Benchmarks
Every result here is measured on recognised, held-out benchmarks and scored with the same production weights the platform serves — held-out scaffold splits for the property models, and genuinely blinded tests for binding. We show the strong results and, below, the limits that qualify them.
Gnosis V · Microsomal stability
On the official public test split, Gnosis V’s microsomal-stability model reaches Spearman 0.000 — above the published leaderboard reference — and ships with validated prediction intervals rather than a bare point estimate.
Spearman, official test split
of test compounds in-domain
scale 0 to 1 · Spearman
Gnosis I · Fused binding
On the public FEP+/JACS binding benchmark, the Fused ranking engine lands in the band between OpenFE and FEP+, the physics-based gold standard, while running in minutes per compound instead of the hours of simulation those methods need. And on a fully blinded oncology target, it self-corrects where a naive approach fails.
On par with physics, in minutes
Public FEP+/JACS set · Pearson correlation
scale 0 to 1 · Pearson r
These congeneric series are in-domain, so this is an optimistic upper bound, not a blinded test. The blinded result is shown alongside.
The blinded test
WRN helicase · blinded held-out compounds · Spearman
A naive co-fold ranked worse than random. Measuring which signal to trust against known actives recovered strong rank agreement, on compounds never seen.
Rank-only. Absolute potency is reported only when a per-target calibration gate passes.
Gnosis III · Systemic safety
Production safety heads on the held-out test split. hERG cardiac-risk is the standout. AUROC is shown on a 0.5-to-1.0 scale, so the bar length is the honest distance above random.
Gnosis V · ADMET (classification)
CYP inhibition liability, blood-brain-barrier penetration and P-gp transport, each on its own held-out scaffold split. These are the inhibition heads; the separate CYP substrate heads are weaker and are not reported here.
Gnosis V · ADMET (regression)
Correlation against measured values on held-out splits. Solubility roughly halves the error of a mean predictor. These are rank/triage reads, not calibrated points.
The honest version: your molecule, every module, the limits stated as you have seen here.
