Changelog
What changed in the registry's content in each release: entries, counts, coding, sources and method, from v1.0.0, the first citable version (2 November 2026). Releases that changed only the site's presentation or its tools are grouped as site and maintenance updates.
Changes to the register
The releases that added or removed an entry, moved one between strict and borderline, or recoded one in a way that moves a published figure (borderline reason, target novelty, furthest phase, status or an approval). Every other change is in the release notes below.
| Release | Strict | Borderline | What changed on the register |
|---|---|---|---|
| v1.3.6 17 November 2026 | 75 | 95 | IAM1363: furthest phase now Phase 1/2 (was Phase 1/1b) |
| v1.3.0 15 November 2026 | 75 (+5) | 95 (+14) | 19 added
|
| v1.2.0 13 November 2026 | 70 (+6) | 81 (+2) | |
| v1.1.0 13 November 2026 | 64 (+9) | 79 (-9) | 9 moved tier, 1 borderline reason recoded
|
| v1.0.16 9 November 2026 | 55 | 88 (-4) | 4 removed, 7 borderline reason recoded, 1 status changed
|
| v1.0.14 8 November 2026 | 55 | 92 (+1) | BXCL701: added as borderline |
| v1.0.13 8 November 2026 | 55 | 91 (+9) | 9 added, 1 borderline reason recoded, 1 target recoded
|
| v1.0.3 5 November 2026 | 55 | 82 (+2) | |
| v1.0.1 3 November 2026 | 55 | 80 | GB-7624: target now known (was undisclosed) |
| v1.0.0 2 November 2026 | 55 | 80 | First citable release: 55 strict, 80 borderline (the baseline for this changelog) |
Downloads: every change to the register (CSV) · every release with its counts and the download that holds its data (CSV).
v1.3.7 — 2026-11-17
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- EXS-73565: one EU trial number removed. It belongs to GTAEXS-617's trial and stays on that entry; EXS-73565's own ClinicalTrials.gov record lists it by mistake, which the entry's note now explains.
- Methodology: the unit of count is now stated. Each entry is one drug substance, and its trials, indications and approvals are recorded on that entry, whoever sponsored them, as the published surveys count. No count changed.
v1.3.6 — 2026-11-17
Register: IAM1363: furthest phase now Phase 1/2 (was Phase 1/1b) · 75 strict, 95 borderline · data as released (zip)
- IAM1363's furthest phase is now Phase 1/2: its trial, first registered as Phase 1, is now registered as a combined Phase 1/2 study with a dose-expansion part. The question page counts it as having moved on from Phase 1 to Phase 2: 18 of 31 strict molecules with a Phase 1 outcome moved on (58%), up from 17 of 30.
- Methodology: a molecule's entry phase is its first trial's phase when it was first dosed, so a Phase 1 trial later amended into a combined Phase 1/2 design counts as reaching Phase 2.
v1.3.5 — 2026-11-17
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- Trial registry IDs added: 97 for 38 entries (8 strict, 30 borderline), found by the weekly searches of WHO's trial registry platform and the UK's ISRCTN. Each record was checked: an interventional trial that names the molecule, not withdrawn before enrolment. 59 other matches were set aside as other products, studies that only shared a word or code, duplicates, observational studies or withdrawn trials.
- Trial records refreshed from ClinicalTrials.gov on 17 November 2026: 402 trials, up from 348. All 54 new records belong to borderline entries; the strict cohort's 143 trials, and every figure computed from them, are unchanged.
- REC-3964's first dosing is now dated from its trial record (Netherlands, first enrolment August 2022) instead of the developer's announcement. The year, 2022, is unchanged.
