Compound Pipeline
Compounds ranked by multi-fidelity Bayesian scoring across 4 layers: Docking (w=0.15), ADMET (w=0.25), MD Stability (w=0.35), Binding (w=0.25). Safety badges from ADMET-AI GNN predictions.
Stage 0: Generative design (GenMol NIM) · Stage 1: ADMET-AI GNN 27 endpoints · Stage 1.5: Retrosynthesis (AiZynthFinder) · Stage 2: Safety gate · Stage 3a: Rigid docking (FlowDock) · Stage 3b: Induced-fit (NeuralPLexer3) · Stage 4: Selectivity filter · Stage 5: MD simulation (OpenMM) · Stage 6: Free energy (BindFlow FEP) · Stage 7: Orthogonal assay · Stage 8: Lab validation
▶How does Compound Scoring work?
Compounds are ranked by a multi-fidelity Bayesian score combining four computational evidence layers. Missing layers widen the confidence interval rather than penalizing the score.
- Docking (w=0.15) — FlowDock pose confidence + Vina binding energy. Low weight: fast but coarse. Min 5 poses per compound.
- ADMET (w=0.25) — ADMET-AI GNN: BBB penetration, hepatotoxicity, hERG liability, solubility, CYP inhibition. Score = fraction of safety-passed endpoints.
- MD Stability (w=0.35)— 100ns OpenMM simulation. RMSD <3A and stable pose over final 20ns required. Highest weight: measures residence time directly.
- Binding Free Energy (w=0.25) — MMPBSA/BindFlow FEP from MD trajectory. Units: kcal/mol. More negative = stronger binding.
score = sum(layer * weight) / sum(weights_present). Status: early = docking only, promising = ADMET + MD, ready = all 4 layers.