Health & biomed · Tracker
Drugs discovered or designed with AI, tracked in the clinic.
Which drugs made with AI are actually being tested in people, how far along they are, and what is really “AI” about each one. Every row links to its sources. Last checked Nov 18, 2026.
Reality check
No AI-discovered drug is FDA-approved yet
As of October 2026, no drug discovered or designed by AI had full FDA approval. The closest is zasocitinib, with an FDA decision due in early 2027. Some earlier AI candidates have already failed: one was dropped in Phase 1, another failed Phase 2a.
Source: IntuitionLabs, AI drug discovery and FDA approvals (updated Oct. 15, 2026)
What the phases mean
- Phase 120 to 100 people. Is it safe, and what dose? Takes several months. About 70% of drugs move on.
- Phase 2Up to several hundred patients. Does it seem to work, and what are the side effects? Several months to 2 years. About 33% move on.
- Phase 3300 to 3,000 patients. Does it work better than the alternative, and is it safe at scale? 1 to 4 years. About 25% to 30% move on.
- FDA reviewThe company files an application. The FDA decides whether to approve it. Approval is the finish line, not Phase 3.
Source: FDA, Step 3: Clinical Research
Why “AI-discovered” is a contested label
AI can help at different steps: finding a new target in the body, designing the molecule, or predicting how it will behave. A drug might use AI for one step or all of them, and there is no agreed bar for the label.
Even the leading candidate is disputed. Nimbus, which discovered zasocitinib, used physics-based simulation plus machine learning. Its R&D president says “first AI-approved drug” is “not the term I would use” for a compound identified in 2020.
Source: PharmaVoice, Nimbus used AI to help develop Takeda's $4B psoriasis bet (Sept. 25, 2026)
The pipeline
| Company | Candidate | Condition | Current phase | Next milestone | What's “AI” about it | Sources | Last verified |
|---|---|---|---|---|---|---|---|
| Insilico Medicine | Rentosertib (ISM001-055), a pill that blocks the protein TNIK | Idiopathic pulmonary fibrosis (scarring of the lungs) | Phase 3 First patient dosed in September 2026. About 320 patients at 47 centers in China, one year of treatment, compared with placebo. In the earlier Phase 2a trial (71 patients, 12 weeks), lung capacity rose 98.4 mL on the top dose and fell 20.3 mL on placebo. | Main results: the registry estimates October 2029. A lead investigator expects 3 to 4 years from the start of Phase 3 to approval, if all goes well. | The Phase 2a paper says both the target (TNIK) and the molecule came from generative AI. A lead investigator says TNIK had never been linked to fibrosis before. | Nov 18, 2026 | |
| Takeda (discovered by Nimbus Therapeutics) | Zasocitinib (TAK-279), a pill that blocks the protein TYK2 | Moderate-to-severe plaque psoriasis | FDA review Phase 3 done. In two pivotal trials, 69% to 71% of patients had clear or almost clear skin at week 16, against 11% to 13% on placebo. The FDA accepted the application for Priority Review. | FDA decision expected in the first quarter of 2027 (no exact date given). Europe's regulator is also reviewing it. | Nimbus used physics-based simulation combined with machine learning to assess more than 13,000 compounds. Nimbus itself pushes back on the label: its R&D president, Peter Tummino, told PharmaVoice that "first AI-approved drug" is "not the term I would use" for a compound identified in 2020. IntuitionLabs notes the "AI-designed" label is disputed: the drug came from physics-based computer chemistry aimed at an already-known target. | Nov 18, 2026 | |
| Generate Biomedicines | GB-0895, an antibody against TSLP, given as a shot every six months | Severe asthma | Phase 3 Two Phase 3 trials (SOLAIRIA-1 and SOLAIRIA-2), each planned for 786 people. Xtalks reports the first patient was dosed in January 2026. | Full enrollment expected in the first half of 2028. The registry estimates main results in late 2028 to early 2029. | Generate says GB-0895 was "optimized using AI" for very tight binding to TSLP, a longer half-life and high specificity. | Nov 18, 2026 | |