- Methodology: a borderline entry's ClinicalTrials.gov records are those in its registry IDs. ClinicalTrials.gov is not searched by name for borderline entries, but the registry searches add any record that names the molecule.
v1.3.4 — 2026-11-17
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- Trial registry IDs added: 10 for 5 entries (ABBV-295, autogene cevumeran, bamlanivimab, intismeran autogene, nirmatrelvir), from the entries' own ClinicalTrials.gov records.
v1.3.3 — 2026-11-16
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- The trials download gains three fields from each trial's ClinicalTrials.gov record:
enrollment, the number of participants, withenrollment_type, which says whether that number is actual (enrolled) or estimated (planned), andconditions, the conditions the trial is registered for. No page shows them yet. Trials in other registries are not in the trials download, so they have none of these fields. - Trial records refreshed from ClinicalTrials.gov on 16 November 2026. No entry, count or coding changed.
v1.3.2 — 2026-11-15
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- A new Charts page shows when the registry's molecules were first dosed in people, year by year, what the developer's claim credits AI with, the therapeutic areas and the kinds of medicine. Each chart shows the strict entries or, with a switch, the borderline ones as well, and each has a table of its figures.
- No entry, count or coding changed.
v1.3.1 — 2026-11-15
Register: no entry added, removed or recoded · 75 strict, 95 borderline · data as released (zip)
- A new field,
modality_class, gives every entry one of seven classes: small molecule, antibody, other protein, peptide, vaccine or antigen product, cell or gene therapy, or other. The rules are in the methodology. The free-textmodalityis unchanged beside it. The registry page's modality column and filter now use the class. - Strict entries by class: 39 small molecules, 12 antibodies, 15 vaccines or antigen products, 3 cell or gene therapies, 3 peptides, 1 other protein, and 2 other (an oligonucleotide and an mRNA encoding a T-cell engager).
- The two T-cell therapies raised against antigens a model selected now count as cell therapies, not vaccines.
- GB-4362 (Generate:Biomedicines) is recorded as a monoclonal antibody, as its developer describes it; it was listed as another kind of biologic.
- No entry, tier or count changed.
v1.3.0 — 2026-11-15
Register: 19 added · 75 strict, 95 borderline · data as released (zip)
- Two reviews under the standing procedure (evidence from primary sources, two blind coders, every relied-on quote read again at its source), signed off by the maintainer and released together. Strict 70 → 75; borderline 81 → 95; broad count 151 → 170; excluded 60 → 68.
- A changed rule (review protocol version 1.8). 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. AI that worked only on a delivery system or formulation, such as a lipid nanoparticle or a new formulation of an existing drug, is now published as borderline under a new code,
delivery; until now it was excluded. AI used only in manufacturing is still not counted. All four surveys this registry compares itself with count METiS's AI-formulated tablet, MTS-004. - Under the changed rule, MTS105 (METiS) moves from the excluded list to strict: an mRNA whose whole sequence the developer credits to its algorithm, at medium confidence because the developer's paper does not call the algorithm a model. Found in the same review, AX-0811 (ProQR), an RNA-editing oligonucleotide that ProQR calls AI-discovered, is also strict; it would have counted under the earlier rule but had not been assessed. Both have the AI step marked as not stated.
