AI-Enabled Drug Discovery, 2020–2026

I The one computational platform that did capture asset value took fourteen years, held 3.8% of one company among many collaborations, captured 2.8% of the upfront, and remained loss-making throughout OPS-0286. recorded, not reconciled The same package states the interval as “thirteen years after the collaboration began” where it reports the economics and “fourteen years” where it argues NR4; both figures are the package’s own and neither is preferred here.

Against that: 23andMe US$305M and Invitae US$239M, both sold out of bankruptcy within sixteen months OPS-0113, OPS-0114, OPS-0140; Insilico’s pipeline upfronts US$58.0M → US$15.3M; AbCellera US$1.0M of milestone revenue across 104 partnered programmes OPS-0122, OPS-0127. This is the only recommendation in the package that survived every correction in the audit without amendment.

I What NR3 costs if the sponsor waits for the covered-person determination. Nothing, if the determination takes weeks. That is the point of putting it in the no-regret set.

I The methodology’s test T1 binds: “A candidate may be described as ‘AI-designed’ only at grade A1.” This track graded zasocitinib A3, on the sponsor’s and vendor’s own language: no generative model proposed the molecule, the 4,000 candidate structures were enumerated by chemists, and the JH1→JH2 pivot was a human judgement on published Bristol Myers Squibb biology. Both available branches give the same answer. If FEP+ is AI/ML, it is A3, and A3 is not AI-designed. If FEP+ is not AI/ML at all — which the peer-reviewed paper’s own wording supports — the molecule is not on the A1–A5 ladder and D1 has no candidate.

The earlier assertion that the methodology “never fixed that boundary” is withdrawn as false: T1 fixed it, and this track applied T1 to reach A3 in the same paragraph.

I The definitional dispute is real and belongs elsewhere. Whether the sector’s phrase “AI-enabled drug discovery” includes physics-based computational chemistry is unresolved and decision-relevant. It is a governance question, not an evidentiary one, and treating it as evidentiary let a word-choice argument masquerade as measurement uncertainty.

attribution The three-way reviewer records that the Claude track made this boundary ruling explicitly and correctly in its own ledger, and that the Codex track never examined the class at all. Those are the reviewer’s findings about other packages; this track has not read them and does not rely on them.

F query.spons matches the collaborator field as well as the lead-sponsor field: NCT04622527 returns under “Relay Therapeutics” with leadSponsor = Mayo Clinic and collaborators = ["Relay Therapeutics, Inc."] OPS-0295.

I That is systematic over-inclusion on collaborator relationships, not a false positive. The earlier claim that “a false positive proves the query is not a filter” is withdrawn — it proves the query filters on a wider field set than assumed.

F The censoring class that actually mattered runs the other way, and here it is measured. query.spons=Nimbus Therapeutics returns 2 studies — sponsored by “Nimbus Wadjet, Inc.” and “Nimbus Saturn, Inc.” — neither of which is a zasocitinib study, while query.term=zasocitinib returns 12 OPS-0296.

I The entire twelve-study zasocitinib franchise is sponsor-invisible: Takeda is the lead sponsor and neither originator appears in any sponsor or collaborator field of any of the twelve records.

The three-way reviewer re-executed the queries live on the as-of date and reproduced them exactly: Insilico totalCount 11; Recursion 9 studies with terminated NCT06005974, NCT06536465, NCT05130866 = 3; query.spons=Schrodinger totalCount 0; Relay 6 returns of which 5 Relay-sponsored across 3 molecules plus 1 Mayo Clinic study.

The reviewer’s verdict: the construction rule is genuinely reproducible from the stated text — the rarest property in this programme. One censoring class was mischaracterised (corrected above) and one property was not stated: whether the interventional filter was applied per sponsor string. Re-execution without the filter returns the same totals for Recursion and Relay, so the effect is nil here, but the rule is not checkable as written. The remedy is to record the exact query string per row.

Counting note: the reviewer’s “seven fail” comprises C1–C7, all graded IND-0 or IND-1. C8 is graded IND-2 and is published in the same table because its failure mode is framing rather than sourcing.

• C9 — headline deal value is not realized value. IND-3, the most robust finding in the programme. Three different companies, three different disclosure regimes. This track’s leg: Insilico audited pipeline upfronts US$58.0M → US$15.3M and AbCellera US$1.0M of milestone revenue across 104 partnered programmes OPS-0122, OPS-0127.
• C10 — do not build or buy an AI discovery platform. IND-3. Three genuinely different evidence bases.
• C11 — AI clinical-operations tools improve speed, not yield. IND-2. Two studies, not three; two packages cite the same randomised trial.
• C12 — the largest measured effects on development success are effects of evidence type, not of modelling. IND-3, and the reviewer records it as this track’s distinctive contribution.
• C13 — foundation models lose to simple baselines on perturbation prediction. IND-2.
• C14 — the sponsor should sell governed data and evidence, not buy models. IND-3.

