AI has sped up drug chemistry for years. This asks the harder question: did AI come up with a biological idea nobody had?
25%likely
Well sourcedNext check
Watching
Why it is different
Not “AI helped make a drug”. The test is whether AI proposed a target humans had missed, the only version of the claim that would be new.
Why it matters
It is the difference between AI as a faster tool and AI as a source of new science.
Next check · 31 Aug 2026
Does Insilico’s mid-2026 report say the rentosertib Phase 3 trial is still running, not paused, stopped or on hold?
How we scored it ·
from how similar cases went
If right, how big5/5How overlooked4/5Upside vs downside2/5Source trust4/5Independent sources2/5
The exact wording we score
On or before 31 Dec 2030, rentosertib (INS018055) reports a statistically significant positive result on the pre-specified primary endpoint in its registrational Phase 3. A China-only NMPA registrational readout counts. Resolves YES only on a qualifying positive topline or publication; NO on an endpoint miss, termination, or absence of a qualifying readout by the deadline. No other asset can be substituted after the call.
PROOF
Evidence and provenance
The links below are the evidence recorded when this forecast was scored. Each note states source quality or bias; a citation is not an endorsement.
Recorded sources · 4 citations / 2 independent origins
2026-07-28 | sell-side: no direct note in the evidence set; mainstream press: sector narrative focuses on AI-drug failures and optimisation; priced market: no relevant contract identified; specialist: trial and sponsor coverage distinguish novel-target discovery from optimisation | E4
Revision history5 entries
Probabilities and interpretations are never silently overwritten. This is the append-only record of what changed and why.
The supposed closed set still contained an unnamed future Isomorphic asset. That is an open rescue path, so it is removed along with its checkpoint. The parent is now the atomic rentosertib Phase 3 result; P returns 30→25 on the already-recorded 0.58 × 0.40 single-asset arithmetic.
(3rd pass) — Eligible set closed and an adjudication rule added; P 25→30. The old wording ('or another end-to-end AI-discovered, novel-target drug') was an open disjunct with no adjudicator in a field where the party claiming membership is the party with the incentive — it could resolve YES for reasons unrelated to the insight. The set is now fixed at rentosertib plus any Isomorphic asset reaching registrational Phase 3, with a three-part test judged on evidence published before readout, the load-bearing part being that the target had no clinical-stage program at the time of the AI hypothesis — which is precisely what separates 'AI as hypothesis generator' from 'AI as cheaper chemist', i.e. the edge. Also fixed prospectively rather than retroactively: a China-only NMPA registrational readout counts, stated now so it cannot be re-litigated later.
(2nd pass) — P 40→25. The internal arithmetic did not survive inspection: this file's own leading indicator prices the timeline alone (Phase 3 topline reaffirmed for 2029) at 58%. P=40 against a 58% schedule gate implies P(pivotal succeeds | on schedule) ≈ 69% — for an indication this file itself calls 'a Phase 3 graveyard', from a Phase 2a with n≈18/arm and a 95% CI 'nearly touching zero', with known dose-limiting GI discontinuations. That is not a reference-class number. 0.58 x ~0.40 ≈ 0.23. Also noted: the disjunct 'or another end-to-end AI-discovered drug' is an open set with no adjudicator and the sponsor claiming membership is the interested party — the eligible-asset list should be named at logging time. Reclassified refclass: phase3-attrition (a shared BASE RATE) rather than corrgroup (a shared EVENT): these five drug bets have five sponsors, five indications and no common data release, so discounting them as correlated exposure was wrong.
v2 rescore. P 45→40: Bio-IT World (2026-07-08) published the sponsor's own Phase 3 schedule — last patient Nov 2028, results Dec 2029, NDA Mar 2030 — leaving only ~12 months of slack before the 2030 deadline, tighter than this file assumed. Also noted: the pivotal is China-only (47 sites); the resolution criteria should state whether a China-only NMPA registrational trial counts. indepsources 4→2 — Nature Medicine, the PR and Drug Target Review all trace to Insilico's own dataset; peer review is the only independent scrutiny.
created from fresh (non-power) research sweep. The Phase 2a CI is wide; treat the single-asset version as fragile and watch for peer AI-discovered novel-target pivotals as the hedge.
Method challenge
Adversarial review record
Admitted after revision
Structured internal role review tied to this frozen claim. It is not independent human peer review. Independent specialist review not yet performed
technical-domain specialist
delivery and operations reviewer
policy or standards reviewer
forecasting-method reviewer
Strongest specialist challenge
The Phase 2a signal comes from roughly eighteen participants per arm in idiopathic pulmonary fibrosis, with wide uncertainty and gastrointestinal discontinuations. AI provenance is scientifically interesting but does not improve the molecule's clinical efficacy base rate.
Outside view
The useful reference class is Phase 3 success in difficult fibrotic lung disease, not the broad success rate of all late-stage drugs and not the frequency with which AI platforms nominate candidates. Timing and efficacy jointly support a probability near one quarter.
Causal rival
The apparent FVC improvement can be sampling noise, baseline imbalance, regression to the mean, or ordinary medicinal-chemistry success unrelated to the AI target-generation narrative. Any of those can explain Phase 2 without validating the claimed discovery edge.
Measurement risk
A China-only result can be disclosed through sponsor channels before a registry or paper provides a complete statistical analysis. Adjudication must use the pre-specified Phase 3 primary endpoint and cannot substitute biomarker, secondary, subgroup, or another asset's result.
Residual risk
The Phase 3 protocol and NMPA disclosure path remain less transparent than a US or multinational filing, and the AI-origin qualification still lacks independent replication. The binary forecast is clean, but the interpretation of a YES remains provisional.
COLDTELL