| Iambic Therapeutics | IAM1363, a pill that blocks HER2 and is built to reach the brain | HER2-driven cancers, including tumors that spread to the brain | Phase 1/2 Early results at ESMO 2025: partial responses (real tumor shrinkage) in 28% of 18 patients with measurable disease. The trial plans to enroll 463 people. | The registry estimates main results in October 2028. | Iambic says it designed the drug with its AI models (Enchant and NeuralPLexer) and went from discovery to first human dose in under two years. | Nov 18, 2026 | |
| Absci | ABS-201, an antibody | Androgenetic alopecia (male- and female-pattern hair loss) | Phase 1/2 First-in-human study that started in December 2025, with up to 227 people. The registry says its effect on hair growth has not yet been established. | The registry estimates main results in July 2028. | Absci describes its approach as "unlocking differentiated biology with generative AI." It has not published how AI shaped ABS-201 in the sources we checked. | Nov 18, 2026 | |
| Recursion | REC-4881, a pill that blocks MEK1/2 | Familial adenomatous polyposis (an inherited condition that causes colon polyps) | Phase 1b/2 In updated TUPELO data (Nov. 2026), 13 of 15 patients (87%) had less polyp burden after 12 weeks, with a median 47% drop. These are company-reported results from a small trial with no comparison group. A separate REC-4881 cancer trial was terminated. | Recursion expects to start a registrational (pivotal) trial in the second half of 2027, depending on talks with the FDA. | AI did not design the molecule: Takeda made it (as TAK-733), and Recursion licensed it. Recursion's AI-driven cell screening picked the new use, FAP. A known drug given a new use, not a new AI-designed molecule. | Nov 18, 2026 | |
| Isomorphic Labs | None disclosed in human trials | Not disclosed | Before human trials We found no public announcement that any Isomorphic drug has been given to a person. | CEO Demis Hassabis said in January 2026 that the first trials should start by the end of 2026. That was already a delay from an earlier target of the end of 2025. | Google-backed company that designs drugs with AI. A promised start date is not a drug in patients. This timeline has slipped once already. | Nov 18, 2026 |
On a phone, swipe the table sideways to see every column.
- Dates credited to “the registry” are the sponsor's own estimates on ClinicalTrials.gov. They often slip.
- We left out companies where we could not verify a candidate in human trials, including Xaira Therapeutics and Chai Discovery. We will add them when that changes.
- Being on this list is not an endorsement. None of these drugs is approved anywhere.
Evidence check: how to read a medical AI claim
Headlines about AI that “beats doctors” or “cures” a disease are common. Four questions sort the strong claims from the weak ones.
Real patients or simulated cases?
Many diagnosis results come from written case studies, not real patients. Microsoft's MAI-DxO diagnosed 85.5% of 304 published medical journal cases, against 20% for 21 doctors. But the doctors could not use colleagues, books or AI, and Microsoft says the tool is not yet approved for clinical use.
Source: Microsoft AI, The Path to Medical Superintelligence (June 30, 2025)
Peer reviewed, or a press release?
A company announcement or a preprint has not been checked by outside experts. MAI-DxO was a preprint when Microsoft announced it. The rentosertib Phase 2a results were published in Nature Medicine after peer review. Peer review checks the methods; it does not guarantee the result will hold up.
Who paid for it?
Most trials in this tracker are run by the company that owns the drug. That is normal, but it means the company chose the design and reports the results. ClinicalTrials.gov lists each trial's sponsor; check it.
How big, how long, and compared with what?
A 12-week trial in 71 people, like the rentosertib Phase 2a, is a promising start, not proof. Look for a placebo or standard-care comparison group, random assignment, and a Phase 3 with hundreds or thousands of patients.
Checked by a human editor on Nov 18, 2026. Spot an error? See our corrections policy. Want the bigger picture on AI myths? Read our Claims check.