• SD1 — no return side anywhere. No NPV, no payback, no discount rate, no hurdle, in any of the three. No board can size any of the three capital asks. This track’s response is §9’s break-even condition.
• SD2 — Medicare-only economics. Every rate quoted anywhere in all three packages is a Medicare national limitation amount. Commercial payer mix, denial rate and collection lag are unmeasured in all three.
• SD3 — attribution taken from sponsors. Three packages, one method, one failure mode. Zasocitinib was missed by two of three precisely because the method reads sponsor self-description. See §5.
• SD4 — China evidence thin in all three, and only one says how thin. Six failed attempts across five artifacts and five mechanisms, quantified here as limitation L4 OPS-0291.

The band “across every AI-platform deal where both figures are disclosed, upfront consideration is 1.2%–6.0% of headline, median ≈3%” appeared in 10 locations across 5 artifacts. It was falsified by a transaction the track itself located late in the programme, and the propagation failure was measured mechanically and found to be 67% larger than the audit that flagged it had found.

F Out-licensing transactions — the population the band was actually drawn from:

F Whole-asset acquisitions — a different population, n=1:

I The withdrawal, stated plainly. The framing that the exception sits “two orders of magnitude away from the rule” is withdrawn. There is no rule and no exception. There are two populations measured under one label, and only one of them has enough observations to carry a band. Three properties make the acquisition row incommensurable: transaction type (an entity purchased, not a relationship licensed); trigger class (sales-based only, against tails spanning research through commercial — and test T2 makes trigger class mandatory, which was applied per row and then not applied across the comparison); and stage (a de-risked Phase 2b asset with positive data, against platform access at signature).

I The conclusion that survives, and it was always the load-bearing part: the upfront share is a function of how de-risked and how separable the thing being sold is, not of whether a computer designed it. F The matched non-AI comparator — Quotient–Merck at 0.91% — is itself an out-licensing transaction and therefore in-population, so the finding it supports is unaffected and strengthened: ~1–6% at signature is the ordinary structure of a discovery alliance, not an AI-specific penalty.

R The sponsor-facing instruction changes with the correction. An earlier artifact named the fragile cell as “whether upfronts of this kind are ~3% sector-wide or specific to AI deals”. Both horns are wrong. The ratio is specific to transaction form. Sell a relationship and expect 1–6% at signature whether or not AI is involved. Sell a separable, de-risked asset that a buyer wants and 66.7% in cash is achievable. Treat every milestone as worth approximately nothing at signature — the position the counterparties’ own auditors take under ASC 606 — and negotiate on separability, not on technology narrative.

I The finding is right and the concession is unqualified. Test T1 states a binding rule in one sentence: “A candidate may be described as ‘AI-designed’ only at grade A1, and only where the sponsor’s own language supports it, quoted.” The track graded zasocitinib A3 and gave three reasons in the sponsor’s own terms. Having done that, the scoreboard then recorded D1 as contested and justified it by asserting that the methodology “never fixed that boundary”.

That assertion is false, and it is withdrawn. T1 fixed the boundary. The same paragraph that recorded the contested status had already used T1 to grade the molecule. A defeater class defined in terms of “AI attribution” is scored by the instrument that grades AI attribution, and that instrument returned A3.

F The second branch is worse for the contested reading, not better. The peer-reviewed discovery paper was re-retrieved by an independent route OPS-0307. The full abstract contains three method statements: “A computationally enabled design strategy, including the use of FEP+”; “computational physics-based predictions used to optimize this series”; and a therapeutic rationale opening on “compelling data from human genome-wide association studies”. No machine-learning, deep-learning, generative or neural method is named anywhere. There is no reading of this track’s own rulebook on which D1 is contested.

R The change makes the package stronger, and that is the reason to accept it rather than defend the original. A scoreboard that flips on a word tells a board nothing. Two scoreboards that each say something true tell it a great deal.

S1 finding O-2 — the executive layer published a ranking the track’s own next artifact had already superseded. Artifact 07 re-scored two economics cells on measured reimbursement evidence this track retrieved itself: F1 (data licensing) D4 from 4 to 3, on one comparator’s Biopharma & Data line growing 9% against 44% for the company and another’s licensing line at 0.40% of revenue OPS-0259, OPS-0260; and F7 (AI prognostic software) D4 from 3 to 2, on the measured CLFS rate of US$706.25 against MRD and CGP codes. D4 carries weight ×2.0. The executive report nonetheless reprinted the pre-correction weighted table.

