AIxBio RegistryAI molecules in the clinic

Methodology

Version 1.3.7 · data as of 17 November 2026

1. Definition

A molecule is in the strict cohort if all three hold:

  1. AI role in discovery, not only in screening logistics, manufacturing, patient selection or trial analytics. Physics-based design with no ML claim and computational repurposing of an existing drug are published as borderline. What counts is what AI designed in the drug itself, the active substance as regulators define it: the therapeutic protein or molecule, a conjugate's payload and linker, a viral vector's capsid, or the full sequence of an mRNA or oligonucleotide. It counts however the drug is given, as a protein, carried in a cell (a CAR-T) or encoded by a vector. AI that worked only on a delivery system or formulation (a lipid nanoparticle, a new formulation of an existing drug) is published as borderline; AI used only in manufacturing is not counted. A vaccine counts when a trained model selected the antigens it carries, whether they are the same for every patient or chosen for each one (from v1.1.0; before that these were borderline). A cell therapy raised against antigens a trained model selected is judged the same way. A phage cocktail is judged like any biologic: it counts when a trained model chose, designed or engineered its phages.
  2. Attribution (the acid test). The developer or a partner credits AI or ML for this molecule in its own words. The statement must name the molecule, or be unambiguously about it, and the thing credited must be AI itself: AI or ML named there, a named AI model, or a platform the developer or partner calls AI. The description of the platform as AI may come from a separate developer or partner statement only when the evidence shows the AI did the work on this molecule; a molecule tied to a platform that only company boilerplate calls AI is platform_only. Third-party statements are not enough. Whether the claim also says what the AI did (chose the target or designed the molecule) is recorded as ai_step_stated; it does not decide inclusion.
  3. Dosed in humans by the as-of date, by any sponsor, in a trial registry (ClinicalTrials.gov, the EU's CTIS and Clinical Trials Register, the UK's ISRCTN, Australia's ANZCTR, Japan's jRCT, China's CTR and ChiCTR, or another WHO primary registry) or per a dated developer announcement.

One entry per molecule. The registry counts drug substances, the unit the published surveys count (§ 6), not trials or indications. Each entry is one active substance, under its code, INN and former names. Its trials, indications, approvals, salts, formulations and routes of administration are recorded on that entry, whoever sponsored them and in whichever registry. A back-up or follow-on molecule gets its own entry once it is dosed. A fixed combination developed as one product is one entry; separate products given together each have their own entry, and a trial that gives both is listed on both. A personalized vaccine or cell therapy is one entry per product, not one per patient. Phase transitions are counted per molecule, by its furthest phase (§ 4), as most of the industry studies on the question page count them.

The tests are applied by the written evidence protocol in review protocol: two blind coders per row, a red team arguing for exclusion on every strict pass, and the maintainer's sign-off on every strict entry. The acid test: if the developer does not claim the molecule as AI-derived in its own words, neither does the registry. Protocol v1.1 (20 October 2026) made the acid test the strict line, v1.2 (25 October) limited physics_only to molecules whose developer names no ML for them, and v1.3 (26 October) judges the method rather than the vocabulary: a data-trained model must have chosen the target or produced or selected the molecule, shown by any developer or partner evidence, with ai_methods recorded for every reviewed row. v1.4 (27 October) applies the test to the therapeutic protein however it is delivered, counts ML-guided optimisation when a trained model identified the modifications that define the clinical molecule, and treats lab-in-the-loop platforms, whose experiments feed or test the model, as AI platforms rather than multi-method ones. Under v1.0, a claim also had to say what the AI did, which cut the count to 32; that requirement is now a quality flag (§ 6).