- Moved from the excluded list to borderline
delivery: MTS-004 (METiS, a new tablet of dextromethorphan with quinidine, in Phase 3), MTS109 (METiS) and Covigenix VAX-002 (Entos). To borderlineplatform_only: Covigenix VAX-001 (Entos) and STX-001 (Strand), each named with a platform its developer says uses AI, in wording that appeared years after first dosing. - Nirmatrelvir (Pfizer, the antiviral in Paxlovid) moves from the excluded list to borderline
other. Pfizer wrote in 2022 that AI and machine learning helped design it; its detailed accounts credit structure-based chemistry and name no machine learning. Its exclusion reason, that the claim was not the developer's own, was wrong. It is the fifth approved drug among the borderline entries; the strict count still has none. - A one-time search of the trial records of AI drug-discovery companies over the previous eleven years, of which 38 products were assessed:
- added as strict, each at medium confidence with its caveat: iNeo-Vac-R01, a vaccine, and iNeo-Vac-T01, a T-cell therapy, both from Hangzhou Neoantigen Therapeutics, which says AI selects their antigens; and GUB-UCN2 (Gubra), a urocortin 2 analogue, whose defining positions Gubra's patent credits to a random-forest model;
- added as borderline
platform_only, each named with an AI platform but with no statement that a model did the work on it: EVM14 (Everest Medicines), Meta10-BCMA and Meta10-TIL (Leman Biotech), BBO-10203 (BridgeBio Oncology Therapeutics), BLX-7006 and BLX-0871 (Biolexis Therapeutics), and BMF-500 and BMF-650 (Biomea Fusion); - ADG206 joins the ADG106 entry: Adagene calls ADG106 its activated form, with the same binder;
- added to the excluded list: LY4292009 / MORF-440 (Morphic, Eli Lilly), BBO-11818 (BBOT), BX002 (BiomX), LTZ-301 (LTZ Therapeutics) and VB10.16 (Nykode), with no developer statement connecting AI to them; EG-501 (Evergreen), five Erasca molecules, EVER001 and VIS-101 (Everest Medicines) and HLX-4310 (Healx), in-licensed or acquired; HZ-A-018 (HealZen), dosed in 2019, with AI credited only from 2026; BX005 (BiomX), never dosed; and IMCY-0098 (Imcyse), where AI was used only to analyse trial data;
- icovamenib (BMF-219, Biomea) stays excluded, now because its AI wording appeared after the series was designed; ALTO-207 and EG-101 join existing exclusions;
- six products not yet dosed are now followed for first dosing.
- 17 of the 75 strict entries are now products whose antigens a model selected (15 before); two, cNeT and iNeo-Vac-T01, are T-cell therapies raised against such antigens. Leaving them out gives 58.
- Trial records: 348 (was 316). The strict cohort has 143 (was 130). iNeo-Vac-R01 also has an investigator-sponsored trial, recruiting since September 2025.
- Corrected: the methodology's Schrödinger sensitivity figure said 53, a figure left over from before the vaccine rule; it is 73 (75 without the two Schrödinger molecules).
v1.2.6 — 2026-11-15
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- Method text only; no entry, count or coding changed. The methodology's Updates section now describes the checks that have run every week since 30 October 2026: registries, approvals, developer news, programs not yet dosed, and a second model's reading of every AI claim. It also covers how each finding is decided, and the searches for molecules the registry does not yet follow, which began on 15 November 2026 and look forward only. The limits section now notes that WHO's registry portal returns only the ten most recent records for each search. The question page no longer says outcomes are recorded quarterly; trial records are checked weekly.
v1.2.5 — 2026-11-15
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- No entry, count or coding changed. DF-006 (Drug Farm) gains its second trial record, ACTRN12623000997651: the multiple-dose part of its Phase 1 study, in people with chronic hepatitis B (enrolment from July 2023; completed). Found by a new search of the WHO trial registry portal for trials sponsored by AI drug-discovery companies.
- PAN-NeoVax (West China Hospital) stays excluded as not yet dosed, checked again today. A second trial record, ChiCTR2600123936 (Phase 1, registered 1 May 2026), is recruiting; neither record reports a start, and no dosing report was found. Its exclusion reason now names both records.
v1.2.4 — 2026-11-15
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- Wording only; no entry, count or coding changed. The home page's reference to Haque et al. 2026 said their list of assets was not available. It is: their supplement lists all 129, and this registry's crosswalk download records how each is treated here. The methodology page already said so.
v1.2.3 — 2026-11-14
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- No entry, count or coding changed. Every recorded AI-claim quote was checked again against the page it cites: of the 149 pages that could be read, each carries its quote; two press releases block automated reading and were not rechecked (AC0176, ENV-6946). Zasocitinib's quote ran to 42 words, over the 40-word limit; it now ends at "(JH)2 domain binding", the same sentence from the same paper.
