AI-designed drugs have now reached late-stage trials, but none has yet been approved on the strength of a completed phase 3 study. What exists is a handful of real, partly published results, a much larger pile of early-stage programs, and a few well-documented failures. This article separates the three.

The two most useful test cases are quite different. One is rentosertib, a molecule from Insilico Medicine that began a phase 3 trial in September 2026. The other is zasocitinib, a psoriasis pill whose phase 3 data are in hand and which was shaped by physics-based computation rather than generative AI.

What “AI-designed” does and does not mean

The label gets stretched. At one end sit programs where a generative model proposed both the biological target and the molecule. At the other end are drugs where software scored thousands of candidate compounds, and chemists then did most of the remaining work. Both get called AI-discovered in headlines, and trackers count them differently.

That matters when you read totals. A conference abstract in the Journal of Clinical Oncology, summarized by a 2026 industry advisory page, counted 117 AI-enabled assets from 63 companies in interventional trials as of December 2025. Other trackers quoted on the same page put the figure above 170. Neither number is wrong. They simply draw the line in different places, so treat any single count as an estimate.

One more point is easy to miss: US regulators do not flag whether AI was used in discovery, so there is no official register to check. Counts come from companies and trackers.

Rentosertib: the clearest AI-to-patient story

Insilico describes rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF), as discovered and designed with its generative AI platform, covering both the target and the molecule. That is the strongest claim in the field, and it is the reason this drug gets so much attention.

The phase 2a trial (GENESIS-IPF, registered as NCT05938920) was published in Nature Medicine on June 3, 2025. It randomized 71 patients across four arms for 12 weeks: placebo (17 patients) and three rentosertib regimens (18 each). The primary endpoint was safety, and adverse-event rates were similar across arms. Liver toxicity and diarrhea were the most common reasons patients stopped treatment.

Efficacy was a secondary question, and the numbers are suggestive rather than settled:

  • In the 60 mg once-daily group, forced vital capacity (FVC) rose by a mean of 98.4 mL, with a 95% confidence interval of 10.9 to 185.9 mL.
  • In the placebo group, FVC fell by a mean of 20.3 mL, with an interval running from a 116.1 mL drop to a 75.6 mL gain.
  • The trial lasted only 12 weeks, enrolled patients at 22 sites in China, and was not powered to prove efficacy.

The intervals are wide. IPF normally erodes lung function over time, so a rise is notable, yet 71 patients is a small sample. The authors and the company both call for larger and longer studies.

The phase 3 that is now underway

On September 10, 2026, Insilico said it had dosed the first patient in GENESIS-IPF-3 (company announcement, registry ID NCT07687459). The design is a 52-week, randomized, double-blind, placebo-controlled study of about 320 participants at 47 centers, all in China. The primary endpoint is the annual rate of FVC decline. Time to first disease-progression event is the key secondary.

This is the study that will tell us whether the phase 2a signal was real. It is also a press-release-level fact at this stage: the enrolment details come from the sponsor and the registry, and no efficacy results exist. The company itself cautions that the drug is investigational and not approved anywhere. A Chinese-only population also means any positive result would still need work before it applies elsewhere.

Zasocitinib: computation without the generative label

Zasocitinib (TAK-279) is the other case worth knowing, because its phase 3 data are already public. It is an oral TYK2 inhibitor. Nimbus Therapeutics designed it in a partnership with Schrödinger that began in 2016, using free energy perturbation, a physics-based method that estimates how tightly a candidate molecule binds its target and how selective it is against related kinases. Schrödinger’s own case study describes the work. Takeda took over the program in 2023.

Takeda reported two phase 3 psoriasis trials in December 2025 and showed fuller data at the American Academy of Dermatology meeting in March 2026. According to Fierce Biotech’s report, the share of patients reaching PASI 90 (a 90% improvement in skin disease score) at week 16 was 61.3% and 51.9% in the two trials. Placebo reached 5% and 4%, and apremilast, an approved comparator, reached 16.8% and 15.9%.

