An engine for discovering cures.

ADDE — Autonomous Drug-Discovery Engine — filters, ranks, designs, and optimizes leads across the entire therapeutic space — with no human in the loop.

Novel Gate

Removes universally inactive molecules before screening even begins. 6–13× better than the best baseline at recovering actives, and 99% of the theoretical maximum on enrichment. Different because it learns bioactivity itself, not drug-likeness rules of thumb.

rfr99 lower is better; EF1% higher is better. Beats the best baseline by 6–13× on rfr99; reaches 99% of the theoretical maximum on EF1%.
Method	Temporal rfr99	Hard rfr99	Hard EF1%
Novel Gate	0.014	0.094	29.7
Chemprop D-MPNN	0.233	0.602	23.3
Best ECFP (XGBoost)	0.214	0.691	19.8
chem-LM (ChemBERTa-2)	0.178	0.782	8.2

Novel Rank

Ranks ligands by predicted binding affinity to a target. Beats Boltz-2 on CASP16 and DAVIS, tops the best live CASP16 result, and runs 40,000–150,000× faster. Different because it predicts affinity directly, without first computing a binding pose.

Tops the best live CASP16 result (SVR_Conjoint, 0.43); 40,000–150,000× faster than Boltz-2 depending on protein length. Lower wall-clock is better.
Benchmark	Boltz-2	Novel Rank
CASP16, N-weighted τ	0.403	0.456
CASP16, N-weighted r	0.595	0.627
DAVIS, per-target τ	0.267	0.462
DAVIS, per-target r	0.401	0.620
DAVIS success rate	92.3%	100%
DAVIS wall-clock	240,266 s	43 s

FoldCraft open source

Designs binders to a specified structural fold — a capability unique to FoldCraft. Matches or beats BoltzProt-1 on OpenMM physics evaluations. The foundation of a next-generation generative family.

See it on GitHub ↗ Read the paper ↗

OpenMM interface ΔE, kcal/mol, lower is better. Raw-pool scores are effectively tied; the win is the delivered set — the designs FoldCraft actually hands off after its selection step.
PD-L1	BoltzProt-1	FoldCraft
Condition on target fold	—	6 folds
Interface ΔE — raw pool	−33.2	−33.1
Interface ΔE — delivered	−36.5	−46.4