v1.2.2 — 2026-11-14
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- No entry, count or coding changed. The maintainer reviewed nine entries whose recorded quote, read on its own, pointed to the other tier, and confirmed each tier as it stands: ABBV-295, BEN-8744, AC0176, surbinsertib (SGR-3515), BEN-2293 and ENV-6946 (strict), and BBO-8520, BX004 and zasocitinib (borderline). Six confirmations are new in the audit log; three were already recorded.
v1.2.1 — 2026-11-14
Register: no entry added, removed or recoded · 70 strict, 81 borderline · data as released (zip)
- Wording only; no entry, count or coding changed. The home page now says the ten Phase 2 outcomes are "enough to start comparing" with the industry's success rate, and gives the figures beside it: 5 moved on, with a 95% confidence interval of 24–76%. The question page says that with ten outcomes the interval is wide and includes the industry median, so no difference can be read into it yet, and that ten outcomes is the floor for showing a comparison, not an adequate sample.
v1.2.0 — 2026-11-13
Register: 8 added · 70 strict, 81 borderline · data as released (zip)
- Eleven products were reviewed under the standing procedure (evidence gathered from primary sources, two blind coders, a red team on every call, every relied-on quote read again at its source) and signed off by the maintainer. Strict 64 → 70; borderline 79 → 81; excluded 57 → 60; broad count 151.
- Added as strict, each a product whose antigens a model selected: EVM16 (Everest Medicines), NEO-PV-01 (Neon Therapeutics, later BioNTech US) and VB10.NEO (Nykode); and, at low confidence, iNeo-Vac-P01 (Hangzhou Neoantigen Therapeutics), cNeT / ATL001 (Achilles Therapeutics, a T-cell therapy raised against model-selected antigens) and NEOVACC (University of Liverpool with myNEO Therapeutics). Each entry gives its basis and the recorded caveat.
- Added as borderline: pEVAC-PS (DIOSynVax and the University of Cambridge), because a university release credits machine learning for the antigen design while the developers' methods paper describes a design with no trained model; and HMBD-501 (Hummingbird Bioscience), which is credited to a discovery platform with no statement of what a model did for it and carries the same antibody as HMBD-001.
- Added to the excluded list, as dosed products with no developer statement connecting AI to them: PGV-001 (Mount Sinai), NOUS-PEV and NOUS-209 (Nouscom).
- 15 of the 70 strict entries are now products whose antigens a model selected (9 before). Every figure is still shown with and without them: leaving them out gives 55.
- The review protocol (version 1.7) writes down three readings for these products. A public prediction program that happens to be a trained network does not count unless the developer or a partner describes the selection as machine learning or AI, or the program as a trained model. A developer's detailed account of a selection method with no trained model is contrary evidence, while a paper that gives no account is only silent. A partner's AI system counts for a product only when the partner says it selected that product's antigens.
- Intismeran autogene: a 2026 paper by Moderna and Merck authors calls its selector "a validated bioinformatics algorithm"; its main text does not mention machine learning. This is recorded in the entry's notes; its tier is unchanged.
- This is still not a complete list of such products: Chinese-language registries and company sites, Japanese and Korean developers and academic trials have not been searched.
- Ten strict molecules now have a Phase 2 outcome (they moved on to Phase 3 or stopped), which is the number at which the question page compares their rate with published industry estimates: 5 of 10 moved on (50%, 95% confidence interval 24–76%). The home page now counts these by the same rule as the question page, where a combined Phase 1/2 trial counts as Phase 2; it had said 7.
- Trial records: 316 (was 299). The strict cohort has 130 (was 114).
v1.1.3 — 2026-11-13
Register: no entry added, removed or recoded · 64 strict, 79 borderline · data as released (zip)
- Registry IDs corrected on five entries, from records waiting in the review queue. No entry changes tier, code or count.
- Removed: ISRCTN18176375 from ADG125. It is a trial of a different drug that shares the code BC-006 (an obesity medicine from another company).
- Added: ISRCTN85112396 to ABBV-295 (its first study in volunteers); ten trials to bamlanivimab and one to bebtelovimab, including platform trials such as ACTIV-2 and ACTIV-3 in which they were one arm; one trial to ADG126 (in the ADG106 entry).
- The rule is now written down in the methodology: an entry lists the interventional trials that name the molecule as an intervention, whoever sponsors them; observational studies, expanded-access programs and records withdrawn before enrollment are not listed.
- Trial records: 299 (was 288). The strict cohort's 114 trials are unchanged.
v1.1.2 — 2026-11-13
Register: no entry added, removed or recoded · 64 strict, 79 borderline · data as released (zip)
- Each strict entry that has an inclusion note now shows it on the registry page, as "Basis for inclusion". Until now only borderline entries showed theirs ("Why borderline").
- All 112 inclusion notes are rewritten in plain language: what the developer said, what is missing or contrary, and a caveat where reviewers disagreed or the evidence is thin. Review shorthand and version references are gone; the earlier wording is kept in the audit log. Five borderline entries that had no note now have one. The same shorthand is removed from the general notes of 15 entries.
- No entry changes tier, code or count, and no new evidence was added: each note restates the entry's existing record.
- On phones, an open entry's details now fit the screen.
v1.1.1 — 2026-11-13
Register: no entry added, removed or recoded · 64 strict, 79 borderline · data as released (zip)
- Inclusion notes corrected on 14 entries. No entry changes tier, code or count.
- Eight notes that had been cut short are restored in full: BBO-8520, PEP08, RTX-117, AC0676, DPTX3186, HMBD-002, emunkitug (HFB200301) and ADG106. The passages they rely on were checked again at their sources.
- Five notes now agree with the entry's recorded "AI step stated" value, which is unchanged: zilemaltib (SGR-1505), surbinsertib (SGR-3515), SNX-281 and IAM1363 (stated), and DSP-0038 (not stated).
- GTAEXS-617 (REC-617): a note now explains the quoted line, which comes from a table in Exscientia's annual report, and cites two plainer statements by the developer: a 2022 conference abstract and a 2024 release. Three sources added.
v1.1.0 — 2026-11-13
Register: 9 moved tier, 1 borderline reason recoded · 64 strict, 79 borderline · data as released (zip)
- Vaccines whose antigens a trained model selected now count as AI-derived. Until now the registry listed them as borderline whatever the evidence, a rule it had never given a reason for. From this version they are judged like every other entry: the developer's or a partner's own evidence must show that a trained model selected the antigens the product carries, whether those are the same for every patient or chosen for each one. The published surveys of AI-derived drugs count such vaccines too.
- 9 entries move from borderline to strict: GRANITE, SLATE, GRT-R910, EVX-01, EVX-02, TG4050, NECVAX-NEO1, intismeran autogene (mRNA-4157) and autogene cevumeran (BNT122). Two of them, GRT-R910 and autogene cevumeran, are at low confidence, and every entry records the argument against it. Tamavaq stays borderline: its developer's own detailed account describes a rule-based selector, and no participant is confirmed vaccinated.
- Strict 55 → 64; borderline 88 → 79; the broad count stays at 143. No new molecule was found and nothing new happened in a clinic: the rise is a change of rule.
- Each of the ten entries was reviewed afresh: evidence gathered again from primary sources, two blind coders, a red team arguing against every call, and every quote checked at its source. The review protocol is now version 1.6.
- A new value in the data: the AI step of these nine is
antigen. Figures by AI step keep their meaning, and the phase-transition figures are shown with and without the nine. Their target type reads "Not classified" (formerly "Undisclosed"), which now also covers products with no single target. - What moves with them: the earliest first dosing in the strict cohort is now 2017 (intismeran autogene and autogene cevumeran; their developers' statements tying AI to them date from 2025); 16 entries have reached Phase 2 or later (was 12) and 4 are in Phase 3 (was 3); 17 are stopped (was 14); vaccines are now 9 of the 64, and small molecules 39.