2. Inclusion tiers

Tier Meaning Counted?
strict Meets all three tests Yes
borderline Dosed, with an AI connection, but not strict. One code in borderline_reason: physics_only (computational design; the developer's statements about this molecule name no ML, even if its platform description does), platform_only (the molecule is tied to a platform that the statement does not call AI, or AI appears only in company boilerplate), third_party_only, delivery (AI worked only on a delivery system or formulation, not on the drug itself), repurposing (a trained model picked a new use, or a new combination, for a drug already dosed in people; a use already tested in people does not count) or other No, but published with the reason in inclusion_note
candidate Plausible but not yet verified No
excluded Considered and rejected; one code in exclusion_code: not_dosed, ai_disclaimed, in_licensed, legacy_retrofit, out_of_scope or no_ai_link No; listed in excluded.csv

AI-derived, broad count and "AI-enabled". On this site "AI-derived" means the strict tier. Strict and borderline together are the broad count: every dosed molecule with an AI connection. The broad count, not the strict count, is the counterpart of published lists of "AI-enabled" assets such as Haque et al.'s, which do not separate AI design of the molecule from repurposing, platform-level claims or physics-based design (§ 6). Borderline is not a lesser kind of AI-derived. In most borderline rows the evidence does not show AI working on the molecule (platform-only, physics-based, contradicted claims); in the rest, AI did a step other than designing it (a new use for an existing drug).

3. Coding rules

4. How the list was built

Six parallel searches: published seed lists (BCG/Jayatunga 2024 and successors); AI-native small-molecule developers in the US, Europe and Israel; Asia-Pacific developers and registries; biologics; pharma-internal assets; and an asset-agnostic registry and literature sweep. Results were merged and coded, then challenged by two independent passes: one trying to disprove entries (about 50 registry records re-checked) and one looking for missing assets. Searches were AI-assisted; every row carries its sources. In October 2026 every stopped, unknown and Phase 2+ status was re-checked against its registry and every source link was verified (20 October 2026).

Evidence-first re-review (October 2026). Every row, strict, borderline and excluded, was re-coded under the review protocol by two coders who could not see the current tiers or each other's work. The protocol was piloted first on 24 rows with known answers (coder agreement κ = 0.93). On all 158 rows the coders agreed on the tier for 145 (91.8%, κ = 0.87). A red team then argued for exclusion on all 38 rows that either coder passed as strict, and defeated 6. After the strict line moved to the acid test (protocol v1.1), the 47 platform-level cases were re-coded by two fresh blind coders. They agreed on all 47 (κ = 1.0), and a red team defeated 7 of the 28 new passes. When v1.2 limited physics_only (25 October), all 17 physics_only rows were re-tested the same way. Coders agreed on 15 of the 17; BBO-8520 and SNX-281 joined SGR-1505 and SGR-3515 in the strict cohort, and red-team evidence kept zasocitinib, iscartrelvir, omilancor, GT-02287 and ATH-063 borderline. Protocol v1.3 (26 October) re-ran the 55 strict rows, the wording-sensitive borderlines and the no_ai_link exclusions in both directions: two fresh blind coders on 112 cases (tier agreement 91.1%, κ = 0.85), then a red team on all 64 passes and demotions. The excluded groups were then re-reviewed molecule by molecule under the same procedure: 19 rows were added (one strict, HXN-1001), four borderline rows were excluded, and follow-up red teams moved COM701 to borderline. Under protocol v1.4 (27 October) GB-5267 and ABBV-295 joined the strict cohort by the maintainer's decision. The five platform rows the new wording could move were re-coded by two blind coders (agreement on 1 of 5) and a red team: JUV-161 became strict, and the other four stayed platform_only. Three molecules on BiopharmaTrend's lists that the registry had not assessed went through the same procedure on 27 October: REC-3599 (ruboxistaurin) and MORF-057 joined as borderline (repurposing and platform_only), and A2A-252 proved to be AO-252, already borderline. Every quoted claim was re-fetched and found on its page. Protocol v1.5 (9 November) applied the same method test to the 19 repurposing rows, with two blind coders on each older row and three red teams: four entries were excluded and LAM-001 was split from LAM-002A and excluded, seven moved to another borderline code, and eight stayed repurposing. The registry was also cross-checked against four published lists (data/crosswalk.csv).

Who did what. AI agents did the searching, the blind coding, the red-team challenges and the re-fetching of quotes. A person, the maintainer John E. Milad, reviewed their output and made every judgement call the written rules did not settle: each rule change (protocol v1.1 to v1.4), each disputed entry, and each scope question. Every decision, with the losing argument recorded as a dissent, is kept in the maintainer's sign-off record; the resulting changes and their evidence are in the changelog and in each entry's record. No entry joined the strict cohort without his sign-off; he reviewed and signed off all 55 on 27 October 2026. No independent expert has yet reviewed the registry. Reviewers will be named here when they have.