I The consequence runs against the more comfortable reading. The two families that fall are the fastest and the most AI-native of the leaders. The family that rises to sole second — companion diagnostic co-development — is the slowest and most capital-hungry. The correction pushes the recommendation further away from “do something quick and AI-adjacent” and further toward “do the slow regulated thing”, which is the opposite of what a track wanting a tidy answer would choose.

I And it leaves a residue the original hid. F7’s margin over F4 falls from 2.5 to 0.5 on a 55-point scale. A half-point margin is not a ranking. They are reported as joint fourth, and the composite is stated to have run out of resolving power there.

S1 finding O-1 and the label defect. The claim “the one authorised AI prognostic device’s product code is paid US$706.25” was published as F while the supporting ledger row itself states “Code-to-product identification is not in the file… Any mapping of a code to a named test in this track is labelled I” — and the methodology forbids an I being “laundered into F in a later artifact”. Relabelled. Separately, “product code” is an FDA device term (here SFH); 0376U is a CPT PLA code. Two different identifier systems, now kept in separate columns.

I Why it happened, stated as a method finding rather than an excuse. This track adopted a convention of correcting forward rather than silently re-issuing earlier tables. That convention is defensible and preserves the record of what was believed when. But a convention of correcting forward imposes an obligation to actually carry the correction forward, and the track failed the forward limb while congratulating itself on not failing the backward one. The procedural remedy: any artifact that re-scores a cell must reprint the affected composite in full at the point of re-scoring.

The Schrödinger arithmetic F

From the issuer’s own release, re-read OPS-0286, corrected in place 2026-08-04: US$111.3M received February 2023; US$36.0M further expected in Q2 2023; US$147.3M total; a 3.8% fully diluted stake as at 2022-12-31; against Takeda’s US$4.0bn upfront payment. I 2.8% of the upfront on the first distribution; 3.7% including the second. The published 2.5% was wrong in the direction of this track’s own prior, which is the direction an adversarial posture is obliged to check hardest — and it is recorded as an instance of declared error class E1 rather than as a typo.

Insilico, restated from the audited filing rather than compressed F

F US$296.7M of the US$352.3M loss — 84.2% — is a non-operating fair-value charge on the conversion of preferred shares at the December 2025 listing. CR status-sensitive The same company guided on 2026-07-09 to H1-2026 revenue of US$102.5–106.5M (+272.7% to +287.3%) with net profit US$33.5–39.5M OPS-0099, unaudited.

I The corrected reading makes the null-first point better than the compressed one did. The finding is not that the AI-native cohort is collapsing. It is that a company’s revenue moved −34.5% and then, on guidance, +280% in consecutive periods while its revenue mix inverted from selling rights to selling services and its geography inverted from the US to the Chinese Mainland. Lumpiness of that magnitude is itself the reason no multi-cycle economics claim can be made in either direction — which is precisely what D4’s unmet limb says. The compressed version was more useful to this track’s prior; the audited document does not support it.

The trial records, paired per trial F

Re-queried per record OPS-0294. Published in this format because the parallel-list format used previously is what made a transposition possible.

F No resultsSection is present for either record at the as-of date OPS-0289. The design, enrolment, completion and pre-registered endpoints are F at Tier 1; the efficacy figures are CR until the registry posts them. Standing item 74.

1. The twelve orphan ledger rows stay. Nine of them are numbered retrieval failures. Deleting them would improve the ledger’s cosmetics and destroy its evidential value: they are the record that the retrieval was attempted and what it would have supported. One clarification is added: nine of the twelve orphans are the same nine failures counted inside the 54-row unverified-blocked total, so the orphan finding and the blocked-retrieval finding are one finding seen twice.
2. The disclosure that inverted the track’s own headline stays in the opening section, in full. The scoring inside it was wrong; the decision to put it first was right, and moving it now would be the exact failure the posture is prone to.
3. The declined Tier 4 row stays. The sector’s information environment — “173 programmes”, “80–90% Phase I success”, “15 drugs in Phase 3”, none traceable to a primary source, several contradicting figures verified here directly — is itself a finding about the diligence problem the sponsor faces, and it is recorded as a numbered row precisely so that it is visible that none of it was used.

