Unrefined ligand occupancy masks protein-binding site heterogeneity

Ligands in co-crystal structures are typically assumed to bind at full occupancy, even though many interact sub-stoichiometrically. This modeling discrepancy obscures the true conformational landscape of binding sites, masking underlying flexibility and the coexistence of bound and unbound states. To gauge how prevalent this oversight is, we re-refined ligand occupancies in ∼10k deposited structures. This revealed that 35% contain partially occupied ligands, 3× more than what is reflected by occupancies in the original structures. We present examples for 3 categories of overlooked structural heterogeneity: alternative side-chain conformations, shifting water networks, and additional ligand poses. Notably, our nuanced refinement approach can reveal unappreciated binding at allosteric sites. These inaccuracies are not always flagged by standard model quality metrics. Ultimately, this study encourages better practices for modeling protein-ligand structures. Increasing the fidelity of models is essential to the success of downstream applications, especially AI/ML methods.