What we've found so far
75 AI-derived molecules dosed in people.* None approved.
The AIxBio Registry is an open record of drug candidates whose target or molecule was materially shaped by AI or machine learning. Each one is counted to a written definition, sourced row by row, and open to correction.
Most have gone after biology the industry already knew how to hit: 38 of the 54 with a single disclosed target. The other 16 (30%) are on novel targets: 9 where earlier drugs had been tried in the clinic, 7 new to the clinic. Of those 16, one (rentosertib) has a randomized, placebo-controlled readout with efficacy data.
For all but one (rentosertib), the only evidence that AI did the work is the developer's own statement. 10 have a Phase 2 outcome so far, enough to start comparing their success rate with the industry's, though the sample is still small: 5 moved on, with a 95% confidence interval of 24–76%.
Data as of 17 November 2026. Trial records retrieved 17 November 2026. Version 1.3.7.
17 of the 75 are vaccines, or in two cases T-cell therapies, whose antigens a model selected. Such products have been counted since version 1.1.0; before that they were listed as borderline. What changed.
Every change to the register, release by release.
* The true number is almost certainly higher. Our searches will have missed some, and we leave out molecules whose AI claim fails our written tests (they are published as borderline). Each one counted has the developer or a partner crediting AI for the molecule in its own words. AI input is now routine across drug discovery and often undisclosed, so this is a cohort, not a census. Why it matters
Each dot is one molecule, placed by the furthest phase it has reached
Colour shows what was known about its target before it was first dosed. Select a dot to open its entry.
- Known target
- Novel: tried in the clinic before
- Novel: new to the clinic
- Not classified
- Stopped
- What the target types mean
What counts
A molecule is in the strict cohort only if all three hold. Entries that miss one test in a way reasonable people dispute are published as borderline, with the reason, but not counted. The full definition is in Methodology § 1.
AI in discovery
Machine learning or generative methods helped find the target or the molecule, not just run screening logistics, manufacturing or trials. Physics-only design, repurposing and AI-designed delivery or formulation are published as borderline. A vaccine counts when a model selected the antigens it carries.
The developer says so
The developer or a partner credits AI for this molecule, in its own words. Each entry shows the quote, and whether it also says what the AI did.
Dosed in humans
In any trial registry worldwide, or per a dated developer announcement, by the as-of date.
How targets are classified
Each molecule's target is judged as it stood on the day the molecule was first dosed in a person. The question is how much the field already knew about hitting that target in people. The rule is set out in Methodology § 3; each entry shows the basis for its call.
Known target
An approved drug, or another drug that had reached Phase 2 or later, already acted on this target. The biology was validated in patients before the AI molecule arrived.
Novel: tried in the clinic before
Other drugs had been dosed in people at this target, but none had got past Phase 1. Known to be druggable, not yet shown to matter in patients.
Novel: new to the clinic
No drug had ever been dosed in people at this target. This is the case that would show AI finding new biology, not just new chemistry. (Coded
first-in-classin the data; in industry usage that term describes a drug with a new mechanism, so the site avoids it.)Not classified
The developer has not said what the target is, or the product has no single target: a vaccine whose antigens a model selected carries many, often different for each patient.
Charts show the cohort by year first dosed, what the AI did, therapeutic area and modality.
Why other counts are bigger
Published counts of AI-discovered drugs in the clinic range from about 50 to more than 100. They differ mainly in what they ask. Lists of “AI-enabled” assets, such as Haque et al.’s, compare with this registry’s strict and borderline entries together (the last step in the table below), not with its strict count.
Where the counts part company
The same evidence gives different totals depending on which kinds of AI claim are counted. Each step adds one kind; the counts are cumulative and computed from this registry's data. The columns show where each published list's molecules fall here, from the crosswalk.
A lower number is not more correct. Each count answers a different question. The uncertainty band on 75 is about −2 to +8, depending on registry coverage and on contested calls. How the band is built.
- Jayatunga MKP et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today 29(6):104009, 2024. doi:10.1016/j.drudis.2024.104009. 75 molecules entered the clinic since 2015.
- Greenblatt E, Geoui T. Dataset and analysis code for "28 Months Later: AI-Discovered, but Not AI-Developed". Zenodo, 2026. doi:10.5281/zenodo.22814924. 83 assets, each with per-field citations.
- Haque W et al. Artificial intelligence-enabled drug discovery: a growing but largely early-stage clinical pipeline. Cancer Treatment Reviews 151:103223, 2026. doi:10.1016/j.ctrv.2026.103223. 129 assets across 69 companies, in trials by 1 March 2026 (their 2025 ASCO abstract counted 117); their supplement lists the assets, and this registry’s crosswalk download gives its reading of each.
The registry
Every entry, with the AI's role, the evidence for it, its registered trials and its sources, is in the registry table.
Read this before citing
- A cohort, not a census. AI input is now routine across drug discovery, often as design suggestions that chemists rework, and much of it is never disclosed. No registry built from public sources can find every molecule AI touched, and our searches will have missed some. This one counts those whose developers say, in their own words, that a trained model chose the target, produced the molecule or selected a vaccine's antigens. Developers choose what to disclose, and those selling AI have reason to credit it when things go well, so if anything this cohort should flatter AI. Each claim is recorded when it is made, so a molecule stays in the cohort if it later fails.
- Re-reviewed. In October 2026 a check found entries counted on AI claims made for a company's platform rather than for the molecule itself, and a headline that briefly counted one such molecule as an approved AI drug. Those entries have been moved to borderline. Every entry has now been re-reviewed against a written evidence test: a strict entry needs developer or partner evidence that a trained model chose its target or produced the molecule. The test is set out in the review protocol.
- Attribution is mostly self-reported. 74 of the 75 strict entries rest on the developer's own statement that AI was used. Only rentosertib has the AI role documented in peer-reviewed or regulatory sources (attribution tier A).
- Second coding covers inclusion only. Every entry's tier was coded by two blind coders and checked by a red team (protocol). Target-type calls follow a written rule but were made by one coder.
- The list was assembled with AI-assisted searches and coded by AI agents working blind to each other, with an AI red team arguing against every strict entry. One person, the maintainer, made every judgement call and signed off each strict entry; no independent expert has reviewed the registry yet (Methodology § 4). Every row carries its sources, and every change since is logged in the changelog.
- Inclusion is not an efficacy rating. Being listed says nothing about whether a molecule works.
Milad, J. E. AIxBio Registry: AI Molecules in the Clinic (v1.3.7, data as of 17 November 2026). https://aixbio-registry.org
Corrections
Spotted an error, a missing molecule or an entry in the wrong category? Send a correction with a source. Each one is checked against its sources and the methodology, and every decision is published in the corrections log.