Structure-based and generative drug design in the clinic: elegant molecules, but whose efficacy?

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. Structure-based and generative design engines (Insilico’s Chemistry42, Isomorphic’s IsoDDE, Schrödinger’s FEP+, Relay’s motion-based platform) have demonstrably done one thing: designed and optimized molecules well enough to enter the clinic, some reaching late-stage trials. But the firm’s narrow question is whether the engines predicted efficacy — and there the picture is more sober. The three strongest clinical signals in this camp (zasocitinib on TYK2, RLY-4008 on FGFR2, RLY-2608 on PI3Kalpha) all sit on already-validated (precedented) targets, so the engines’ contribution is selectivity and chemistry, not efficacy.
• So what. “AI designed the drug” collapses two different claims. On precedented targets, the design engine explains the molecule’s selectivity and safety margin, while the direction of efficacy is set by the target biology — and the two cannot be separated from the clinical data. The only case where AI opened the target itself, all the way through, is Insilico’s rentosertib (TNIK inhibitor), which entered Phase III on 2026-07-07 — but its Phase 2a primary endpoint was safety, not efficacy.
• Now what. Watch the capital-versus-evidence gap. Isomorphic Labs holds the most capital in the camp (~$3B in partnership biobucks) yet has zero clinical assets and no named clinical candidate. Two corrections carried through: the rentosertib placebo arm was FVC −20.3 mL (not −62.3 mL, an earlier secondary citation), and the “ISM8969” attribution to Isomorphic is unconfirmed — the “ISM” prefix belongs to Insilico, so it is not stated as fact. Verification status is VERIFIED (15 confirmed / 1 refuted / 4 unverified).

[demo-gap note] Reaching Phase 1/2/3 is a real, verified fact for these molecules. But a late-stage efficacy signal on a precedented target is not proof that the design engine predicted efficacy — the target biology is the confound. Rentosertib is the one asset where AI reached the target itself, and its efficacy is not yet demonstrated: the Phase 2a primary endpoint was safety/tolerability, and the hard-endpoint efficacy test (52-week annual FVC decline) only began in Phase III. Isomorphic’s efficacy-prediction power can be neither confirmed nor refuted today, because there is no clinical data at all.

The five-minute read

The one question: did the design engine predict efficacy, or just build a better molecule?

This is Part 2 of the AI-drug clinical-readout series. Part 0 mapped the landscape; here the structure-based and generative camp is examined against that map’s Q2 — did the design engine move beyond “design” to “efficacy”? The through-line from Part 0 (“strong in Phase 1, ordinary in Phase 2”) is especially sharp in this camp, because the engines are strongest exactly where chemistry lives — selectivity, free energy, residence time — which is precisely what Phase 1 screens for (druggability, safety, PK). The late-stage efficacy signals, by contrast, all sit on targets that were already clinically validated, so “the engine predicted efficacy” and “the validated target produced efficacy” cannot be told apart in the data.

Four engines, one attribution problem

Insilico Medicine — the only end-to-end case. PandaOmics (target discovery) plus Chemistry42 (generative chemistry) derived both target and molecule by AI. Rentosertib (ISM001-055, TNIK inhibitor, IPF) reported in the GENESIS-IPF Phase 2a (n=71, 22 sites in China, 12 weeks, placebo-controlled; Nature Medicine 2025, PMID 40461817): 60 mg QD arm FVC +98.4 mL vs. placebo −20.3 mL at 12 weeks. The primary endpoint, though, was safety/tolerability, with FVC secondary/exploratory. Phase III began 2026-07-07 (320 patients, 47 sites in China), with the primary endpoint now the efficacy measure — 52-week annual FVC decline. Efficacy validation, in other words, is only starting.

Isomorphic Labs — most capital, zero clinic. IsoDDE, an AlphaFold3-based design engine (disclosed 2026-02), carries the thickest capital and technology narrative in the camp — a $600M external round (Thrive Capital-led, GV, Alphabet, 2025-03-31) and Lilly/Novartis partnerships totaling ~$3B in biobucks. Yet it has zero clinical assets and no named clinical candidate; 17 active programs are at the preclinical-nomination stage, with a first AI-designed oncology asset targeting first-in-human by end of 2026 (forward-looking, moved from an earlier end-2025 target).

