The landscape of AI drug-discovery platforms: the design is proven, but is the clinic?

Evidence-first notes on bioscience and deep tech, at the edge of the lab and the market. Information only — not investment advice.

The 30-second version

• What. AI drug-discovery platforms (Recursion, Insilico, Isomorphic, Schrödinger, Relay, Nimbus and others) have partly proven that they can design and optimize molecules — many candidates have entered the clinic and cleared Phase 1. But whether those molecules cure disease (Phase 2 efficacy) is still indistinguishable from traditional drugs, and no company has closed the evidence loop showing AI changed a hard clinical outcome. The number to hold onto: the count of FDA-approved drugs that originated from AI is zero.
• So what. The data draw the gap precisely. AI-derived candidates show a visibly higher Phase 1 success rate (roughly 80–90% vs. a historical 40–65% in a secondary tally), but by Phase 2 the rate falls to ~40%, statistically indistinguishable from the traditional ~37% (Jayatunga et al., Drug Discovery Today 2024). Phase 1 mostly measures chemistry and design (druglikeness, safety, PK); Phase 2 measures whether the target and biology actually move the disease. AI/physics engines win where they are strong (design) and become ordinary where they are not (biology).
• Now what. Read “reached the clinic” as separate from “worked in the clinic.” Watch the attribution problem: rentosertib’s Phase 2a (Insilico, Nature Medicine 2025) is widely called “positive,” but its primary endpoint was safety, not efficacy — the FVC gain was secondary/exploratory; zasocitinib’s strong Phase 3 sits on a pre-validated target (TYK2), so the computational contribution was chemistry selectivity, not target discovery. Verification status is PARTIAL (15 confirmed / 0 refuted / 4 unverified); case-by-case adversarial checks are finalized in the sub-parts.

[demo-gap note] A design win is not a clinical win. What AI platforms have demonstrated is molecular design and preclinical/Phase 1 druglikeness; the hard clinical outcome (Phase 2/3 efficacy on disease) is the open question. The clinical version of the demo-gap: proven design and preclinical performance are not proven efficacy on a hard clinical endpoint. Company self-framing (“first clinical validation of our OS”), trial-registry facts, and independent efficacy readouts must be kept apart.

The five-minute read

The one question: does AI design translate into clinical cure?

AI drug-discovery platforms get compressed into headlines about “designing medicines in a fraction of the time.” One layer in, the useful question is narrower: on a real disease, does an AI-derived molecule beat placebo or standard of care on a hard endpoint? This is the successor question to the firm’s biology-foundation-models series, which concluded that models are hypothesis generators, not prediction oracles, and that bio-FM-derived marketed drugs number zero. Here we look not at the models but at the candidate molecules’ clinical report cards.

Design is proven; efficacy is ordinary

Phase 1 — the design win. AI-derived candidates clear Phase 1 at a markedly higher rate than the historical baseline (an aggregate 80–90% vs. 40–65%). Phase 1 largely gates on druglikeness, safety and PK — the chemistry/design layer where generative and physics engines are genuinely strong. This much is a real, if small-sample, signal.

Phase 2 — the efficacy plateau. By Phase 2 the AI-derived rate lands at ~40%, statistically indistinguishable from the traditional ~37% (Jayatunga et al., 2024). Phase 2 gates on whether the target and biology actually move the disease — the layer where AI has not yet shown an edge. The sample is small and early (survivorship-bias exposed), so the precise percentages are secondary/unverified, but the pattern — win at design, ordinary at biology — is the series’ through-line.

The frontier case, read carefully. rentosertib (Insilico, TNIK inhibitor, IPF) is the leading end-to-end AI clinical proof-of-concept — target found by PandaOmics, molecule designed by Chemistry42. In GENESIS-IPF (Phase 2a, n=71, 12 weeks, placebo-controlled), the 60mg arm showed FVC +98.4 mL vs. −20.3 mL for placebo (Nature Medicine 2025). But the primary endpoint was safety (TEAE rate), the FVC change was secondary/exploratory, and hepatotoxicity/diarrhea drove early discontinuations. It is an efficacy signal, not efficacy proof.

