Abstract
The knowledge of protein-ligand binding sites is vital prerequisite for any structure-based virtual screening campaign. If no prior knowledge about binding sites is available, the ligand-binding site prediction methods are the only way to obtain the necessary information.
Here we introduce P2RANK, a novel machine learning-based method for prediction of ligand binding sites from protein structure. P2RANK uses Random Forests learner to infer ligandability of local chemical neighborhoods near the protein surface which are represented by specic near-surface points and described by aggregating physico-chemical features projected on those points from neighboring protein atoms.
The points with high predicted ligandability are clustered and ranked to obtain the resulting list of binding site predictions. The new method was compared with a state-of-the-art binding site prediction method Fpocket on three representative datasets.
The results show that P2RANK outperforms Fpocket by 10 to 20 percentage points on all the datasets. Moreover, since P2RANK does not rely on any external software for computation of various complex features, such as sequence conservation scores or binding energies, it represents an ideal tool for inclusion into future structural bioinformatics pipelines.