Where a genetic testing company should participate in AI-enabled drug discovery

I The single methodological decision that produced this package’s distinctive content was to stop reading about facts and go to the system a state compels someone to populate. Each artifact escalated to a new register, and the verification rate rose with it.

The Phase 04 lesson, and the two it generalises

I A retrieval failure must be attributed to its actual layer — a 404 is not evidence the URL is wrong. An unreadable payload is not evidence the data is absentSRC-0152capped. Phase 04 adds: a sample is not evidence about a population when the population is published. The thirteen-code gapfill median that drove the previous report’s largest correction was correct and was the best available answer; the whole CY2026 fee schedule, holding 194 algorithm-priced codes, was one file away the entire timeSRC-0157.

What the audit changed

F Of 21 findings raised against this package by an independent audit: 13 accepted in full, 6 partially accepted, 2 rejected. Eleven material changes followed, of which nine cut against this package’s own prior conclusions. I That is the expected distribution when an audit is doing its job, and it is reported this way deliberately: a revision in which everything came out in the author’s favour would be evidence of a bad revision.

F Six decisive claims changed. D1 withdrawn — “exactly three AI-attributed Phase 3 programmes” was a universal this package could not support. D38 restated after the rentosertib full text was retrieved. D4, D28, D30, D32 and D37 demoted, because their sole supporting rows are capped at verified-metadata. A mechanical harness now enforces the rule that produced the error: it ran 52 checks with zero failures when the audit was written and 74 checks with zero failures after.

I The audit’s central sentence — “the package is better at finding evidence than at propagating it” — is correct and is now a rule in the methodology rather than an observation about one document.

The two rejections, stated so they can be argued with

1. The interested-party objection to the commoditisation source is rejected. A company that open-sources under an MIT licence the capability it sells is producing evidence against its own commercial interest, and the finding rests on the licence and the compute ratio rather than on the accuracy claim — though the accuracy claim is restored to its weaker published form throughoutSRC-0025capped.
2. The proposal to raise the AI-target-selection evidence score is rejected. It reads on whether AI target selection has clinical support; it has been tested twice and failed twiceSRC-0051SRC-0060. Importing a genetics result to lift it would be a category error.

What crossed

F Boltz-2 was released under an MIT licence, reported as approaching free-energy-perturbation accuracy on small-molecule affinity at ≥1,000× the compute efficiencySRC-0025capped: preprint, abstract-level.

What did not change — the part that matters to a capital decision

I No approval. No completed pivotal readout. No measured cohort-level improvement in phase transitions attributable to AI. The only systematic peer-reviewed attempt stops in 2023, is self-selected and survivorship-biased, and its own authors treat 80–90% Phase 1 success as an upper boundSRC-0015capped. Against it, the industry base rate is 7.9% Phase I to approval across 9,704 programmes and 12,728 transitionsSRC-0163capped.

The only measured lever, and it is not an AI result

F Drug mechanisms with human genetic support show a 2.6× greater probability of successSRC-0018capped. E Against a 7.9% Phase I-to-approval base rate that is ≈20.5% absolute: roughly four in five genetically supported programmes still failSRC-0163capped.

I This report uses that finding against the AI thesis, not for it — which is why the absolute framing strengthens the argument rather than weakening it. The multiplier is a fact about genetics. Importing it to lift an AI-target-selection score would be the category error this package exists to avoid.

Capital raised has decoupled from evidence produced, and the decoupling is measured

F Isomorphic Labs raised US$2.7bn in fourteen months against US$82.5m of disclosed upfronts, with zero registered interventional trialsSRC-0084. insitro likewise stands at zero. F Six terminations carry sponsor-supplied stop reasons and actual enrolments of n=3 to n=54SRC-0051SRC-0060.

F And where consolidation has happened it has been distress-driven, not strategic: Exscientia’s audited cumulative revenue was £93.4m against £342.5m of losses, with revenue peaking in 2021, before the combination with RecursionSRC-0080SRC-0007capped.

F Rentosertib’s Phase 2a, retrieved in full textSRC-0003:

F There were four arms, not two. The 30 mg once-daily arm performed worse than placebo. There is no monotonic dose-response on the efficacy endpoint, and no between-group statistical test or p-value was reported for FVC. The trial ran at 21 sites.