One audit finding was rejected outright. The audit recomputed the ledger’s verification composition on a “strictest-state-wins” rule over whole-row text and got 223/9/54/7 against the published figures. Extracting column 11 of the row schema — the single-valued verification field — reproduces this track’s published figures exactly, with zero residual. The rows the audit re-assigned are substitution rows, governed by an explicit rule: “the row records both URLs, the tier of what I actually read, and the fact that the primary was blocked… The claim inherits the tier of what I read.” “Strictest state wins” is a different rule; applying it silently changes what the field means. No change to the figures.

One was half rejected on the text of the rule the audit invoked (a permitted-verb objection to “outscored… blind”: CASP16 is not a retrospective benchmark, so a prospective verb is the correct stage description). And one was accepted and then extended against this track’s own interest, because the mechanical count showed the propagation defect was 67% larger than the audit found.

F OPTIONS OPS-0198, OPS-0308: single-centre randomised trial, N=20,707, 2:1 allocation (13,802 / 6,905). The randomised intervention is an email notification to the treating oncologist about genomically matched trials, for patients whose imaging an AI model flagged as progressing with elevated probability of treatment change. Primary outcome, enrolment in any therapeutic trial: 2.20% (95% CI 1.97–2.46) vs 2.03% (95% CI 1.72–2.39); difference 0.18 pp; P=.41. Authors’ stated limitations include a 48–72 hour notification delay from EHR data flow and a 50.33% per-clinician survey response.

I Corrected statement. A randomised test of AI-triggered clinician notification was null; matching accuracy itself was not the randomised variable. What this is evidence for is more useful than a domain verdict: the AI models did identify the patients, and surfacing that to oncologists did not change enrolment. The bottleneck in trial enrolment is not detection. A sponsor selling enrolment services should price that.

This remains the only randomised, disinterested, adequately powered test of any AI clinical-operations intervention located anywhere in this programme.

attribution The three-way reviewer’s adjudication of the reimbursement anchor: this track never separates coverage from payment as cleanly as the Claude package does, and never reaches the ADLT mechanism at all — and the reviewer records that no ADLT grant had published in seventeen months at the as-of date, which routes this track’s plan toward MRD/CGP rates obtainable through a mechanism whose recent throughput is zero. Where this track is right and the others thin: the direction of travel. Others measure levels; this track measures the derivative and finds it negative.

These are the reviewer’s findings about packages this track has not read. They are reported as the reviewer’s, not adopted as this track’s own evidence. What this track does adopt: gate G3 in §11 makes coverage a separate and binding gate from payment.

• The 2032–2035 approval cohort is being determined now, in Phase 1. Molecules dosing first patients in 2026 have registered Phase 1 primary completions in 2027–2029 OPS-0215, OPS-0216, OPS-0219. On the industry likelihood of approval from Phase I of 7.9%, or 15.9% biomarker-preselected OPS-0151, the ~29 non-Phase-3 molecules in the cohort floor imply an expectation of roughly two to five approvals across 2032–2038. I That is the first cohort large enough that a rate rather than a count could in principle be computed — and it will only be computable if the denominator problem is solved, which cannot be done from the public record as it currently exists.
• The European Health Data Space is the structural threat to the sponsor’s core asset, and it commences in the window’s last year. Regulation (EU) 2025/327 applies from 26 March 2027, with Chapter IV (secondary use) applying from 26 March 2029 and staged provisions to 2031 and 2035; the categories available for secondary use expressly include human genetic, epigenomic and genomic data OPS-0166. I The erosion of proprietary-cohort rent is more likely to come from a regulation than from a competitor — which is why the outcome layer’s value is in the linkage and outcomes, which EHDS does not commoditise, rather than in the genomes, which it may.
• Prognosis versus prediction is the line to watch in the sponsor’s own product category. FDA has authorised an AI device for prognosis (product code SFH, one device OPS-0190) and has never authorised a machine-learning companion diagnostic across 235 device–indication rows OPS-0188. I Whether that line moves by 2035 is the single most important regulatory question for the sponsor’s product strategy, and it is entirely observable at zero cost.
• China’s regulator has published a two-horizon plan to rebuild itself around AI by 2035 — an initial regulatory-AI integration system by 2030 and a smart drug-safety governance framework by 2035 OPS-0022. Critical scope finding: it addresses AI applied by the regulator to review and inspection, not AI applied by industry to drug discovery. It is evidence about regulator capability, and is not evidence of a Chinese regulatory pathway for AI-designed drugs.

F Forty-three contradictions were recorded across artifacts 01–08 and none was averaged away. One (C41) is now resolved and removed from the carried set. Those that most affect the decision: