-versus-ageing question unresolved.

The dose result is particularly interesting

The 30 mg twice-daily regimen produced the most consistent ageing-clock signal. But 60 mg once daily had the stronger lung-function response in the original clinical analysis, including the clearest forced vital capacity signal compared with placebo.

This weakens the simplest explanation that “the ageing clocks only improved because lung function improved”. It does not eliminate that possibility, and dose-response relationships in small exploratory datasets can be unstable. The observation needs replication.

What did the original trial actually show?

The ageing-clock paper should not overshadow the actual clinical trial. The Phase 2a Nature Medicine study evaluated safety and exploratory efficacy in 71 people with IPF. Fifty-five participants completed the 12-week placebo-controlled period, and 16 discontinued before the end of treatment.

Treatment-emergent adverse events were reported in substantial proportions of all treatment groups and placebo: 83.3% in the 30 mg twice-daily group, 83.3% in the 60 mg once-daily group, 81.3% in the 30 mg once-daily group and 70.6% in the placebo group. Treatment-related serious adverse events were low and comparable across groups. The most common events leading to discontinuation were related to liver toxicity or diarrhoea.

The 60 mg once-daily group had a mean forced vital capacity change of +98.4 ml, compared with −20.3 ml for placebo, but the authors described the result as warranting further investigation in larger and longer trials.

Phase 2a studies help researchers understand safety, dose, biological activity and early signals of efficacy. They are not definitive proof that a drug improves long-term clinical outcomes, still less proof that it extends healthspan or lifespan.

This was a post-hoc, secondary analysis

The blood samples already existed. Researchers later asked a different question of them. That makes the ageing-clock work exploratory.

Exploratory analyses are not worthless, but they are more vulnerable to multiple comparisons, chance findings, model choices and selective interpretation. This paper used false-discovery correction and permutation testing, which strengthens confidence that the observed pattern was not simply random. Replication remains essential.

The sample size is small

The ageing-clock analysis involved 42 participants with suitable baseline serum data. That is tiny compared with the evidence needed to establish a longevity intervention.

Small samples can exaggerate effect sizes, produce unstable estimates and be disproportionately influenced by individual participants. The right summary is:

The Age Watchers Evidence Ladder

The Evidence Ladder is an Age Watchers editorial framework, not a validated scientific grading system. Applied to this story:

Age Watchers claim check

The AI part is actually two stories

AI appears in this research in two different roles.

1. Discovery: AI-based systems helped identify TNIK as a target and supported the development and optimisation of rentosertib.
2. Measurement: computational models helped interpret patterns in thousands of proteins and produce predicted ageing outputs.

So AI sits at both ends: discovery and measurement. That is genuinely significant for the future of medicine.

But “AI-assisted” is more accurate than “AI independently invented a medicine”. Human scientists designed experiments, synthesised compounds, ran preclinical work, designed trials, treated patients and analysed results.

The bigger idea: dual-purpose clinical trials

The most important result may not be that one drug made ageing clocks move. It may be that future drug trials start measuring ageing at all.

Traditionally, a lung-drug trial measures lung outcomes and a diabetes-drug trial measures diabetes outcomes. The Nature Biotechnology paper points towards a more ambitious possibility: trials for age-related diseases could include ageing-related endpoints alongside safety, symptoms, lung function or other disease measures.

An IPF trial might measure lung function and safety, plus proteomic ageing, epigenetic ageing, physical function, frailty and other validated ageing biomarkers. That could help researchers investigate whether a disease treatment also modifies broader ageing biology.

It could make geroscience more efficient. Humans live a long time, and a trial designed to prove that a drug extends lifespan could take decades. Researchers therefore need intermediate endpoints that are validated against outcomes people actually care about.

Potential candidates include proteomic clocks, epigenetic clocks, frailty, physical function, immune markers and organ-specific biomarkers. But measurable does not automatically mean clinically useful. These endpoints only become valuable if they reliably predict better health, better function, fewer hospitalisations, less disability or longer life.

Biological age may be more useful in research than in a personal score

Age Watchers has deliberately resisted telling members: “You are 57 but biologically 49.” Commercial biological-age scores can create false precision.