I Absence of dose-response is the standard reason to treat a small-trial efficacy signal as chance rather than pharmacology. There is also a statistical trap here that a reader will walk into unwarned: one confidence interval excludes zero and one does not, which reads as a demonstrated between-group difference that was never tested. The counter-consideration is real and is kept: the hepatotoxicity pattern is dose-ordered (0% / 22.2% / 17%), so the trial is not free of dose-related structure — it is free of it on the endpoint being cited.

I The previous version of this report described this as an “encouraging-but-exploratory” result with correct caveats about sample size, duration, geography and authorship. Every one of those caveats was right and none of them was the problem. The problem was that the source had not been read. The separate and stronger claim — that rentosertib carries the sector’s only A1-level attribution — is untouched and stands.

Attribution runs inversely to clinical maturity

F The most clinically advanced computationally designed molecule is zasocitinib, and its Phase 3 superiority release makes no computational or AI claim whatsoeverSRC-0063; the design-attribution leg rests on a vendor case studySRC-0066capped and is demoted out of the decisive register accordingly. I The AI label is applied most loudly where clinical evidence is thinnest. Treat it as a marketing variable in diligence, not an evidentiary one.

The negative record, from registry documents

F The two direct clinical tests of platform target selection were both negative: VRG50635 in ALS, with the neurofilament biomarker moving the wrong waySRC-0051SRC-0060, and the EXS21546 combination terminated at n=6. F A second, undisclosed Western Phase 2a of rentosertib has a primary completion date that has silently elapsed. F Foundation models lose to simple baselines on perturbation predictionSRC-0021, and pathology foundation models are unrobust to medical-centre differencesSRC-0020capped — the specific failure mode a multi-site validation is designed to catch.

CR For balance, the one AI-attributed programme with a randomised mid-stage efficacy claim outside IPF is MDR-001, whose sponsor reports 8.2–10.3% weight reduction at 24 weeks against 2.5% for placebo in a randomised Phase 2bSRC-0162capped. It is a company report, it is not independently validated, and it may not alone support a decisive claim.

F Headline versus realized, audited. Recursion discloses ~US$19.4–20.5bn of collaboration headline value and states it has “fully constrained” the remaining milestone consideration under ASC 606, booking US$382.7mSRC-0073; realized across all partners since 2021 is “more than US$500m”SRC-0074 — ≈2.5–2.6%, and the ratio reproduces at 1–4% across four further dealsSRC-0084SRC-0086. I The denominator is not fully reconciled — two readings of the same 10-K differ by roughly US$1bn — so the ratio is published as a range, not a point, and the governance rule that rests on it is unaffected: only upfronts count.

F The natural experiment that settles the business model. Schrödinger, FY2025, one company, one auditor: software revenue US$199.5m at 74.4% gross margin; drug discovery US$56.4m at −10.4% gross margin, with discovery revenue growing 107% and widening the gross loss; and the annual report states the company will not initiate further independent clinical trialsSRC-0075. I The sponsor should stand on the software-and-data side of that line.

F And the audited counter-case is the sponsor’s exact business model. Ten listed genomic-diagnostics registrants earned combined FY2025 revenue of US$9,928m at gross margins of 40.6–79.4% — and nine of the ten lost moneySRC-0141. I The category has solved gross margin and not operating leverage. That is the correct discount to apply to every revenue line in the capital plan.

I One reconciliation the audit was right to demand. This report has said both that the model layer commoditised and that physics-based simulation is the most durable moat observed — held by the vendors whose capability Boltz-2 is said to have matched. The reconciliation is that what commoditised is the prediction; what remains durable is validation against physical measurement with regulatory acceptance attached. The corroborating datum is uncomfortable for the optimistic reading: Schrödinger’s count of customers above US$1m annual contract value fell from 29 to 27 while spend per customer roseSRC-0075.

What buyers actually pay — transaction prices, not market sizes

F Published 2025 sector funding totals span US$2.11bn to US$11bn with no shared inclusion criteriaSRC-0091capped, so this report publishes none. It publishes six observed prices instead.

I There is no jurisdiction from which U.S. and Chinese genomic data can lawfully be pooled. That kills any cross-border data structure at concept stage, consuming no legal budget.

The demand signal, and the prohibition attached to it

F Phase 1–3 interventional starts rose 8,416 → 8,985 (+6.8%) from 2020 to 2025, while starts whose eligibility criteria invoke a biomarker, mutation, genotype or gene expression rose 939 → 1,221 (+30.0%)SRC-0138. Split by site country: U.S.-sited such trials were flat (493 → 490, −0.6%); China-sited doubled (241 → 480, +99.2%)SRC-0139.

