Evaluation of Small-Molecule Binding Site Prediction Methods on Membrane-Embedded Protein Interfaces.

Increasing structural and biophysical evidence suggests that many drug molecules bind to the protein-membrane interface region in membrane protein structures. An important starting point for drug discovery is the determination of a ligand's binding site; however, this information is missing for many membrane proteins, especially for their membrane-embedded parts. Therefore, we tested the performance of computational methods for ligand binding site prediction in the protein intramembrane region. We compiled data sets containing GPCR- and ion channel-ligand complexes and compared method performance relative to a soluble protein data set obtained from PDBBind. We tested state-of-the-art geometry-based (Fpocket, ConCavity), energy probe-based (FTSite), machine learning-based (P2Rank, GRaSP), and deep learning-based (PUResNet, DeepPocket, PUResNetV2.0) methods and evaluated them using the center-to-center distance (DCC) and discretized volume overlap (DVO) between the predicted binding site and the actual ligand position. The three best-ranking methods based on success rates on GPCRs were DeepPocket, PUResNetV2.0, and ConCavity, and for ion channels, these were DeepPocket, PUResNetV2.0, and FTSite. However, average DCC and DVO values were lower for all methods compared to the soluble protein data set, for which DVO and normalized DCC values ranked between 0.33 and 0.72 in their best case, respectively. In conclusion, this study provides an overview of the performance of state-of-the-art binding site prediction methods on their ability to identify pockets in the protein-membrane interface region. It also underscores the need for further method development in the prediction of protein-membrane ligand binding sites.