- These nine are not a complete list of such vaccines. A first search found three more likely candidates, which will be reviewed.
- Trial records refreshed: 288 trials (was 277). The strict cohort now has 114 ClinicalTrials.gov trials (was 80), including 8 more for intismeran autogene and 2 for autogene cevumeran found by name, and one new record for ABBV-295.
v1.0.16 — 2026-11-09
Register: 4 removed, 7 borderline reason recoded, 1 status changed · 55 strict, 88 borderline · data as released (zip)
- The 19 repurposing entries have been reviewed under a stricter rule: a repurposed drug counts only when the developer's evidence shows a trained model picked the new use, or selected the drug for it. A use already tested in people before the AI step does not count, and people choosing from a model's output is assisted review, not an AI pick. Two blind coders and red teams reviewed each entry.
- 4 entries excluded: SOM3355 (bevantolol) and PXT3003, because the developers' own accounts from the time describe non-AI methods and the AI wording came later; REC-2282, because it had been tested in people with NF2 in 2012–13; and azelaprag, because the developer credits its obesity use to an unexpected mouse result. LAM-001 (inhaled sirolimus), previously part of the LAM-002A entry, now has its own exclusion, as nothing ties AI to it.
- 7 entries stay borderline under another reason: OC514, baricitinib, INV-001 and INVA-8001 (other), OC201 + OC202e and REC-3599 (platform-only), and bavisant (third-party only). 8 stay repurposing.
- Borderline 92 → 88; the broad count, strict and borderline together, 147 → 143. Strict stays at 55.
- LAM-002A now leads with C9ORF72 ALS and is recorded as stopped: its own trials are complete and its developer now focuses on LAM-001. BXCL701 adds Point Therapeutics' earlier pancreatic cancer trial (NCT00116389). Each reviewed entry records the methods its developer describes.
v1.0.15 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 92 borderline · data as released (zip)
- In the comparison of published counts, each list's heading now links to its full citation in the bibliography. McKinsey's column is labelled as 9 examples: its exhibit is a non-exhaustive set of AI-enabled assets chosen because their discovery times were disclosed, each already in clinical trials, rather than a survey.
- No entry, count or coding changed.
v1.0.14 — 2026-11-08
Register: BXCL701: added as borderline · 55 strict, 92 borderline · data as released (zip)
- The last two molecules on Haque and colleagues' list that this registry had not yet assessed have been reviewed under the written protocol (two blind coders, a red team and the maintainer's sign-off). Every entry on their list is now assessed.
- BXCL701 / talabostat (BioXcel Therapeutics) added as borderline (repurposing). BioXcel says its AI platforms found new uses, with checkpoint inhibitors in prostate and pancreatic cancer, for a compound Point Therapeutics first dosed in 1999. It is recorded as stopped, because BioXcel has deprioritized it. Borderline 91 → 92; the broad count, strict and borderline together, 146 → 147. Strict stays at 55.
- Tolcapone (SOM0226, SOM Biotech) excluded. The developer repositioned this approved Parkinson's drug for transthyretin amyloidosis by virtual screening and docking. Its AI wording came years later and never names the molecule.
- Haque and colleagues' list now places 38 of their entries as strict here and 96 as strict or borderline; the other 33 are excluded.
- Meta10-19: the developers' Lancet Haematology paper names no AI, and the IL-10 approach behind it predates the AI partner's involvement. The B-ALL trial first enrolled patients on 16 May 2023. Its coding is unchanged.
- 6 trial records added for BXCL701 (292 in all).
v1.0.13 — 2026-11-08
Register: 9 added, 1 borderline reason recoded, 1 target recoded · 55 strict, 91 borderline · data as released (zip)
- The 27 molecules on Haque and colleagues' list that this registry had not yet assessed have been reviewed under the written protocol (two blind coders each, a red team on every disagreement, and the maintainer's sign-off). None qualifies as strict.