[diagram: AI drug platforms — proven at design, ordinary at efficacy]

  Stage / question       AI-derived        vs traditional     What it gates on
  ────────────────────────────────────────────────────────────────────────────
  Phase 1 (design)       80-90%  ███████    40-65%  ████       druglikeness,
  "does the molecule                                           safety, PK
   behave?"                                                    = CHEMISTRY

  Phase 2 (efficacy)     ~40%    ████       ~37%    ████       target/biology
  "does it move the                         (indistinguishable) moves disease
   disease?"                                                   = BIOLOGY

  Approved AI-origin drugs:  0   (none to date)
  ────────────────────────────────────────────────────────────────────────────
  through-line:  a design win  !=  a clinical (efficacy) win

Deep dive

1. Background — porting the “real bottleneck” lens to the clinic

This Part 0 transplants the firm’s “the real bottleneck next to the headline” lens onto the clinical layer of AI drug discovery. It directly takes up the question the biology-foundation-models series explicitly handed off — does AI translate into actual therapy? — and looks at the candidates’ clinical results rather than the models. Verification status is PARTIAL: clinical events, discontinuations and readouts were cross-checked against primary press, peer review and SEC filings, but company forward-looking items (Isomorphic first-in-human timing, Recursion milestone valuations) and precise success-rate figures are attributed and marked unverified, with case depth deferred to the sub-parts.

2. What this landscape newly establishes

Core answer: AI platforms have proven they can design molecules and put them into the clinic; they have not proven those molecules change hard clinical outcomes, and no case cleanly separates the AI contribution from ordinary drug success.

• Positive signals (verified) [primary/peer-review, high]: rentosertib Phase 2a FVC signal (Nat Med 2025, PMID 40461817); REC-4881 (Recursion, MEK1/2, FAP) rapid, durable polyp-burden reduction in TUPELO Phase 1b/2 (company-framed as “first clinical validation of Recursion OS,” readout webinar 2025-12-08); SGR-1505 (Schrödinger, MALT1) Phase 1 ORR 22% in 45 B-cell malignancy patients (EHA 2025); RLY-4008 lirafugratinib (Relay, FGFR2) ORR 35% in an FGFR2-fusion subgroup.
• Failures / discontinuations (verified) [primary/high]: Recursion halted four programs in May 2025, including REC-994 (CCM, SYCAMORE Phase 2) and REC-2282 (NF2, Phase 2); Exscientia discontinued EXS-21546 (A2A, 2023) and DSP-1181 (OCD) — the original “AI-designed clinical candidate” banners; Schrödinger discontinued SGR-2921 (CDC7, AML) amid a reported safety signal; BenevolentAI’s BEN-2293 (Pan-Trk, atopic dermatitis) missed its Phase 2a primary endpoints (2023-04), a symbol of the hype correction.
• The best-in-class case is a target problem [press, high]: zasocitinib (Nimbus/Schrödinger, TAK-279, TYK2) advanced to Phase 3 psoriasis under Takeda with strong late-stage efficacy — but TYK2 is a pre-validated target (deucravacitinib precedent), so the computational contribution was chemistry selectivity, not target discovery. It is the attribution problem in a single example.
• The best-capitalized platform has zero clinical assets [forward-looking, unverified]: Isomorphic Labs (Alphabet-affiliated, AF3-based) has no clinical asset; first-in-human is targeted for late 2026 (slipped from late 2025). Some secondary outlets attribute an “ISM8969 FDA clearance (2026-01)” to Isomorphic, but the ISM prefix conventionally belongs to Insilico — attribution is unconfirmed and is not stated as fact here.

3. Strengths and limits of the methodology

Strengths: clinical facts are attributed to primary press, peer review and SEC filings; platforms are grouped by approach (phenotypic vs. structure/physics-based vs. generative); and the success-rate claim is deliberately split from the efficacy claim. Limits: the Phase 1 > traditional figure rests on a small, early sample exposed to survivorship bias (discontinued programs drop out of the tally), and the “AI-derived” definition varies by company (target discovery only? molecular design only? end-to-end?), destabilizing the denominator of any comparison. rentosertib’s “positive Phase 2a” headline exceeds its evidence grade because the primary endpoint was safety. The upshot: on currently public data, no case statistically separates “AI contribution changed a hard clinical outcome” from ordinary drug success.

4. Neighbouring domains

The cross-domain hook is methodological: the same “headline metric is not usefulness” structure recurs from the computing and biology-foundation-models series, and the survivorship-bias problem here mirrors benchmark-contamination debates in general ML evaluation. Whether these design engines translate to CKM (cardio-renal-metabolic) indications is deferred to the sub-parts; the through-line recorded now is only the structural similarity, not a claimed clinical payoff.

5. Commercialization and market context (actors, approaches)

Clinical facts are attributed to primary press, peer review and SEC filings. No ranking of the companies is implied.

Three approach axes: phenotypic (Recursion — reverse-engineering from imaging phenotypes), structure/physics-based (Isomorphic, Schrödinger, Nimbus, Relay — predicting and optimizing binding from structure and physics), and generative (Insilico, Exscientia — target discovery plus molecule generation). That the best-capitalized platform (Isomorphic) has zero clinical assets echoes the “capital-does-not-track-validation” pattern from the biology-foundation-models series, now recurring at the clinical layer. No single-winner conclusion is asserted; company figures are attributed.