This study illustrates a potentially better use. Biological clocks may be more valuable for comparing groups in clinical trials than for telling one person exactly how old their body “really” is.

That is also why this article does not add a biological-age score to the Age Watchers dashboard. Our member approach remains function and actionable measures first: blood pressure, strength, fitness, activity, waist and weight trends, sleep, resting heart rate, functional capacity and appropriate clinical biomarkers.

Biological-age clocks may eventually be useful if they become validated, reproducible, actionable, affordable and clinically interpretable. Until then, they belong in research coverage rather than the core dashboard.

Do not major in the minor

Someone concerned about ageing should not respond to this paper by looking for rentosertib, proteomic-clock testing or unapproved anti-ageing drugs while ignoring smoking, blood pressure, activity, strength, sleep, diet, social connection and appropriate medical care.

The frontier is exciting. The fundamentals remain more actionable.

Commercial involvement is context

The Nature Biotechnology paper declares that Alex Zhavoronkov is founder and CEO of Insilico Medicine, that several authors are Insilico employees and that Insilico developed rentosertib.

This does not automatically invalidate the study. It does mean that readers should know the company developing the drug is closely involved in the research. The paper was peer reviewed by Nature Biotechnology, but peer review is not the same as independent replication.

Follow the evidence. Follow the money. Do not confuse the two.

What would actually prove something big?

We would want to see:

1. Larger trials: more participants and more stable effect estimates.
2. Pre-specified ageing endpoints: not only a post-hoc analysis of stored samples.
3. Longer follow-up: do clock changes persist?
4. Functional outcomes: do participants walk better, become less frail or remain healthier?
5. Clinical outcomes: fewer diseases, hospitalisations, disability or deaths.
6. Independent replication: other researchers, drugs and diseases.
7. Healthy and broader populations: if a general geroprotective effect is being claimed.

Until then, we have an intriguing biomarker signal — not an anti-ageing treatment.

What should you do with this information today?

Nothing medically.

Do not seek rentosertib, buy biological-age tests, alter treatment or look for unapproved anti-ageing drugs because of this article. Rentosertib is an investigational drug being developed for idiopathic pulmonary fibrosis and is not an approved anti-ageing treatment.

Instead, follow the science and understand why the research matters. Continue using interventions already supported by much stronger human evidence.

When someone says “biological age reversed”

The exciting part

This paper is not proof of human age reversal. But it demonstrates several potentially important developments at once:

• AI-assisted drug discovery has reached meaningful human trials.
• Researchers can measure thousands of proteins simultaneously.
• Multiple ageing models can be applied to ordinary clinical-trial samples.
• Ageing biology can potentially be studied alongside conventional disease treatment.

Instead of waiting decades for dedicated longevity trials, researchers may eventually be able to ask two questions at once:

Did this treatment help the disease?

And did it also shift biology associated with ageing?

That could accelerate geroscience enormously.

Conclusion

Did rentosertib make people younger?

We cannot honestly say that.

What we can say is more precise: in a small exploratory analysis of a randomised lung-disease trial, six proteomic ageing clocks shifted towards younger predicted biological ages in people receiving the drug. Some age-related protein pathways moved in the opposite direction to normal ageing.

That deserves attention.

But the clocks cannot yet tell us whether ageing itself changed, the lung disease improved, or both happened together.

The real breakthrough may therefore be methodological rather than pharmaceutical. We may be learning how to look for ageing effects inside ordinary clinical trials.

Interesting? Absolutely.

An anti-ageing pill? Not yet.

MEASURE.
UNDERSTAND.
IMPROVE.
MAINTAIN.

Sources / Further reading

• Nature Biotechnology: Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment.
• Nature Medicine: A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial.
• Insilico Medicine: rentosertib development context. Company material is included for context, not independent validation.
• UK Biobank, including the proteomic reference cohort used for normal ageing trajectories.
• How Well Are You Really Ageing?
• Can AI Help Older People Age Well?
• Start With a Baseline
• Chronic Stress, Ageing and Allostatic Load