- 9 borderline entries added: NEU-111, NEU-627, ATV-1601, DF-006 and Meta10-19 (platform-only); OC201 + OC202e, INV-001 and INVA-8001 (repurposing); GTA182 (other). Borderline 82 → 91; the broad count, strict and borderline together, 137 → 146. Strict stays at 55.
- 18 molecules excluded, in 13 rows: 10 in-licensed (including Alto Neuroscience's five), 3 with no AI claim about the molecule, 3 not yet dosed (Path-001, EG-007, EG-301), 3 where the AI claim is about the delivery system rather than the vaccine (Entos's Covigenix vaccines, now one row), and 1 where the developer's account names a non-AI origin (ERAS-007).
- Zasocitinib is recoded from physics-based to borderline "other", with tier A attribution: Takeda's peer-reviewed paper credits an AI-assisted design tool, which rules out physics-only, but the developers' accounts of the design credit physics-based calculations and docking. It stays borderline.
- DF-003's target is recoded from first-in-class to novel: DF-006, from the same developer and against the same target, entered Phase 1 first.
- Haque and colleagues' list now places 38 of their entries as strict here and 95 as strict or borderline; 32 are excluded and 2 not yet assessed.
- 11 trial records added for the new entries (286 in all).
v1.0.12 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Comparison with Haque and colleagues' list corrected: four of their entries were already in this registry under other names (ENT-H-1 is IAM1363; amelenodor is NX-13; nezutatug is HMBD-001; LAM-001 is part of the LAM-002A entry). 38 of their entries are strict here and 86 are strict or borderline (was 37 and 82).
- Strict entries on none of the published surveys: 18 of 55 (was 19).
- No entry, count or coding changed.
v1.0.11 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- The data-as-of date is corrected to 5 November 2026, the last day the register's data changed (it had stayed at 2 November). It now moves with every change to an entry, exclusion, trial record, readout or developer detail, and the methodology defines it.
- No entry, count or coding changed.
v1.0.10 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Haque and colleagues' 129 AI-enabled assets are now placed in the comparison of published counts, from the list in their paper's supplement: 37 of their entries are strict here and 82 are strict or borderline. Most of the rest are molecules this registry has not yet assessed.
- The methodology now explains that lists of "AI-enabled" assets compare with this registry's strict and borderline entries together (the broad count, 137), not with the strict count.
- Strict entries on none of the published surveys: 19 of 55, down from 29, now that Haque's list is included.
- No entry, count or coding changed.
v1.0.9 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Bibliography: Alacrita's white paper on the clinical record of AI-derived molecules, their valuation and the regulatory instruments added under Further reading.
v1.0.8 — 2026-11-08
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Haque and colleagues' count of AI-enabled assets now cites their full paper in Cancer Treatment Reviews (October 2026): 129 assets from 69 companies in trials by March 2026, up from 117 in their 2025 conference abstract. Its place in the comparison of published counts is unchanged. No entry, count or coding changed.
v1.0.7 — 2026-11-05
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Bibliography: Lindus's account of a Phase 2a trial in ME/CFS run for under $1 million with AI-assisted trial operations added under Further reading, as a case of AI lowering trial costs rather than designing drugs.
v1.0.6 — 2026-11-05
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- PHOE-01, tested in the ReMEdi Phase 2 trial in ME/CFS (ISRCTN74088209), added to the exclusions: its AI claim covers running the trial (documents, recruitment and data checks), not choosing the target or designing the drug, which is undisclosed and likely an existing medicine. Strict and borderline counts unchanged.
v1.0.5 — 2026-11-05
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Inclusion notes on 66 entries no longer say "pending sign-off": the re-review they record was signed off on 27 October 2026. Wording only; no entry, inclusion or coding changed.
v1.0.4 — 2026-11-05
Register: no entry added, removed or recoded · 55 strict, 82 borderline · data as released (zip)
- Borderline entries now carry their ClinicalTrials.gov records: 195 trials for 79 borderline molecules, fetched by the trial IDs already listed for each entry, so their rows show trial status and the next data due. The trial outcomes and the industry comparison still use the strict cohort only.