6. The skeptic’s bottom line

• Verified-clean: the clinical events themselves — rentosertib’s Nat Med data, the REC-994/REC-2282 discontinuations, the EXS-21546 discontinuation, the BEN-2293 failure, SGR-1505’s 22% ORR, zasocitinib’s Phase 3 — are cross-checked and factually settled.
• Proceed-with-caveats: the success-rate superiority claim. The Phase 1 edge is real, but the small-sample, survivorship-bias, secondary-tally conditions are mandatory, and it must not be extended to Phase 2 efficacy. Any listed-company pipeline description carries a mandatory condition: no share-price implication; the investment layer is kept separate.
• Hold: (1) any complete “AI has clinically validated a drug” narrative — approvals are zero and Phase 2 is indistinguishable; (2) headlining rentosertib’s “positive Phase 2a” as efficacy — the primary endpoint was safety; (3) unconfirmed attributions such as Isomorphic “ISM8969.”
• Escalated: success-rate/attribution claims to Skeptic; listed-company neutral framing to Principal. Fan-out tally: 15 confirmed / 0 refuted / 4 unverified.

7. What to watch (falsifiable predictions)

1. Over the next 24 months, the number of AI-derived candidates that pass a Phase 3 hard endpoint versus placebo/comparator and reach an approval filing will be 0–1. Falsifiable by zasocitinib (though on a pre-validated target) or rentosertib Phase 2b disclosures.
2. As the sample grows, the AI-derived Phase 2 success rate will converge toward the traditional ~37% while the Phase 1 edge persists (design win, efficacy ordinary). Falsifiable by larger downstream tallies.
3. Isomorphic Labs will not report a public Phase 1 efficacy/PoC readout from its own pipeline within 18 months (design and capital lead, clinical translation lags). Falsifiable by company disclosure.

References

• 2025. “A generative AI-discovered TNIK inhibitor (rentosertib, ISM001-055) for idiopathic pulmonary fibrosis: a Phase 2a trial.” Nature Medicine (PMID 40461817). https://www.nature.com/articles/s41591-025-03743-2
• Jayatunga, Madura K. P., et al. 2024. “How successful are AI-discovered drugs in clinical trials?” Drug Discovery Today. https://www.sciencedirect.com/science/article/pii/S135964462400134X
• Jayatunga, Madura K. P., et al. 2022. “AI in small-molecule drug discovery: a coming wave?” Nature Reviews Drug Discovery. https://www.nature.com/articles/d41573-022-00025-1
• Recursion. 2025. “Recursion halts four pipeline programs.” GEN / company IR. https://www.genengnews.com/topics/drug-discovery/recursion-halts-four-pipeline-programs/
• Recursion IR. 2025. “REC-994 SYCAMORE Phase 2 (ISC, 2025-02); REC-4881 TUPELO readout webinar (2025-12-08).” https://ir.recursion.com/
• Schrödinger IR. 2025. “SGR-1505 Phase 1 (EHA 2025, ORR 22%); SGR-2921 discontinued.” https://ir.schrodinger.com/
• Relay Therapeutics IR. “RLY-4008 lirafugratinib (FGFR2) data; RLY-2608 ReDiscover-2 Phase 3.” https://ir.relaytx.com/
• Takeda. 2022. “Acquisition of Nimbus NDI-034858 (TAK-279 / zasocitinib, TYK2), 2022-12.” https://www.takeda.com/newsroom/
• Exscientia pipeline reductions; EXS-21546 discontinued (2023). Fierce Biotech. https://www.fiercebiotech.com/

Source knowledge asset: knowledge-base/deep-dives/ai-drug-clinical-readout/part0-landscape.md (generated 2026-07-12, verification PARTIAL — 15 confirmed / 0 refuted / 4 unverified). This draft inherits the figures and source attributions of the original part and creates no new figures, numbers or sources.

Disclosure

This post is for information only and is not investment advice. The author holds no position in, and no financial interest in, the companies mentioned (Recursion RXRX, Schrödinger SDGR, Relay RLAY, BenevolentAI, Alphabet, and others).

COI note: This document describes listed AI drug-discovery companies (Recursion RXRX, Schrödinger SDGR, Relay RLAY, BenevolentAI) and private actors (Insilico Medicine, Isomorphic Labs, Nimbus Therapeutics) in a factual, neutral clinical/platform context, including negative facts such as program discontinuations, hepatotoxicity and safety signals. There is no buy/sell implication. Quantitative claims (Phase 1/Phase 2 success rates, ORR figures, rentosertib’s FVC change, Isomorphic’s clinical-entry timing) are attributed as “vendor/author/registry/preprint claims” and are not independently verified beyond the primary sources cited.