I The growth market for genomic patient selection is in China, and — under assumption A — the sponsor cannot lawfully move genomic data across that border in either direction. The opportunity and the prohibition are the same fact seen from two sides. Under assumption A falsified, this row inverts.

Four dated items that make existing internal positions stale

• F The USPTO rescinded its AI-inventorship guidance in its entirety on 2025-11-28SRC-0114.
• F EU AI Act Art. 6(1) high-risk in-vitro-diagnostic obligations begin 2027-08-02SRC-0124.
• F European Health Data Space secondary use begins 2029-03-26SRC-0125 — the date before which EU positioning should not be converted to capital.
• F The laboratory-developed-test rule was vacated and 21 CFR 809.3(c) now reads “[Reserved]”SRC-0114SRC-0115. I This matters more than it did: the LDT route is the one the completed precedent actually used to reach Medicare payment.

F The completed precedent is inside the low band, and authorisation did not move it. PLA code 0376U — the DEN240068 asset — is paid US$706.25, effective 2024-01-01, and US$706.25 still on the CY2026 Q3 scheduleSRC-0157SRC-0158. Its comparator in the same disease, 0047U, the 17-gene mRNA prostate classifier, is paid US$3,873.00 — 5.5× moreSRC-0157.

Two strategic consequences follow, and they run in opposite directions.

1. Against the recommended route as previously stated. An FDA authorisation is not the thing that unlocks payment. Its value must be argued on durability, PCCP updatability, follow-on velocity and the ADLT limb.
2. For the fallback the previous report treated as a downgrade. “Re-scope to an LDT launch” was framed as what happens if the predicate fails to generalise. The precedent says the LDT route reached Medicare payment first, on the same asset, from the same companySRC-0158.

The ADLT limb, measured — and it is the whole economic case

F CMS publishes the complete list of tests granted Advanced Diagnostic Laboratory Test statusSRC-0159, on a page last modified 2026-08-03SRC-0160. Eighteen tests have ever been granted ADLT status, from 2018-05-18 to 2025-03-10, and fifteen carry a published new-ADLT initial-period payment amountSRC-0159SRC-0160.

F And the rate persists. Comparing each test’s self-set initial-period list charge against its CY2026 CLFS rateSRC-0157SRC-0159: eleven of fifteen are paid at exactly the same amount, years later. One rose 111% (0108U, TissueCypher, US$2,350 → US$4,950), one rose 2.6%, one fell 1.6%, one fell 25%.

I 42 CFR 414.522(a) says the sponsor sets its own rate for three calendar quarters and is then reset to the PAMA weighted median with a 130% clawbackSRC-0149SRC-0150. In the observed record, the rate a laboratory set for itself has become the rate it keeps. That is the strongest single argument for Track 2 anywhere in this work, and this package published the rule for months without ever testing it against the outcome.

Three limits, stated with equal force

1. The grant is rare and may currently be unavailable. Eighteen grants in just under seven years, and none published in the seventeen months since 2025-03-10SRC-0159SRC-0160. I An ADLT grant is not a planning assumption.
2. No listed ADLT is an AI or machine-learning device. The closest analogue is 0108U, TissueCypher — whole-slide digital imaging with computer-assisted quantitative immunolabeling — granted 2022-03-24 at US$2,350 and since repriced to US$4,950. That is a real precedent for an image-based computational score obtaining ADLT status, and it is the concrete thing to ask CMS and FDA about.
3. The one FDA-authorised AI diagnostic is not on the list. It took the CDLT route and earns US$706.25.

F B2’s economics score falls 4 → 3 on the pricing measurement above, moving its composite 30 → 29. Both rows previously omitted from the published table are restored.

I The composite is rewarding optionality. The top four families score 4–5 on capital, time and reversibility and only 2–3 on defensibility; the two families carrying the highest defensibility and evidence scores rank ninth and seventh. The high-composite families are good first moves and poor destinations; the low-composite, high-defensibility families are good destinations and poor first moves. That asymmetry is the entire argument for the barbell. A ranked list read as a shopping list would fund the wrong things in the right order.

I The single highest-leverage structuring decision is not a market event — it is a contract. Contracting outcome-linked cohort access rather than building it moves A2 from 26 to 32 by removing the time and reversibility penalties without touching its defensibility score of 5. It also survives a restrictive consent finding, which building does not.