- Trial records re-fetched on 5 November 2026. One strict trial closed (AU-409's NCT06374485, completed): of the strict cohort's closed trials, 8 of 25 were terminated (was 8 of 24); of those matched to the industry comparison, 8 of 21 (was 8 of 20), against 18.6% across the industry (comparator re-run, unchanged).
v1.0.3 — 2026-11-05
Register: 2 added · 55 strict, 82 borderline · data as released (zip)
- Two molecules added as borderline
physics_only(computational design; the developers' statements about each name no machine learning): AJ1-11095 (Ajax Therapeutics, now Eli Lilly, designed with Schrödinger), a Type II JAK2 inhibitor in Phase 1 for myelofibrosis, and LTSE-2578 (Structure Therapeutics), an LPA1 antagonist for idiopathic pulmonary fibrosis whose Phase 1 completed in July 2025 and for which the developer is now considering strategic alternatives. Both went through the standing review (two blind coders, who agreed; a red team on each). Strict unchanged at 55; borderline 80 → 82. Counting physics-based design, the sensitivity figure is 72 (was 70).
v1.0.2 — 2026-11-03
Register: no entry added, removed or recoded · 55 strict, 80 borderline · data as released (zip)
- The changelog now opens with the changes to the register: every release that added or removed an entry, moved one between strict and borderline, or recoded one in a way that moves a published figure (target novelty, furthest phase, status or an approval), from v1.0.0. Each release's notes start with the same line, and the home page names the latest change.
- New downloads: every change to the register (
register-changes.csv), every release with its strict and borderline counts (releases.csv), and each release's data exactly as released, from v1.0.0 (a zip per release). Data for the pre-release versions before v1.0.0 is available on request. No entry, count or coding changed.
v1.0.1 — 2026-11-03
Register: GB-7624: target now known (was undisclosed) · 55 strict, 80 borderline · data as released (zip)
- GB-7624 (Generate:Biomedicines) is an anti-IL-13 antibody, as Generate's February 2026 prospectus states; its target was recorded as undisclosed. IL-13 already had an approved drug (tralokinumab, 2021), so GB-7624 now counts as a known target: 35 of the 51 strict molecules with a disclosed target are on known targets (was 34 of 50). Novel targets are unchanged at 16.
- ISM3312's AI claim now quotes the sentence from Insilico's 2025 paper that names the molecule: "the novel Mpro inhibitor ISM3312 generated by a generative AI engine". Its inclusion is unchanged.
v1.0.0 — 2026-11-02
Register: First citable release: 55 strict, 80 borderline · 55 strict, 80 borderline · data as released (zip)
The registry is formally released as version 1.0, alongside the essay Turning the Crank Faster in Backed By Reality. The pre-release notice is removed and the figures may be cited. Cite it as: Milad, J. E. AIxBio Registry: AI Molecules in the Clinic (v1.0.0, data as of 2 November 2026). https://aixbio-registry.org - Data refreshed to 2 November 2026: trial records re-fetched from ClinicalTrials.gov (no changes), registries searched (14 EU and ISRCTN trial IDs added to six borderline molecules), developer news for the past 45 days reviewed. No count moved: strict 55, borderline 80, approved 0; Phase I→II 11 of 22 (50%), Phase II→III 3 of 6.
Before v1.0.0
The registry was drafted in public from 16 October 2026 as pre-release versions (0.x), marked as not for citation. Its figures settled at v1.0.0, the first citable version and the baseline for this changelog. The draft history is kept on file and is available on request.