Schrödinger and Relay — physics- and motion-based, on precedented targets. Schrödinger’s free-energy perturbation (FEP+) engine co-invented zasocitinib (TAK-279, TYK2) via Nimbus, later transferred to Takeda; the LATITUDE Phase 3 met its primary endpoints against placebo and apremilast. Relay’s motion-based platform produced RLY-4008 (lirafugratinib, FGFR2, ORR 35%/40%) and RLY-2608 (zovegalisib, PI3Kalpha, Phase 3). All three late-stage signals sit on precedented targets (TYK2, FGFR2, PI3Kalpha), so the engines’ contribution is selectivity and safety margin — chemistry — not the direction of efficacy.

[diagram: did the design engine predict efficacy? — by company]

  Company        Engine            AI reached target?   Efficacy attributable
  ─────────────────────────────────────────────────────────────────────────
  Insilico       Chemistry42 +     YES (TNIK)  █████░░   PENDING: Ph2a primary
                 PandaOmics                             = safety; Ph3 (2026-07)
  Isomorphic     IsoDDE (AF3)      n/a — no clinic ░░░░  UNDECIDABLE: 0 clinical
                                                        assets, no data
  Schrödinger    FEP+ (physics)    NO (TYK2 precedented) chemistry/selectivity;
                 zasocitinib       ██░░░░░░               efficacy = target
  Relay (RLAY)   motion-based      NO (FGFR2/PI3Ka       chemistry/selectivity;
                 RLY-4008/-2608    precedented) ██░░░░    efficacy = target
  ─────────────────────────────────────────────────────────────────────────
  through-line:  design elegance  !=  clinical efficacy prediction

Deep dive

1. Background — porting the “design vs. efficacy” lens to structure/generative engines

Part 0’s cross-cutting proposition — a Phase 1 edge but ordinary Phase 2 — lands hardest in the structure-based and generative camp. The design engines are strongest in chemistry (selectivity, free energy, residence time), which overlaps precisely with the druggability, safety and PK that Phase 1 screens. This Part tests, at the camp level, whether that strength extends to predicting clinical efficacy. Verification status is VERIFIED, with each candidate’s phase and figures cross-checked against primary press, peer review, SEC filings and clinical registries.

2. What this Part newly establishes

Core answer: the engines demonstrably design and optimize molecules into the clinic, but late-stage efficacy is attributable to the target, not the engine — with a single end-to-end exception whose efficacy is still unproven.

• Insilico rentosertib — the only end-to-end clinical case [peer-review, cautious]: Nature Medicine 2025 (PMID 40461817) reports 60 mg QD FVC +98.4 mL vs. placebo −20.3 mL at 12 weeks, but the primary endpoint was safety/tolerability; Phase III (2026-07-07, 320 patients, 47 sites) makes 52-week annual FVC decline the primary endpoint. Chemistry42 has additionally nominated 11 preclinical candidates, including ISM3830 (CBLB, immuno-oncology).
• Isomorphic — capital-versus-evidence at its extreme [fact / second-hand]: $600M raise (2025-03-31), Lilly (upfront $45M + milestones >$1.7B) and Novartis (upfront $37.5M + biobucks $1.2B) partnerships totaling ~$3B, yet zero clinical assets. The efficacy-prediction power of IsoDDE is currently undecidable — no public clinical data exists to confirm or refute it.
• Schrödinger zasocitinib — the camp’s strongest signal, and the clearest attribution example [company/press, high]: Nimbus co-invented zasocitinib via FEP+, transferred to Takeda. LATITUDE Phase 3 (NCT06088043, NCT06108544; N=693/1108; 21 countries) in plaque psoriasis met the week-16 sPGA 0/1 and PASI 75 primaries superior to both placebo and apremilast, with all 44 ranked secondary endpoints met (Takeda, 2025-12); sPGA 0/1 71.4%/69.2% vs. placebo ~11–13% and apremilast ~30%. A head-to-head reportedly outperformed deucravacitinib (Sotyktu) per BioWorld. But TYK2 is a precedented target (deucravacitinib precedent), so FEP+’s contribution is selectivity/chemistry, and most of the efficacy is target biology.
• Schrödinger in-house pipeline — activity and a safety halt coexist [press, high]: SGR-1505 (MALT1, Phase 1) showed early B-cell malignancy activity (EHA 2025, ORR ~22%); SGR-3515 (Wee1/Myt1) is in Phase 1 with early solid-tumor data expected in Q4; SGR-2921 (CDC7, AML) was discontinued, having been deemed to have contributed to two patient deaths (FierceBiotech). Clinical assets narrowed to two.
• Relay (RLAY) — motion-based, same attribution limit [company/press, high]: RLY-4008 (lirafugratinib, FGFR2) reported ORR 35% (FGFR2 fusion) and 40% (HR+/HER2- breast), “NDA-ready,” globally licensed to Elevar Therapeutics; RLY-2608 (zovegalisib, PI3Kalpha) is in Phase 3 (ReDiscover-2, + fulvestrant). Both targets are precedented, so the engine’s contribution is selectivity and safety margin, not the direction of efficacy.

