The First Drug Designed Entirely by AI Just Improved Lung Function in Humans. Here's What That Actually Means.

The Molecule That Shouldn't Exist Yet

In January 2025, a paper appeared in Nature Medicine describing the Phase 2a results for a drug called rentosertib (ISM001-055, later INS018_055). Seventy-one patients with idiopathic pulmonary fibrosis, a disease that gradually turns lung tissue into scar tissue, received either the drug or a placebo for 12 weeks. In the highest dose group, lung function improved by 98.4 mL as measured by forced vital capacity. Placebo patients declined by 20.3 mL. The separation was statistically significant at p<0.001.

Those numbers require context. IPF is a disease where the standard trajectory is decline. Patients don't get better. The two existing FDA-approved treatments, pirfenidone and nintedanib, slow the deterioration. They don't reverse it. A 98.4 mL improvement in 12 weeks, if it holds, would represent something qualitatively different from current therapy.

But the numbers are not why this trial matters to anyone outside pulmonology. Rentosertib matters because of how it was found. Both the disease target, a kinase called TNIK (TRAF2 and NCK-interacting kinase), and the molecule itself were identified by generative AI systems built by Insilico Medicine, a Hong Kong-based company. No human chemist proposed TNIK as a fibrosis target. No human medicinal chemist designed the compound. The entire pipeline from target hypothesis to preclinical candidate took approximately 18 months.

Traditional drug discovery timelines for the same journey: four to six years. Traditional cost to reach a preclinical candidate: roughly $400 million on average across the industry. Insilico's reported spend through the preclinical stage: $2.6 million. A caveat: that figure represents the marginal cost of one program running on platforms (PandaOmics, Chemistry42) that cost substantially more to develop. The per-program economics improve with every additional candidate — which is itself part of the thesis.

What AI Actually Did Here

Insilico used two of its internal platforms. PandaOmics, a biological target discovery engine, analyzed omics datasets (transcriptomic, proteomic, epigenomic) to identify TNIK as a novel target implicated in fibrotic pathways. Chemistry42, a generative chemistry platform, then designed molecules optimized to inhibit TNIK with the right pharmacokinetic properties: oral bioavailability, metabolic stability, selectivity.

This is worth being precise about, because most "AI drug discovery" headlines describe something different. Many AI-assisted programs use machine learning to optimize a molecule that humans already identified against a target that humans already validated. Insilico's claim is more specific: both the target and the molecule were AI-originated. TNIK was not in the existing IPF literature as a therapeutic target. The molecule had no close analogs in known chemical space.

That distinction matters. Finding new targets and designing new molecules from scratch is the hardest creative step in drug discovery. It is also, unfortunately, the step that has the least to do with whether a drug ultimately works in humans.

The Honest Assessment

Drug development has a famous geometry. It is shaped like a funnel, and the funnel is brutal. Approximately 90% of drugs that enter Phase 1 clinical trials never reach approval. The attrition rate from Phase 2, the stage rentosertib is now entering at scale, is roughly 70%. These numbers have been remarkably stable for decades, through every previous revolution in drug design: combinatorial chemistry, high-throughput screening, structure-based design, fragment-based discovery.

AI appears to be improving the early stages. A comprehensive review published in Pharmacological Reviews found that AI-designed molecules show 80-90% success rates in Phase 1 trials, compared to roughly 52% for historically developed compounds. That's a real improvement: it suggests AI is better at designing molecules that are safe and well-tolerated in humans, which is not trivial.

But Phase 1 primarily measures safety, not efficacy. It answers "does this molecule poison people?" not "does it cure the disease?" The harder question, does the drug work, is answered in Phase 2 and Phase 3. And here, early evidence suggests AI drugs fail at roughly the same rate as everyone else's drugs.

As of early 2026, there are 173+ AI-originated drug programs in clinical development across the industry. Zero have received FDA approval. The first approval is projected for 2026-2027, but projections in pharma are aspirational by nature. Rentosertib itself still faces Phase 2b and Phase 3 trials before any approval decision.

The $2.6 Million Question

If AI doesn't yet change the probability that a drug works, what does it change? Two things: speed and cost at the front end.

Insilico's 18-month, $2.6 million path to a preclinical candidate is genuinely remarkable when set against industry averages. If you can fail faster and cheaper, you can run more experiments. More experiments mean more shots on goal. More shots on goal, at the same per-shot success rate, still mean more total successes. The math is straightforward even if the biology isn't.

This is the sober version of the AI drug discovery thesis. It is not that AI designs better drugs. It is that AI lets you try more drugs, at lower cost per attempt, with faster iteration cycles. The funnel stays just as narrow. You just pour more candidates into the top of it.

