AI-Designed Drugs Just Cleared Phase 2. Phase 3 Is Next

The First AI-Designed Drug Just Cleared Phase 2. Phase 3 Is Next

For two decades, the pharma industry chased a question that sounded simple but proved nearly impossible to answer: can a generative AI model identify a disease target, design a molecule against it, and produce a drug that actually works in humans? In June 2025, the answer arrived in *Nature Medicine*, and in early 2026 it became a regulatory storyline that no large pharmaceutical company can afford to ignore.

Insilico Medicine’s rentosertib, formerly known as ISM001-055, is the first drug in history where both the target (a kinase called TNIK) and the molecule itself were identified and designed end-to-end by generative AI, with no human hypothesis guiding either step. Phase 2a results in idiopathic pulmonary fibrosis (IPF) showed patients on the highest dose gained a mean of 98.4 milliliters of lung function. Patients on placebo lost 62.3 milliliters. That is a 160-milliliter swing in a disease where every approved therapy only slows the rate of decline.

Rentosertib is now heading into a Phase 3 trial in China with 500-plus patients and a Phase 2b in the United States with 200-plus, with regulatory engagement underway. It is one of roughly 173 AI-discovered drug programs in clinical development worldwide, and industry forecasts suggest 15 to 20 of those will enter key Phase 3 trials before the end of 2026.

This is the year the AI drug discovery thesis stops being a thesis.

Why Rentosertib Matters Beyond IPF

The temptation is to read the rentosertib data as a single drug story. That misses the point. The technical claim that Insilico has been making since 2021, and that competitors like Iambic, Generate Biomedicines, Recursion, and Insitro have made in different forms, is that AI can compress the early stages of drug discovery from five years to under eighteen months while still producing molecules that survive human biology.

Rentosertib is the first datapoint that tests the second half of that claim. Compressing chemistry is the easy part. Producing a molecule that actually moves a clinically meaningful endpoint in patients is the hard part. The Phase 2a result does that.

The mechanism is also notable. TNIK was not on most pulmonologists’ radar as an IPF target. Insilico’s PandaOmics platform identified it by mining 1.5 million publications, 9 million patents, 3 million grants, and a long list of multi-omics datasets. The Chemistry42 platform then designed and prioritized molecules against the target. A team of human medicinal chemists refined the lead, but the conceptual leaps came from the model.

If Phase 3 confirms the Phase 2a signal, every kinase that has been parked as undruggable, every disease without a clear genetic anchor, and every rare condition where there is too little data for a traditional pharma program becomes a candidate for the AI-first workflow.

The 2026 Pipeline: Who’s Actually in the Clinic

The 173 figure that gets quoted in industry decks hides a wide quality range. Most of those programs are early-stage assets that may never see Phase 2. The companies with assets that matter for 2026 are a much shorter list.

Insilico Medicine has rentosertib heading into Phase 3 plus a USP1 inhibitor, ISM3091, in Phase 1 oncology trials and at least eleven additional clinical-stage programs.

Iambic Therapeutics combines physics-based AI with a high-throughput wet lab that generates new biological data on a weekly cycle. The company’s lead programs in oncology and a discovery collaboration with Takeda focus on what the industry calls “undruggable” targets — proteins where structure-based design has failed for decades.

Generate Biomedicines uses generative protein models to design biologics, including monoclonal antibodies. Several programs are in IND-enabling studies and the first Phase 1 entries are expected this year.

Recursion Pharmaceuticals runs an industrial-scale phenomics platform that has generated more than 65 petabytes of cellular imaging data. Its lead asset, REC-994 for cerebral cavernous malformation, has Phase 2 data, and a deep partnership with Roche covers an additional 40 oncology and neuroscience programs.

Insitro focuses on machine-learning-driven target discovery using human genetics. Its first clinical asset is now in Phase 1 and a partnership with Bristol Myers Squibb covers neurodegeneration programs.

The point is not that any one of these programs will succeed. The point is that the failure rate in Phase 2 for AI-designed drugs in 2026 will determine the cost of capital for AI-first biotech for the next decade. A few clean wins, and capital floods in. A string of failures, and the sector compresses.

