BioinvestGPT ApS · Research use only
Platform · BVCT and pharma AI co-scientists ·
BVCT called No. 1959 (mavacamten, non-obstructive HCM) on : partial SUCCESS (statistically maybe significant yet weak clinical benefit in KCCQ compared with placebo); ODYSSEY-HCM failed , True Negative (TN). No. 2094 (aficamten), called : SUCCESS (statistically significant clinical benefit in KCCQ CSS compared with placebo); ACACIA-HCM succeeded , True Positive (TP).
BioinvestGPT ApS is not affiliated with, endorsed by or partnered with the other AI companies named on this page.
BVCT wraps every existing pharma agentic-AI workflow on both ends: derisks the target before generation; derisks the asset before clinical capital.
The New York Times reports that courts are drowning in AI-generated lawsuits that look authoritative but fail verification. Pharma should act now to prevent AI co-scientists from overwhelming R&D with hypotheses that look plausible but fail to translate into revenue.
The agentic AI co-scientist, and the enterprise AI platforms around it, have arrived inside top-20 pharma and beyond. Google Cloud’s Gemini Enterprise agentic AI platform across MSD, including R&D. Anthropic’s Claude across BMS, including R&D. OpenAI partnering with Novo Nordisk on drug discovery and operations. Isomorphic Labs with Lilly and Novartis. Owkin with Sanofi. Edison Scientific with Incyte. — the list keeps growing.
In our view, each one promises the same thing: hypothesis explosion. Targets, indications, designs, binders, biomarkers, trial-design tweaks — surfaced at machine speed. The bottleneck is the same one it has been for forty years: most preclinical hypotheses don’t translate into superior clinical benefit versus Standard of Care, as the clinical success rate of preclinical hypotheses has sat below 10% for decades (BIO: 9.6% from Phase I to approval for 2006–2015; BIO / QLS / Informa: 7.9% for 2011–2020).
Agentic AI gives every pharma a hypothesis explosion. BVCT is the ex-ante clinical filter that turns explosion into prioritised pipeline — one Phase III hazard ratio at a time.
Every pharma agentic-AI loop, regardless of which co-scientist is generating, follows the same shape. BVCT layers slot into four specific points in that sequence — before generation, after wet-lab validation, before translational spend, and at Phase 1/2 GO/NO-GO. Below: the full loop, with the BVCT layers marked “BVCT —”.
| Step in the Loop | Owner / Role |
|---|---|
| BVCT — TARGET DERISKING | BVCT. Rank candidate targets by simulated Phase-III HR vs. SoC. Deprioritise those unlikely to clear HR < 0.6 before any compute, chemistry, or wet-lab spend. |
| Target | Sponsor selects from the BVCT-derisked shortlist. |
| Constraints | Sponsor + co-scientist specify functional, structural, and developability constraints. |
| Embeddings + generation | Co-scientist generates candidates — Claude, GPT, Gemini, Edison, Isomorphic, Iambic, Recursion, Exscientia, Insilico, Iktos, Cradle, BenevolentAI, Generate, Biohub’s ESMC, or any combination. |
| In silico ranking | Co-scientist ranks candidates on biochemical and structural scores. |
| Structure prediction | AlphaFold, ESMFold2, RFdiffusion, or equivalent. |
| Developability filter | Manufacturability, solubility, immunogenicity, off-target binding screened. |
| Synthesis → wet-lab validation → iteration | Survivors synthesised; binding, function, cellular activity confirmed; loop iterates for next-generation candidates. |
| BVCT — ASSET DERISKING | BVCT. For each wet-lab-validated asset, simulated head-to-head Phase-III HR / OR / RR vs. SoC at 0.1 precision, with mechanistic causal rationale — before any IND-enabling spend. |
| BVCT — INDICATION + TRIAL DESIGN | BVCT. For each surviving asset, rank plausible indications; compare endpoints; surface the patient-stratification + SoC-comparator combination beating SoC. |
| BVCT — PHASE 1/2 GO/NO-GO | BVCT. The world’s largest-scale ex-ante effect-size evidence for the sponsor’s GO/NO-GO committee at the point clinical uncertainty peaks. Final decision remains sponsor-owned. |
BVCT is knowledge-rich and patient-data-free: causal biology decompressed directly from the human genome into 2M virtual patients. No ML training on past trial outcomes. No proprietary patient-level data required. Public NCT-level trial design and drug structure go in; head-to-head clinical effect sizes come out. That orthogonality is what makes BVCT compose with every co-scientist on the market.
Typical use cases (hypothetical): Gemini surfaces a target inside MSD — BVCT scores the Phase-III HR of every drug design against it. Claude proposes an indication expansion inside BMS — BVCT returns the differential effect size vs. SoC in that population. GPT writes a trial protocol inside Novo Nordisk — BVCT pre-simulates the readout HR / OR / RR vs. SoC therapies. Isomorphic designs a binder for Lilly or Novartis — BVCT tests whether that binder beats SoC at the Phase-III endpoint. BVCT does not replace any of these systems. It tells the sponsor which of their outputs deserve clinical capital to generate blockbuster revenue.
Pattern-matching AI tells you what looks like the answer. Structural AI tells you what binds the answer. BVCT tells you what beats the answer the standard of care already gives the patient.
BVCT’s performance is measured on metrics that FDA’s Jan 2025 draft AI guidance lists as examples — PPV, NPV, F1 — plus DOR, across a timestamp-audited prospective cohort delivered before each readout:
A worked example, on the public record: BVCT Prediction No. 1959 — mavacamten in non-obstructive HCM, recorded technical call partial SUCCESS (statistically maybe significant yet weak clinical benefit in KCCQ compared with placebo) — timestamped on . ODYSSEY-HCM missed both primary endpoints 12 days later; the dashboard classifies the readout as True Negative (TN). Prediction No. 2094 — aficamten will succeed in the same indication, recorded technical call SUCCESS (statistically significant clinical benefit in KCCQ CSS compared with placebo) — timestamped . ACACIA-HCM confirmed it 170 days later; the dashboard classifies the readout as True Positive (TP). Two co-scientist-grade hypotheses, opposite verdicts, both right ex ante, from molecular differences alone. The full case: aficamten vs mavacamten.
Visit bvct.bioinvestgpt.com to test a small set of wet-lab-validated candidates from your existing co-scientist loop, evaluated across 2–3 indications and proposed trial designs. BVCT returns simulated registrational-endpoint HR vs. SoC per asset / indication / design, with mechanistic rationale. ~1 week per asset, < 0.1% the cost of a head-to-head Phase III.
Audit the 705-prediction track record and the No. 1959 / No. 2094 cases at data.bioinvestgpt.com: ODYSSEY-HCM and ACACIA-HCM.
Clinical Foresight Before Decisions. Earlier is Kinder for Patients.