AF2BIND predicts small-molecule-binding residues on a target protein by leveraging AlphaFold2's internal pairwise representation. It uses 20 'bait' amino acids to extract the binding signal in the absence of an actual ligand. The model is a small linear head trained on top of AF2 pair features; per-residue p(bind) probabilities are produced as well as a per-residue x per-bait-amino-acid score matrix correlated with the chemical properties of the unknown ligand.