BioinvestGPT ApS · Research use only
Perspective · Mathematics, not data ·
Mathematics is Enough
In a note, BioinvestGPT set out how BVCT predicts Phase-3 hazard ratios from mathematics derived from the genome, with no patient data. On the frozen cohort to 24 Sep 2026, 674 of 705 scored predictions were correct (95.6%); 82.1% of the cohort is first-in-class, where statistical AI has no historical exemplars to learn from.
BioinvestGPT ApS is not affiliated with, endorsed by or partnered with the other AI companies named on this page.
A TRL-9, deployed clinical superintelligence.
The protagonist of BVCT is not a model, not a transformer, and not a billion-row dataset. It is mathematics, in the strict sense: an analytical apparatus, derived from first principles, with closed-form behaviour and falsifiable predictions.
BioinvestGPT has built — and is operating today — the world's first TRL-9, deployed, patient-data-free clinical superintelligence. The system, BVCT (BeatSoC Virtual Clinical Trials), predicts Phase-3 hazard ratios for live pharma programmes — a substantial majority of them in oncology — with no patient data of any kind. It does, in timestamp-audited evidence, what, in our view, Isomorphic Labs has aspired to since its November 2021 launch.
The substrate is the genome and the laws governing how a body unfolds from it — not a corpus, not a benchmark, not a billion-row dataset.
Why Mathematics, Not Data, Wins Here.
Clinical drug discovery — especially in oncology — is the domain in which statistical AI is structurally, not contingently, blind. Roughly four in five cancer drug projects in clinical development were potentially first-in-class (Analysis Group/PhRMA, 2017): for a first-in-class drug, by definition, no historical exemplars exist for any model to learn from. AlphaFold 3 solves a structural-biology problem; BVCT solves the very different problem of clinical effect size against the standard of care, where the signal is genuinely novel.
BVCT works in this region because its substrate is mathematical, not empirical. We de-compress causal disease and drug biology directly from the genome, instantiating two million human-equivalent virtual patients, and simulate any molecule, any modality, any indication head-to-head against any control. Inputs are public trial-design parameters; outputs are PFS/OS HR vs. SoC at 0.1 precision for specific patient subgroups, with a mechanistic causal rationale per prediction.
The mathematical apparatus has no counterpart in the published statistical-learning literature.
The Proof.
Ex-ante, prospective, cryptographically timestamped before readout, timestamp-audited.
- 705 PREDICTIONS — Sent to pharmaceutical clients across sponsors, drug assets, therapeutic areas (oncology dominant) and modalities; frozen cohort as of 24 Sep 2026.
- 95.6% ACCURACY — 674 of 705 correct; two-sided p < 0.0001. F1 = 93.8%, DOR 953, 99.1% NPV, 89.7% PPV.
- 82.1% FIRST-IN-CLASS — Share of the ledger: the system operates exactly where statistical AI is structurally blind, because no historical data exists.
- EU TOP-20 PHARMA — Auditing live predictions in an exclusive data room since April 2024; independent due diligence in April 2026 returned 9 / 10 correct on the partner's own oncology trials, median lead time ~9 months before readout.
- 2−181 NULL EXCLUDED — The probability that the track record is a lucky guess is below 2−181 — more than 9.0 × 1015 (253) times stronger than the 128-bit cryptographic standard (2−128). The null is, in a precise mathematical sense, excluded.
In our view, Isomorphic Labs has been chasing this since 2021 with statistical AI. We did it with mathematics. The interesting question is why mathematics was, all along, the only path that could close.
Live Evidence.
- Public ledger of ex-ante predictions: data.bioinvestgpt.com — updated regularly, re-validated against every new global trial readout.
- Public alpha (run a simulation yourself): bvct.bioinvestgpt.com