Trial by Fire

In a hospital in China, a patient with idiopathic pulmonary fibrosis — a disease that quietly turns the lungs to scar tissue — took a small pill twice a day. The molecule inside it had never existed in nature, or in any chemist's notebook. Both the disease target it aimed at and the drug itself had been dreamed up by software: a generative artificial intelligence that proposed a protein worth attacking and then designed, from scratch, a compound to attack it. After twelve weeks, the patients on the highest dose were not just declining more slowly. Their lung function had measurably improved. For the first time, AI had walked a medicine from an empty screen to a real benefit in a real human body — and that was only the opening act of what 2026 has become.

This is the year the experiment grows up. After a decade of dazzling demos and breathless promises, roughly fifteen drugs discovered or designed with the help of AI have entered Phase III clinical trials — the final, largest, and most punishing stage of human testing, where a candidate must prove not merely that it is safe and biologically plausible, but that it actually helps thousands of sick people more than a placebo or the existing standard of care. Phase III is where most drugs die. It is also the only door to approval. And so 2026 has become a referendum: was AI drug discovery a genuine revolution in how we find medicines, or a very expensive way of arriving at the same disappointments faster?

The drug that proved the concept

The story above belongs to rentosertib, a compound from the Hong Kong– and New York–based company Insilico Medicine. Its Phase IIa results, published in Nature Medicine, became the field's foundational milestone for a simple reason: it was the first drug in which both the biological target and the molecule itself were identified and designed by a generative AI platform. The target is an enzyme called TNIK, surfaced by Insilico's software as a node worth disrupting in the tangled biology of fibrosis. The molecule, a TNIK inhibitor, was generated and refined by the same system.

In the placebo-controlled GENESIS-IPF trial, 71 patients across 22 sites in China were randomized to rentosertib or placebo. Those on the 60-milligram once-daily dose saw their forced vital capacity — the standard measure of how much air a damaged lung can move — improve by a mean of 98.4 milliliters over twelve weeks, while the placebo group declined by 20.3 milliliters. In a disease defined by relentless, irreversible decline, an actual improvement is striking. Exploratory biomarker analyses backed up the mechanism, suggesting the AI had not merely stumbled onto a lucky chemical but had correctly reasoned about the underlying biology.

It is essential to be precise about what this does and does not prove. Phase IIa is a small, early efficacy hint, not a verdict. Rentosertib still has to clear a pivotal Phase III before anyone speaks seriously about approval, which is unlikely before 2028 at the earliest. But the result reframed the debate. The question was no longer whether AI could generate a plausible-looking molecule on a screen — it plainly could — but whether such a molecule could survive contact with human biology. Rentosertib delivered the first real evidence that it might.

A whole pipeline comes of age

Rentosertib is the furthest along, but it is far from alone. The field has matured from a single hopeful candidate into a genuine pipeline. Recursion Pharmaceuticals, which built its reputation on industrial-scale automated biology — robots running millions of cellular experiments to train its models — filed a new investigational-new-drug application this spring for REC-7221, a CDK4/6 inhibitor aimed at solid tumors, with primary trial completion targeted for 2027. Isomorphic Labs, the drug-discovery company spun out of Google DeepMind on the back of the AlphaFold protein-structure breakthrough, has signaled that its own AI-designed candidates should begin human trials by the end of this year.

Behind these names sits a broader shift that industry analysts have been tracking closely. By the start of 2026, well over a hundred AI-driven drug programs were in active clinical development across the sector, and surveys found that roughly 80 percent of biopharma organizations planned to increase their AI budgets within the year, with nearly a quarter expecting to double their spending or more. The center of gravity has moved from isolated digital tools bolted onto traditional R&D toward what the industry now calls "AI-native" discovery — companies redesigning their data, their labs, and their org charts so that machine learning is the default, not an add-on.

Where the skeptics are right

For all the momentum, sober voices inside the field urge caution, and they have good reasons. The first is that AI's claimed advantages have mostly been demonstrated at the early stages — generating candidates, predicting structures, prioritizing targets. The hard, slow, expensive part of drug development has always been the clinic, and there is no shortcut through a three-year, multi-thousand-patient trial. An AI that gets you to Phase I faster has not necessarily improved your odds of surviving Phase III. The real test of the technology is not speed but the success rate — whether AI-discovered drugs ultimately win approval more often than the dismal industry baseline, where roughly nine of ten candidates that enter human trials fail.

The second caution is about what "AI-discovered" even means. The label is applied generously, sometimes to drugs where AI played a supporting role and sometimes to those, like rentosertib, where it drove both target and molecule. Pulling apart genuine AI-native breakthroughs from clever marketing will take years and honest bookkeeping. And the third caution is regulatory: the FDA's draft guidance on AI in drug development is expected to be finalized this year, while the EU AI Act's high-risk provisions take effect in August 2026, potentially classifying some drug-development AI as high-risk and subjecting it to new scrutiny. The rules of the road are being written even as the cars accelerate.

Why this matters beyond the lab

Step back from the trial data and the stakes are enormous in human terms. Developing a single new drug the traditional way takes well over a decade and routinely costs more than a billion dollars, and the overwhelming majority of attempts fail. That brutal arithmetic is why thousands of diseases — many rare, many devastating — have no treatment at all: the economics simply never close. The promise of AI drug discovery is not merely faster pills for profitable markets. It is the possibility of bending that cost curve far enough that medicines become viable for conditions the old model left behind.

Idiopathic pulmonary fibrosis is a fitting place for the story to begin. It is relentless, poorly understood, and starved of good options — exactly the kind of problem where a fresh way of reasoning about biology could matter most. Whether rentosertib and its cohort ultimately cross the finish line, the deeper shift is already underway: for the first time, the search for new medicines is being guided by systems that can reason over more biology than any human mind can hold at once.

The verdicts will not all arrive at once. Phase III trials run for years, and 2026's readouts are early chapters, not the ending. Some of these drugs will fail, and their failures will be just as instructive as the successes — teaching the field where its models are still blind. But the direction is set. The molecule the machine made is no longer a thought experiment on a screen; it is in the bloodstream of patients who are waiting, like the rest of us, to find out if the future of medicine works. By the time this year closes, we will know far more than we did when it began. That, in a field built on uncertainty, is its own kind of breakthrough.

Sources

1. Nature Medicine — "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial"
2. Insilico Medicine — "Insilico Announces Nature Medicine Publication of Phase IIa Results of Rentosertib"
3. PR Newswire — "Insilico Medicine Announces Nature Medicine Publication of Phase IIa Results"
4. Drug Discovery Trends — "Insilico's AI-designed rentosertib shows promise in first Phase 2a results"
5. HumAI Blog — "AI-Discovered Drugs Reach Phase III. And 2026 Will Determine Whether All the Promises Were Real."
6. Drug Target Review — "AI in drug discovery: predictions for 2026"
7. Drug Target Review — "2026: the year AI stops being optional in drug discovery"
8. Axis Intelligence — "AI Drug Discovery 2026: 173 Programs, FDA Framework & Market"
9. BuildMVPfast — "AI Drug Discovery: What's Working, What's Failing in 2026"
10. Ardigen — "AI in Biotech: Lessons from 2025 and the Trends Shaping Drug Discovery in 2026"
11. Healio — "Q&A: AI-established drug 'breakthrough in the treatment of IPF'"
12. Pulmonary Fibrosis Foundation — ISM001-055 (rentosertib) pipeline profile