3. Strengths and limits of the methodology

Strengths: each efficacy claim is separated into “engine contribution (chemistry)” versus “target biology,” with clinical events (phase, endpoints, ORR, discontinuations) cross-checked against peer review, registries and company disclosures, and precedented-target status foregrounded so that selectivity is not mistaken for efficacy prediction. Limits: much of what is confirmed is “the accuracy of the claimed clinical events,” and the central efficacy-attribution question cannot be resolved with current data — separating engine contribution from precedented-target biology is deferred to Part 4. Isomorphic’s forward-looking figures (first-in-human end-2026, biobucks) are attributed, not verified; the “ISM8969” attribution to Isomorphic is a prefix confusion and is not stated as fact; and the corrected placebo FVC (−20.3 mL, not −62.3 mL) is carried per the primary source.

4. Neighbouring domains

The cross-domain hook is methodological. Physics-based simulation (FEP+, molecular dynamics) contributing to chemical optimization is the drug-discovery instance of the computing series’ pattern — an edge on narrow, verifiable tasks. The capital-versus-evidence direction mismatch (Isomorphic: most capital, zero clinic) repeats the observation from the bio-FM series Part 5 at the clinical layer. And the attribution-separation problem (success on precedented targets) becomes a central confound for Part 4 (are the reported success rates real?), where an “AI-derived” definition that lumps together target discovery, chemical optimization and end-to-end could inflate the apparent success rate.

5. Commercialization and market context (actors, engines)

This is the engine/platform angle; individual trial designs are deferred to Part 3.

Deployment-readiness spectrum: preclinical-only (Isomorphic’s IsoDDE) to Phase 1 in-house (Schrödinger SGR-1505/-3515) to late-stage/partnered efficacy readouts (zasocitinib Ph3, RLY-4008 NDA-ready, RLY-2608 Ph3, rentosertib Ph3). No single-winner conclusion is asserted; biobucks and private valuations are attributed to company or second-hand sources; listed-company (SDGR, RLAY) pipelines are described neutrally with the investment layer kept separate.

6. The skeptic’s bottom line

• verified-clean — the clinical facts. Rentosertib Nature Medicine figures and Phase III start (2026-07-07); zasocitinib LATITUDE Ph3 numbers; SGR-1505 ~22% ORR; SGR-2921 discontinuation after two deaths; RLY-4008 35%/40% ORR; RLY-2608 Phase 3; Isomorphic $600M and partnerships. Cross-checked against primary sources.
• proceed-with-caveats — the engine’s efficacy contribution. Mandatory caveat: success on precedented targets (TYK2, FGFR2, PI3Kalpha) is limited to a chemistry contribution — efficacy attribution must not be extended to the design engine. Listed pipelines (SDGR, RLAY) are described with no stock implication and the investment layer separated; biobucks and private valuations are attributed to company or second-hand sources.
• hold — three narratives. (1) Any completed story that “structure/generative AI predicted and proved efficacy” — the late-stage signals are target-attributable and rentosertib’s primary endpoint was safety. (2) Any positive or negative statement about Isomorphic’s efficacy-prediction power — the data is zero. (3) The “ISM8969” and similar unconfirmed attributions to Isomorphic.
• Corrections logged: (1) rentosertib placebo FVC corrected from an earlier “−62.3 mL” secondary citation to −20.3 mL per Nature Medicine and Insilico; (2) “ISM8969 FDA clearance” is not attributed to Isomorphic — the “ISM” prefix belongs to Insilico — and is not stated as fact.