Sensitivity 1 — the predicate-generalisation risk, finally scored

Downside assumption S-D5 — that a sequence-based AI score does not fit the DEN240068 predicate — was this package’s largest named structural risk and had no representation in the scoring. Adding it:

I Under S-D5 the destination scores below the engine. That is the information a board needs before it funds Tranche 2, and the previous report computed its portfolio as though S-D5 had resolved favourably.

Sensitivity 2 — somatic versus germline (assumption B)

E Under a germline / hereditary asset base — blood-derived DNA, no tissue, no treatment linkage — A2 falls to ~20, B2 to ~19 and B3 to ~14, because the DEN240068 template requires tissue with treatment and outcome linkage that the business does not holdSRC-0136. I Under assumption B falsified, Track 2 is not merely harder — it requires acquiring a modality, which is a different decision from the one this report analyses. Track 1 is substantially unaffected.

Genetically anchored target work, decomposed

R The previous report scored one family where there are two businesses, and scored the reader into the wrong conclusion.

What FDA required, and what it cost

F The DEN240068 decision summary: training on whole-slide images from completed multi-centre prospective randomised trials using actual patient events as ground truth — 1,133 distant-metastasis and 931 prostate-cancer-specific-mortality events; a pivotal clinical validation in 886 patients across three U.S. sites; and an analytical reproducibility study of 52 samples at three sites over five daysSRC-0136. No new prospective trial was run.

F The follow-on breast indication cleared in 136 days against 248 for the De NovoSRC-0151, and FDA reviewed and cleared a Predetermined Change Control Plan, so conforming model updates require no new premarket notificationSRC-0137.

I This remains the only asset class identified anywhere in this package that improves with accumulated data while staying inside its authorisation, and the only moat that does not erode with open weights. The route ratio on the entry ticket is 24.1× — De Novo US$191,020 against a prescription-drug application fee of US$4,600,753SRC-0147SRC-0148.

R Coarse bands are better calibrated and give a board less to govern with, so the bands move only on observations stated in advance.

The four indicators to read first

I Four things are deliberately not indicators, each for a stated reason: sector venture-funding totals (~5× spreadSRC-0091capped), counts of “AI-discovered” clinical assets (~15% spread, no denominatorSRC-0043capped), platform share prices (they move with capital availability and get read as science), and announced deal headline values (realized runs 1–4%).

I An instrument read against a stale threshold is not an instrument. L11’s published base-case reading was superseded by this package’s own measurement and was not updated for months. That is the defect the audit named and it is fixed above.

How the independence of the two axes is defended

I Zero AI drug approvals ran alongside a completed diagnostics path — but that path ran evidence → LDT reimbursement → authorisation → no rate change, not the order previously stated. The two axes still moved independently for six years; the diagnostics leg is a weaker demonstration of a working pathway than this report previously claimed. The strongest objection is conceded in full: the decoupling is confounded with a pandemic, a rate cycle and a biotech capital contraction, so independence is held as a working structure with a stated falsifier, not as an established property.

I Three things this frame sees that a one-dimensional frame cannot. Cell II is the trap, which is why governance rule 2 gates therapeutic exposure on Axis 2 evidence. Cell V is better for the sponsor than “the sector collapsed” sounds. And the sponsor’s real downside — CMS rate-setting and predicate generalisation — is entirely on Axis 2 and has nothing to do with Phase 3 results. Both Axis 2 readings landed adverse.

Three consequences, all carried into the horizon table. The envelope is reduced to US$64.5–174m — published as the computed sum of the tranches, unrounded. Tranches 2 and 3 require balance-sheet capital or a funding partner. And whether to own the diagnostic asset or trade it to a pharmaceutical partner for funding becomes a board decision with numbers attached, not an analytical preference.

F For a test that qualifies as an ADLT, 42 CFR 414.522(a) sets payment during the new-ADLT initial period “equal to its actual list charge” — a period 42 CFR 414.502 defines as “a period of 3 calendar quarters”SRC-0149SRC-0150. F The ADLT criteria are drafted around the object this sponsor would build: a test qualifies if it analyses multiple biomarkers of DNA, RNA or proteins “when combined with an empirically derived algorithm” to predict disease development or therapy response and provides new information — or, alternatively, if it is FDA-cleared or approvedSRC-0149. I An FDA-authorised AI-derived multi-analyte score qualifies under both limbs at once, and the application form is publicSRC-0154capped.