Registries searched. ClinicalTrials.gov; the EU Clinical Trials Information System (CTIS, trials from 2023) and the EU Clinical Trials Register (EudraCT, earlier trials); the UK's ISRCTN; Australia's ANZCTR; Japan's jRCT; and China's CTR and ChiCTR. WHO's International Clinical Trials Registry Platform (ICTRP) is a cross-check that also reaches other national registries, such as the Pan African Clinical Trials Registry. Dated developer announcements fill gaps where no registry record exists. An entry's registry IDs are the interventional trial records that name the molecule as an intervention, whoever sponsors them; observational studies, expanded-access programs and records withdrawn before enrollment are not listed. ClinicalTrials.gov was searched for every molecule in the build and audit (13–20 October 2026), and ANZCTR, jRCT and the Chinese registries in the Asia-Pacific search. On 25 October 2026 every strict and borderline molecule was searched by each code, alias and INN, and by developer name, in EU CTIS, the EU Clinical Trials Register, ISRCTN and ICTRP. This added 63 registry records to 32 entries (7 strict) and changed no count, phase or first-in-human date. Each entry's records are in registry_ids.

Trials (trials.csv) are every ClinicalTrials.gov record matched to a strict-cohort molecule by NCT number or verified intervention name, plus the ClinicalTrials.gov records in each borderline entry's registry IDs (by NCT number: ClinicalTrials.gov is not searched by name for borderline entries, but the registry searches in § 8 add any record that names the molecule; API v2, retrieved 17 November 2026). Only the strict cohort's trials enter the trial outcomes and comparisons below. Twelve strict molecules have no ClinicalTrials.gov record. Posted protocols and statistical analysis plans are linked where the sponsor has uploaded them. Each trial also carries its enrolment as ClinicalTrials.gov records it (enrollment, with enrollment_type: ACTUAL for the number enrolled, ESTIMATED for the number planned) and the conditions it is registered for (conditions, separated by semicolons, as the sponsor wrote them). A field the record leaves out is blank. trials.csv holds ClinicalTrials.gov records only. Records in other registries are listed in each entry's registry_ids and linked from the site, but they are not in trials.csv, because other registries do not report status, dates and sponsor type in a comparable form; so no record in another registry has an enrolment or conditions in the registry's data.

Trial outcomes and the industry comparison use ClinicalTrials.gov records only. The strict cohort has 143 ClinicalTrials.gov trials. A trial is closed if it is completed or terminated; withdrawn trials (stopped before enrolling anyone) are left out. Of the 45 closed trials, 14 were terminated. For the industry comparison the registry's trials are matched to the comparator's filter (industry-sponsored, Phase 1, Phase 1/2 or Phase 2, started 2018–2024): 13 of 35 closed trials were terminated. The comparator is every ClinicalTrials.gov trial matching the same filter, retrieved through API v2 on 13 November 2026 with filter.advanced = AREA[StudyType]INTERVENTIONAL AND AREA[LeadSponsorClass]INDUSTRY AND AREA[InterventionType]DRUG AND AREA[StartDate]RANGE[2018-01-01,2024-12-31] AND (AREA[Phase]PHASE1 OR AREA[Phase]PHASE2): 18,200 trials, of which 10,474 completed, 2,387 terminated and 594 withdrawn. That gives 18.6% of closed trials terminated, or 22.2% if withdrawn trials are counted as stopped. Some comparator trials will also involve molecules with AI input their developers did not disclose, so the comparison is between disclosed AI design and an industry where AI is common, not between AI and no AI.

