PIGEON: Pocket-Inferred Geometric Ensemble Flexible Docking

Molecular docking is a central tool for structure-based drug discovery, but realistic docking remains challenging because receptor conformations are flexible, a receptor-ligand pair may admit multiple stable holo states, and the true binding-pocket residues are often unknown in prospective applications. We introduce PIGEON (Pocket-Inferred Geometric Ensembles for Flexible Docking), a manifold flow-matching framework that performs ensemble flexible docking without ground-truth pocket residues. PIGEON represents receptor-ligand complexes with a time-conditioned heterogeneous graph and predicts residual motions for ligand translation, ligand rotation, ligand torsions, receptor residue frames, and receptor side-chain torsions. To guide ligand translation in the absence of pocket labels, PIGEON constructs a receptor-intrinsic buriedness field, a ligand-centered spherical harmonic pocket field, and a compact bank of pocket-aware candidate directions, then uses a learnable staged selector to choose between pocket-ingress and local-fitting motions. Empirically, PIGEON recovers accurate ligand poses while maintaining physically plausible full-complex geometry, providing a practical route to generating coupled ligand-receptor ensembles under realistic pocket-unknown conditions.