I Two structural changes. G2 no longer releases capital — it authorises a programme, and G2a releases the money once the ADLT question has an answer. The previous ladder released US$35–90m against an ADLT assessment folded in among five conditions; the 4.2× spread sits entirely on that one condition. And G3’s old pass condition — “realized rate inside US$3,782–8,330” — was set against a survivor band that no route reaches by default. Until an ADLT determination exists, G3 is not releasable.

R One of them has already fired in draft: the reimbursement criterion, on this package’s own measurement.

R The consolidated register holds 23 items requiring management interview, paid data or legal review; eighteen close inside the 90-day window for Tranche 0’s US$1.5–4.0m — about 2% of the five-year plan. I That ratio is the strongest argument for the sequencing.

I Item 6 is the clearest instance in the package of buying an answer being cheaper than reasoning to one, and it is open whether MDUFA VI introduces a pre-submission fee from FY2028 — a dated reason to make the purchase now.

Six things a counterparty will claim, and what the record supports

Ten stated limitations of this report

1. Public sources only. No management interviews, no paid databases, no data-room access. Deal terms, pre-IND attrition and private financings are under-observed by construction.
2. Publisher access controls skewed the mix. I This track is better sourced on what companies did than on what auditors and independent analysts concluded — a bias running toward company narrative and against this report’s own conclusions, so the net effect is conservative. Phase 04 is the exception that proves the cost: one paywalled full text, once retrieved, changed a decisive claimSRC-0003.
3. SEC-registrant skew. Well sourced on the listed sector, poorly sourced on the best-capitalised private portion — Isomorphic, insitro, DoveTree.
4. U.S.-federal price skew. Every price anchored in the opportunity and capital sections is a U.S. federal price. The budget is inapplicable rather than approximate outside the U.S. This is assumption A.
5. China visibility asymmetry, and the specific gap that matters most: this package cannot price the China diagnostics opportunity it identifies as the growth market for genomic patient selectionSRC-0139, and will not fill it with a consultancy figure.
6. Attribution opacity. Whether a model contributed causally to a given asset is rarely externally verifiable. This report can usually establish what a company claims, and only sometimes what AI did.
7. The revenue build is illustrative and one of its four lines has no published anchor. It is falsifiable, which is its purpose; it is not a forecast.
8. The ADLT findings measure published rates, not claim volume or realized revenueSRC-0159, and cannot distinguish a CMS pause from a publication lag.
9. It is corporate strategic research, not personalised securities advice. Listed companies appear as evidence about industry structure. Nothing here is a view on any security.
10. It is a snapshot. The sentences that decay fastest: trial statuses, fee schedules (reset annually and quarterly), the ADLT list, the L7 device count.

Independence and provenance

F No material under the Codex track was read, searched or referenced at any point — during the original research, during the revision, or in producing the final report. That is a stronger statement than the approved brief requires, and it was made deliberately: where the cross-review named a fact, this track retrieved it from its own primary source rather than adopting it. Four such retrievals were madeSRC-0157SRC-0159SRC-0161SRC-0003. I A track cannot retroactively hash-lock its own history; per-phase completion commit hashes are recommended for future programmes and are outside this track’s control.

Claim taxonomy

Two discipline rules, both frequently violated in this sector’s literature

• AI-designed versus AI-adjacent. A compound whose structure or target was generated or selected by a model is categorically different from a programme that used a model somewhere in its development. This track will not aggregate the two.
• Headline value versus realized economics. Partnership totals, biobucks and “up to” milestone figures are CR at best and are never reported as revenue, as committed capital, or as evidence of validation.

Source hierarchy

Retrieval protocol

A source is not admitted on the strength of a search snippet. Chinese-language sources: where a Chinese regulation, standards document or filing is material, the original text is the source of record and the English secondary analysis is a Tier 3 triangulation, never a substitute.

Three rules were added by the audit and are now mechanically enforced by a test harness that runs 74 checks with zero failures: no decisive claim may rest solely on a verified-metadata row; a correction must propagate to every artifact that carries the claim, not only the one where it was found; and a capped row must be flagged inline at the point of use, not only in the ledger.

Research as-of date

2026-08-04. All present-tense statements are asserted as of that date and nothing later. Claims whose truth decays quickly are marked status-sensitive in the underlying report so a reader consuming this package later knows exactly which sentences to re-verify: trial statuses, fee schedules (reset annually and quarterly), the ADLT list, and the device count under 21 CFR 864.3755.