Phase transitions (the question page) follow the industry studies' convention for combined phases (Wong et al. 2019: "Phase 1/2 and Phase 2/3 trials are to be considered as Phase 2 and Phase 3, respectively"). Each strict molecule records its entry_phase, the phase of its first trial in people: Phase 1, or a combined Phase 1/2 trial. A molecule that entered in a combined trial enters at Phase 2, so it is not part of the Phase 1 → 2 step. A molecule that entered at Phase 1 moved on once it reached Phase 2 or later, including through a later Phase 1/2 trial. A Phase 2/3 trial counts as Phase 3. Entry phase comes from the first trial in registry_ids; where the first trial was never registered (US Phase 1 studies need not be), it comes from the developer's announcement, cited in the audit log. It is that trial's phase when the molecule was first dosed: a Phase 1 trial later amended into a combined Phase 1/2 design (IAM1363's, in 2026) counts as the molecule reaching Phase 2, not as entering there. Review: docs/reviews/2026-11-01-phase-transition-method.md. The home page chart still shows Phase 1/2 as its own column.

Approvals (approvals in molecules.csv) record each regulatory approval as agency, date, application or approval reference, indication and a link to the regulator's own record (Drugs@FDA for US approvals). A molecule can be approved and still ongoing when trials in other indications continue.

Randomized readouts (readouts.csv) record each randomized, placebo-controlled trial with published efficacy data for a strict molecule: trial, design, efficacy measure, result, date and source. A Phase 2+ trial on a novel target that has passed its primary completion date with nothing published is also logged, as randomized_efficacy = no with the date checked, and is re-checked after 180 days. Coverage is complete for strict molecules on novel targets (checked 21 October 2026); molecules on known targets are not yet recorded. A readout is a published result, not a trial in progress, and efficacy may be a secondary endpoint.

5. Uncertainty

The strict count is about 75, −2/+8: - dosing and attribution errors: about ±2. Six strict entries (DSP-0038, DSP-1181, ENV-308, ENV-6946, EXS-4318, HXN-1001) rest on developer statements of dosing with no registry record read, and the maintainer's sign-off may still move contested entries; - assets probably missing, mainly from China CTR and recent US first-in-human starts: +0 to +6; - definition sensitivity, each computed from borderline_reason in molecules.csv: - leaving out the vaccines whose antigens a model selected gives 58: the seventeen entries with ai_step = antigen (§ 1.1), nine counted since v1.1.0, six added in v1.2.0 and two in v1.3.0. Two of the seventeen, cNeT and iNeo-Vac-T01, are T-cell therapies raised against such antigens. Five are at low confidence: GRT-R910, autogene cevumeran, iNeo-Vac-P01, cNeT and NEOVACC. These rows are not a census of such products: Chinese-language registries and company sites, Japanese and Korean developers and academic trials have not been searched; - excluding Schrödinger's entries gives 73: SGR-1505 and SGR-3515 are strict because Schrödinger names machine learning for each, but its design is physics-led (FEP+); - counting physics-based design gives 91: adding the 16 borderline physics_only entries; - counting platform-level attribution gives 120: adding the 45 borderline platform_only entries.

Why the two Schrödinger molecules count and Relay's do not. Both companies design mainly with physics: Schrödinger with free-energy calculations (FEP+), Relay with long-timescale molecular dynamics (its Dynamo platform). Both say their platforms also use machine learning. The test in § 1.2 is not the company or the method mix; it is whether the developer, in its own words, credits machine learning for this molecule. - Schrödinger does, for both of its strict molecules. Its paper on SGR-1505 (J Med Chem 2025, doi:10.1021/acs.jmedchem.5c01494) describes "combining free energy calculations with machine learning methods" to find the hit series. Its case study for SGR-3515 describes Active Learning FEP+, which it calls machine learning, "en route to development candidate SGR-3515". - Relay does not, for any of its molecules. Its own discovery papers for lirafugratinib (PNAS 2024, doi:10.1073/pnas.2317756121) and zovegalisib (Cancer Discovery 2024, doi:10.1158/2159-8290.CD-23-0704) credit molecular dynamics and structural biology; neither full text mentions machine learning. Machine learning appears only in Relay's description of the platform, so its molecules are borderline. - Schrödinger's SGR-2921 is borderline for the same reason as Relay's: Schrödinger's statements about it name physics-based modeling and the platform, not machine learning. MORF-057, which Morphic designed with Schrödinger, is borderline platform_only: Schrödinger's case study lists machine learning among the programme's tools, but its account of the design credits free-energy and quantum-mechanics calculations. AJ1-11095 (Ajax, co-discovered with Schrödinger) and LTSE-2578 (Structure Therapeutics, designed with Schrödinger's FEP) are physics_only: their developers' accounts credit structure-based design and free-energy calculations, and machine learning appears only in Schrödinger's description of its platform.