Those are company-reported results, and the full peer-reviewed publication was not something I could confirm. Takeda has said it plans to seek approvals in its 2026 fiscal year. Whether regulators agree is not yet known.

Why include a drug that predates the current wave of generative models? Because it shows the pattern that probably matters most. Computation did not replace medicinal chemistry in this program. It narrowed the search, so that chemists made and tested fewer compounds before reaching a selective candidate. If AI-designed drugs eventually earn approval, the route may look a lot like that: faster, more targeted iteration, with the final judgment still coming from human trials. It is also a reminder that a compound can succeed in phase 3 for reasons unrelated to how it was invented, which is why the invention story should never be mistaken for evidence of benefit.

What the aggregate data say

Zooming out, the best-known analysis is a 2024 paper in Drug Discovery Today by researchers at Boston Consulting Group. As summarized by Health Tech World, AI-discovered molecules passed phase 1 at 80 to 90 percent, well above the historical 40 to 65 percent. In phase 2 the rate fell to roughly 40 percent, close to the industry average, on a limited sample.

Read that carefully. Phase 1 mostly tests whether a molecule is safe enough and behaves sensibly in the body, which is where better chemistry should show up first. Phase 2 asks whether the drug works against the disease, and that depends on whether the biological target was right. Better molecules do not fix a wrong target.

Stage What it mainly tests AI-discovered molecules (BCG analysis)
Phase 1 Safety, dosing, drug behavior About 80 to 90 percent success
Phase 2 Early evidence of efficacy About 40 percent, similar to average
Phase 3 Confirmed benefit at scale Not yet demonstrated

The setbacks matter too

Recursion offers a clean example. In May 2025 it ended or paused five programs, as Pharmaceutical Technology reported. One of them, REC-994 for cerebral cavernous malformation, had been safe and met its phase 2 primary endpoint, yet follow-up data did not confirm a meaningful clinical benefit. Passing a trial is not the same as helping patients.

The same secondary summary lists other discontinuations, including programs from BenevolentAI, Exscientia and Schrödinger. I have not independently confirmed each of those, so check the primary announcements before citing them.

Myth versus fact

Claim you may have seen What the record supports
AI has already produced approved drugs. Industry trackers say no AI-discovered drug has full marketing approval. Definitions vary, so the answer depends on what you count as AI.
AI cuts failure rates in half. Early phase 1 success looks higher. Phase 2 success is about average, and the sample is small.
AI discovers drugs in months instead of years. Companies report faster preclinical stages. Clinical trials still take years, and nothing in AI changes how long it takes to measure lung function over a year.
A positive phase 2a proves the drug works. It suggests a direction. Seventy-one patients over twelve weeks cannot settle effectiveness, as the confidence intervals above show.

The speed claim deserves a second look. Preclinical timelines, meaning the stretch from choosing a target to nominating a candidate, are the part software can plausibly shorten. The long stretch is human testing, which is gated by recruitment, safety monitoring and the need for enough events to measure. A drug for a slowly progressing disease like IPF cannot be rushed by a better algorithm, because lung function has to be followed for a full year before the trial can answer its question.

There is also a practical reason the field draws scrutiny. Drug development is expensive and mostly fails, and a technique that nudges success rates upward even a little is worth a great deal. That economic pull explains the investment and the optimism. It also explains why independent readers should keep asking for numbers with intervals attached.

How to read the next headline

  1. Check the phase. A phase 1 success says little about efficacy.
  2. Look for the registry ID and the endpoint. Safety-only primary endpoints cannot show a drug works.
  3. Separate peer-reviewed results from press releases and conference slides.
  4. Ask what the AI actually did: pick the target, generate the molecule, or rank existing compounds.
  5. Look at confidence intervals, not just averages.

The honest summary is modest and still interesting. AI methods seem to speed up early discovery and may improve the odds of clean phase 1 results. Whether they raise the odds of a drug that works in patients will be decided by trials like GENESIS-IPF-3, which will not report for some time.

Sources