7. What to watch (falsifiable predictions)

1. Until rentosertib’s Phase III (52-week annual FVC decline) passes a hard endpoint with statistical significance versus placebo, the claim that “AI end-to-end proved efficacy” remains incomplete. Falsifiable by the Phase III readout (success would be the first efficacy proof where AI reached the target itself).
2. The late-stage efficacy of the precedented-target assets (zasocitinib, RLY-4008, RLY-2608) will not be statistically distinguishable in effect size from non-AI precedents on the same targets (deucravacitinib, prior FGFR/PI3Kalpha inhibitors); the differentiation is confined to selectivity and safety margin. Falsifiable by head-to-head trials and meta-analysis.
3. Isomorphic Labs will not report public Phase 1 efficacy or proof-of-concept data for its own IsoDDE assets within the next 18 months (most capital and design capability, slowest clinical translation). Falsifiable by company disclosure.

References

• Insilico Medicine et al. 2025. “AI-discovered TNIK inhibitor rentosertib in idiopathic pulmonary fibrosis (GENESIS-IPF Phase 2a).” Nature Medicine. DOI 10.1038/s41591-025-03743-2 (PMID 40461817). https://www.nature.com/articles/s41591-025-03743-2
• Insilico Medicine. 2026-07-07. “Insilico initiates rentosertib Phase III (320 patients, 47 sites in China; primary = 52-week annual FVC decline).” https://insilico.com/news/xmjsn4l091
• Insilico Medicine. “Chemistry42 nominates ISM3830 (CBLB) preclinical candidate.” https://insilico.com/news/h0ftslh6s1
• Isomorphic Labs. 2025-03-31. “Isomorphic Labs announces $600M external investment round (Thrive Capital, GV, Alphabet).” https://www.isomorphiclabs.com/articles/isomorphic-labs-announces-600m-external-investment-round
• FierceBiotech. “Alphabet’s Isomorphic stacks two new deals with Lilly and Novartis (~$3B combined biobucks).” https://www.fiercebiotech.com/biotech/alphabets-isomorphic-stacks-two-new-deals-lilly-novartis
• Takeda. 2025-12. “Zasocitinib Phase 3 plaque psoriasis data (LATITUDE; NCT06088043, NCT06108544).” https://www.takeda.com/newsroom/newsreleases/2025/takeda-zasocitinib-phase-3-plaque-psoriasis-data-once-daily-pill
• BioWorld. “Zasocitinib outperforms Sotyktu (deucravacitinib) in head-to-head Phase 3.” https://www.bioworld.com/articles/731826
• FierceBiotech. “Schrödinger abandons CDC7 inhibitor SGR-2921 after two patient deaths.” https://www.fiercebiotech.com/biotech/schrodinger-abandons-cdc7-inhibitor
• Relay Therapeutics. “RLY-4008 (lirafugratinib, FGFR2) and RLY-2608 (zovegalisib, PI3Kalpha) pipeline.” https://ir.relaytx.com

Source knowledge asset: knowledge-base/deep-dives/ai-drug-clinical-readout/part2-structure-generative.md (generated 2026-07-12, verification VERIFIED — 15 confirmed / 1 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 (Schrödinger SDGR, Relay Therapeutics RLAY, and others).

COI note: This document describes listed AI drug companies (Schrödinger SDGR, Relay Therapeutics RLAY), private actors (Insilico Medicine, Isomorphic Labs, Nimbus Therapeutics) and sponsors (Takeda, Eli Lilly, Novartis, Elevar) in a factual, neutral clinical/platform context. There is no buy/sell implication. Quantitative claims from company materials (biobucks, private valuations, Isomorphic’s clinical-entry timing) are attributed as “vendor/company/second-hand claims” and are not independently verified.