AI Found a Fibrosis Drug and Put It in Humans. Aging Is Next on the Slide Deck.

The full story

Something genuinely new happened in 2025. A molecule that a generative model designed, aimed at a target an AI pipeline nominated, completed a randomized double-blind placebo-controlled phase 2a trial in patients with idiopathic pulmonary fibrosis and reported in Nat Med. That is not a slide deck. That is a drug in people, with a control arm and a blind. Anyone who has watched this field promise things for a decade should register it as a milestone.

Now read the trial's own design, because it says something the coverage mostly did not. The prespecified primary endpoint was safety - the percentage of participants with at least one treatment-emergent adverse event - and the lung-function number everyone quoted was a secondary endpoint measured in 18 patients. Both of those facts are true at once, and holding both is the whole skill required to read this beat.

Three different things get called "AI drug discovery"

Structure prediction is the most mature and the least clinical. AlphaFold, reported in 2021 in Nature, predicted protein structures with atomic accuracy competitive with experimental structures in the majority of CASP14 cases, even where no homologous structure existed. That work is in silico only - no animals, no humans - and the paper makes no claim about drug efficacy or clinical outcomes. Knowing a protein's shape is upstream of everything and equivalent to nothing.

Molecule generation is the famous one, and the famous number needs an asterisk. A 2019 paper in Nat Biotechnol reported that a generative model produced DDR1 kinase inhibitors in 21 days. In cells and mice, only four compounds were active in biochemical assays, only two of those validated in cell-based assays, and exactly one lead was tested for pharmacokinetics in mice. The 21 days covers generating molecules, not developing a drug. It is a real capability, described accurately in the paper and inaccurately almost everywhere else.

Target discovery is the piece most likely to matter for aging, and the hardest to audit. A 2025 paper in Nat Biotechnol describes an AI pipeline nominating TNIK as a fibrosis target and generating the compound against it, going from target discovery to preclinical candidate nomination in roughly 18 months. It also reports safety, tolerability and pharmacokinetics from a randomized double-blind placebo-controlled phase 1 in 78 healthy participants - a phase 1 in healthy volunteers, with no efficacy endpoint, which is exactly what a phase 1 is for.

The phase 2a, read properly

The phase 2a in Nat Med tested rentosertib in 71 patients with idiopathic pulmonary fibrosis across four arms of 17 to 18. Treatment-emergent adverse events - the primary endpoint - were similar across arms at 72.2%, 83.3% and 83.3% against 70.6% on placebo. The number that travelled was a forced vital capacity change of 98.4 mL at 60 mg once daily, with a confidence interval running from 10.9 to 185.9. That was a secondary endpoint in 18 patients, and a confidence interval that wide is telling you so. Discontinuations were driven by liver toxicity and diarrhoea.

The registry entry removes any ambiguity: sponsor InSilico Medicine prespecified percentage of participants with at least one treatment-emergent adverse event as the primary outcome. The lung-function figure was never the question the trial was built to answer. A 2026 network meta-analysis in BMC Pulm Med, covering 35 fibrosis trial reports and 8,983 participants across 21 strategies, then placed the drug among the top three for preserving lung function - while grading the certainty of evidence as mostly low to very low because of imprecision and indirect comparisons, and finding that no regimen significantly reduced all-cause mortality. Promising, provisional, and honest about which it is.

The counterexample nobody puts on a slide

In January 2020, Sumitomo Pharma announced DSP-1181, the first AI-created molecule to enter a phase 1 study, for obsessive-compulsive disorder, with an exploratory research phase that took less than 12 months against a conventional 4.5 years. That speed claim went everywhere. A 2025 critical review in Pharmaceuticals (Basel) records what happened next: DSP-1181 was discontinued after phase 1 despite a favourable safety profile, and the review concludes that the acceleration of discovery timelines by AI does not guarantee clinical success.

That is the sentence to keep. AI compresses the front of the pipeline. The attrition lives in the middle and the end, where biology decides.

Why aging is the hardest possible target for this

Fibrosis was a favourable proving ground: a defined disease, a measurable organ function, a trial that reads out in little more than a year. Aging offers none of that.

There is no approved regulatory endpoint for aging, and the field says so in its own documents. A 2019 review in Trends Pharmacol Sci describes deep-learning predictors of chronological and biological age proposed by their developers for target identification and drug discovery - biomarker predictors, explicitly not validated regulatory outcome endpoints. A 2024 multi-institution consensus paper in Aging (Albany NY) argues that integrating generative AI, biomarkers and clocks is essential for extending healthspan while simultaneously treating those biomarkers and clocks as an unfinished problem; it does not claim a validated aging endpoint exists. Without one, there is nothing for a generative model to optimise toward that a regulator would accept.

The architecture problem is similar. A 2019 review in Ageing Res Rev describes the AI-for-aging pipeline, as characterised by its own proponents, as a chain of generative-adversarial and reinforcement-learning steps producing candidate targets, synthetic molecular data and candidate geroprotectors - work that is in silico only, with the biological validation step entirely downstream. And a 2025 review in Acta Pharm Sin B on AI in Alzheimer's and delirium drug development notes that intricate biological mechanisms have produced numerous clinical trial setbacks in exactly those indications. The translation failure AI is being asked to solve is the failure it inherits.

What the evidence does not show

No AI-discovered compound has been tested against an aging endpoint in humans. Not one. There is no aging trial, no aging readout, no aging safety database. The phase 2a that made the headlines was in a specific lung disease, with a safety primary endpoint, in 71 patients. Nothing here demonstrates that an AI pipeline picks targets that survive contact with human biology better than chemists do - only that it reaches candidate nomination faster, a different and lesser claim. The regulator is roughly where the science is: the FDA's drug centre drew on experience with over 500 submissions containing AI components between 2016 and 2023 for its 2025 guidance on using AI in regulatory decision-making - still at draft stage.

How to read the next headline about this

• Ask which of the three things happened. A predicted structure, a generated molecule and a nominated target are wildly different achievements. Press releases blur them deliberately.
• Find the primary endpoint and check whether it was safety. A safety-primary trial that reports an efficacy number has produced a hypothesis, not a result. Both can be true; only one can be quoted as a finding.
• Count the patients behind the headline figure. An effect measured in 18 people in one arm, with a confidence interval spanning an order of magnitude, is a reason to run the next trial.
• Check whether the speed claim covers discovery or development. Twenty-one days to generate molecules and eighteen months to candidate nomination are front-of-pipeline milestones. Approval is a decade downstream, and most candidates never arrive.
• For aging specifically, ask what the endpoint would even be. If the answer is a clock or a biomarker panel, no regulator has accepted it, and the trial being described is a biomarker study.

The interesting thing about this field right now is that its actual achievement and its marketing achievement point in different directions. A generative model produced a molecule that cleared a phase 1 in healthy volunteers and a phase 2a with a safety primary endpoint, and that is a real, checkable, hard-won result. Another AI-created molecule with a favourable safety profile was discontinued after phase 1, and that is equally real. Aging asks for something neither has delivered: an endpoint a regulator will accept, over a timescale nobody wants to fund, in a population that takes decades to develop the outcome you care about. Computation is not the rate-limiting step there. It never was.