Insilico has 31 programs in its pipeline, across fibrosis, oncology, immunology, and CNS. No single company using traditional methods would have 31 parallel programs at this stage of development. The portfolio strategy itself is the innovation.

The 33% Signal

Buried in the rentosertib Phase 2a data is a number that deserves attention: 33% of patients in the highest dose group discontinued treatment due to liver enzyme elevations. This is a safety signal. It is the kind of finding that can narrow or kill a program in later trials, where the risk-benefit calculus gets stricter and the patient populations get larger.

It does not invalidate the results. Drug programs routinely adjust dosing based on Phase 2 safety data, and the lower dose groups showed better tolerability with still-meaningful efficacy. But it is a reminder that the molecule AI designed, however creatively it was found, must still survive the same gauntlet every other molecule faces: proving it is safe enough, effective enough, and manufacturable enough at scale.

AI contributed nothing to the liver toxicity problem. It contributed nothing to the manufacturing challenge. It contributed nothing to the regulatory pathway. Those constraints are biological and institutional, and they are where most drugs actually die.

The Landscape Beyond Insilico

Rentosertib is not alone. In July 2025, Recursion Pharmaceuticals reported Phase 2 data for REC-994, an AI-identified compound for cerebral cavernous malformations, showing roughly 50% reduction in brain lesion volume. Recursion had also completed a $688 million merger with Exscientia, creating the largest AI drug discovery company by market capitalization.

Absci is using generative AI not for small molecules but for de novo antibody design, a harder problem with potentially higher rewards. Their platform generates novel antibody candidates from scratch, with no starting template, with claimed hit rates of 10-20% versus <1% for traditional screening. If those numbers hold in clinical translation, antibody development timelines could compress by years.

The FDA itself published draft guidance in early 2025 acknowledging AI's role across the drug development lifecycle, from target identification through clinical trial design. The regulatory infrastructure is adapting, slowly, but it's moving.

Meanwhile, the broader numbers remain sobering. An IntuitionLabs analysis of AI drug pipelines found that of all AI-originated drug candidates that entered clinical development, 97% had not progressed beyond Phase 2. The revolutionary technology, in other words, is producing revolutionary candidates that then fail at historically normal rates.

What We Don't Know

This analysis has significant blind spots. Insilico's $2.6 million cost figure is self-reported and excludes platform R&D amortization; independent verification is not available. The 80-90% Phase 1 success rate for AI drugs comes from a relatively small sample compared to the historical dataset, and selection bias is likely — companies with failed AI candidates have less incentive to publish. The 173+ pipeline count includes programs at widely varying stages of maturity, many of which are preclinical and may never reach human testing. Most critically, the question of whether AI-designed molecules are finding genuinely novel biology or simply rediscovering known chemical space faster remains open. If the targets AI identifies turn out to be the same ones human researchers would have found eventually, the contribution is efficiency, not discovery — a meaningful but narrower claim.

What This Actually Means

Rentosertib's Phase 2a results are a milestone by any reasonable definition. A fully AI-originated drug — novel target, novel molecule, no human hypothesis at the origin — demonstrated meaningful efficacy in a randomized, placebo-controlled trial in humans. That sentence was not possible two years ago.

It is also a milestone that illuminates the limits of what AI currently contributes. Drug discovery has always had two distinct problems: finding candidates (creative, computationally intensive, amenable to pattern recognition) and proving candidates work in the irreducible complexity of human biology (slow, expensive, stubbornly resistant to computational shortcuts).

AI is solving the first problem. It has not yet touched the second one. A 98.4 mL improvement in FVC is a reason for cautious optimism, not celebration. The celebration, if it comes, is years of Phase 2b and Phase 3 data away.

What's changed is the economics of trying. If AI can turn a $400 million, six-year preclinical program into a $2.6 million, 18-month one, the implications are profound even if the failure rate stays the same. More targets explored. More mechanisms tested. More shots at diseases that traditional economics deemed too risky to pursue. Rare diseases, complex multi-target conditions, antibiotic resistance — categories where the expected return couldn't justify the traditional cost of the search.

Rentosertib might work. It might not. The liver toxicity signal might be manageable, or it might not. But the 30 programs behind it in Insilico's pipeline, and the hundreds behind those across the industry, represent something that actually is new: a drug discovery apparatus that can afford to be wrong at unprecedented scale.

That's the real proof of concept. Not one molecule's Phase 2a data. The economics that made it possible to find that molecule in the first place.