Models

Ranks by molecular weight.

A property-based baseline that scores molecules solely by molecular weight. Larger molecules tend to have more intermolecular contacts, so this tests whether models beat the trivial heuristic of size-based ranking.

Ranks by lipophilicity (cLogP).

A property-based baseline that scores molecules by computed lipophilicity (cLogP via RDKit). Greasier compounds often bind more strongly through the hydrophobic effect, but this is undesirable in drug design.

Physics-based molecular docking.

Deterministic molecular docking using Smina (a fork of AutoDock Vina). Ligands are docked into the known protein binding site and scored by a physics-based energy function. Uses the reference crystal structure from PDB.

Deep learning protein-ligand co-folding.

Uses Boltz2, a deep learning model for joint protein-ligand structure prediction (co-folding). Confidence scores from the model's predicted structures are used as binding affinity proxies.