What the FDA Is Doing About It

Regulators do not like the word “black box.” That has been the central friction in AI drug development since the first ML-derived programs entered the clinic. The FDA’s response is the AI/ML credibility framework.

In January 2025, the agency published its first draft guidance on AI in drug development. The framework is risk-based: the higher the regulatory consequence of an AI-derived decision, the more rigorous the credibility evidence required. A model used to prioritize candidates in early discovery faces a low bar. A model whose output is used to support a clinical safety claim faces a much higher bar, including independent validation, documented training data provenance, and ongoing performance monitoring.

The final guidance is expected in Q2 2026. That is the calendar quarter that will define how AI-designed drug applications get reviewed for the rest of the decade. Sponsors that have been preparing under the draft framework — including Insilico — are positioning for the first wave of NDAs that explicitly cite AI in target ID and molecule design.

The European Medicines Agency is moving in parallel. The EMA’s reflection paper on AI in the medicinal product lifecycle, updated in late 2025, takes a similar risk-tiered view but with an additional emphasis on data minimization and post-market surveillance.

This is the part of the story that gets undersold. The reason 2026 matters is not just that the science is delivering. It is that the regulatory plumbing is being built in real time to handle the science, and the companies that align early will have a structural advantage.

The Economics: $6 Million Models Beating $100 Million Drugs

Pharma’s traditional cost curve assumes that bringing a new molecule into Phase 1 costs somewhere between $100 million and $300 million, accumulated over five to seven years of preclinical work. AI-first programs are running an order of magnitude below that.

Insilico’s published cost for rentosertib’s preclinical campaign was approximately $6 million from target identification to IND filing, executed in roughly 30 months. Other companies cite similar numbers. Iambic claims it can take a target from in silico hit to development candidate in under twelve months. Recursion has industrialized the screening side enough that adding a new disease area to its phenomics map is largely a question of cell type acquisition, not new infrastructure.

If those numbers hold across a broader sample, the economics of pharma R&D shift. The question stops being “can we afford to develop this drug?” and becomes “can we afford not to develop this drug?” Indications that were too small for traditional pharma — ultra-rare diseases, narrow oncology slices, geographic or demographic subsets — become viable.

This is also the implicit threat to the contract research organization industry. If half of the preclinical workflow can be done with a wet lab attached to a model rather than a 200-person CRO program, the CRO model has to change.

What Could Go Wrong

The bear case is straightforward. AI-designed drugs may have the same Phase 2 and Phase 3 attrition rates as conventional drugs once you control for therapeutic area and target novelty. If that turns out to be true, the speed advantage in early discovery does not translate into a higher overall success rate, and the cost savings get eaten by the same late-stage failures that have always plagued the industry.

There is also the data quality question. Models trained on public chemistry and biology databases inherit every bias and error in those datasets. A target that looks AI-validated against historical data may simply be a target the literature happens to like, and the model is recommending it because the model has read everything the field already believes.

And there is the talent and IP question. The companies with the best models are not the largest pharma companies. The largest pharma companies have responded with a wave of partnerships and acquisitions, but it is not yet clear whether the integration works or whether the culture clash kills the value before it can be captured.

What to Watch Through 2026

Three datapoints will determine whether the AI-first thesis holds.

The first is the rentosertib Phase 3 readout, which should arrive in 2027 or 2028 if the trial starts on schedule this year. Until then, the next two years of clinical updates will set sentiment.

The second is the FDA’s final AI/ML guidance in Q2 2026 and how the first AI-designed drug NDAs are reviewed under it. The agency has signaled flexibility, but flexibility in regulatory practice is hard to predict in advance.

The third is the broader pipeline performance. If even five of the 15 to 20 AI-designed Phase 3 starts produce positive readouts, the field has a new normal. If three or fewer succeed, the consensus narrative resets.

For coverage of related shifts in clinical AI, see IssueByte’s reporting on AI liquid biopsy, AI pharmacogenomics, and AI drug repurposing. The clinical-trial side of the story is the one most people in tech overlook, and it is where the real enterprise value of AI in healthcare is being settled this year.