In both strict Schrödinger molecules the machine learning works alongside physics, which is why they are the sensitivity case above: excluding them gives 73.

The ±2 and +0 to +6 components come from the verification and completeness passes of the 13 October 2026 build; their working log is not yet published.

6. Other counts, and why they differ

Published counts differ mainly in which kinds of AI claim they count, not in what they find. The ladder below adds one kind at a time. Counts are cumulative and computed from molecules.csv. The list columns show where each published list's entries fall in this registry, from data/crosswalk.csv. A script checks every number here against the data before each release.

Step This registry Jayatunga 2024 (75) Greenblatt & Geoui 2026 (83) Wellcome 2023 (72) Haque et al. 2026 (129) McKinsey 2026 (9 examples)
Strict, AI step stated 49 17 19 12 30 8
Strict 75 27 30 18 44 9
+ platform-only 120 42 46 27 67 9
+ physics-based 136 50 55 35 72 9
+ repurposing, delivery, other 170 62 67 50 94 9
Excluded here, or not in this registry – 13 16 22 35 0
Count Number Definition
Jayatunga et al. 2024 (BCG), Drug Discovery Today 75 Clinical assets of AI-native companies since 2015
Greenblatt & Geoui 2026 (Zenodo) 83 BCG list plus additions
McKinsey 2026, Exhibit 2 9 Examples of "AI-enabled assets" with disclosed discovery cycle times (to a preclinical candidate, IND or first dosing), each in Phase 1 to 3 at publication; non-exhaustive
Haque et al. 2026, Cancer Treatment Reviews 129 "AI-enabled" assets in trials by March 2026: any meaningful AI contribution to discovery, design or optimisation, on company or third-party evidence
This registry 75 Developer credits AI for the molecule, dosed in humans, written rules
This registry, broad count 170 Strict and borderline: every dosed molecule with an AI connection; the counterpart of "AI-enabled" lists

Lower is not more correct. Each count answers a different question.

These counts, and the other literature, benchmarks and guidance this methodology relies on, are listed with links on the bibliography page.

7. Limits

8. Updates

Since 30 October 2026 an automated watch has run every Monday. It changes no data; what it finds goes to the maintainer's review queue.

How a finding is decided. Each finding gets a proposal from an AI model (Claude, from Anthropic) that searches the web; since 14 November 2026 the quote a proposal relies on is checked against the page it cites, and a proposal whose quote is not there is marked low confidence. The maintainer accepts or rejects each proposal. Two routine kinds are settled without him, stay visible and can be reopened:

A daily update applies approved changes of a few routine kinds as a logged release, after checking each source again: trial IDs, indications, and a clearer quote from the same source where the AI step is already recorded as stated. Anything that could move a count, a tier or a code, and every new candidate, goes through the standing review in the review protocol and the maintainer's sign-off. Readers' corrections take the same route.

By hand, each quarter: China's CTR (which is not in ICTRP) and jRCT block automated access, so every China- and Japan-linked code is searched on their websites.

Every row carries a last_verified date: the day each of its fields was last checked against primary sources. The "data as of" date on the site and in each row's as_of is the day the register's data last changed: an entry, an exclusion, a trial record, a readout or a developer detail. Updating the comparisons with published lists does not move it. The evidence is in data/audit-log.csv. It has one line per row and field, with the verdict (confirmed, changed or flagged), the old and new value, the source URL and a short verbatim quote. The log is append-only, so a later check adds lines and keeps the earlier ones. From v1.0.0, the first citable version, every release's data can be downloaded as released (data/releases.csv lists each release and its download; data for the earlier, pre-release drafts is available on request), and data/register-changes.csv lists each entry added or removed, each move between strict and borderline, and each recoding that moves a published figure. The first full field-by-field audit covered every row on 20 October 2026.