Transfer learning improves the prediction of ligand binding sites at protein–membrane interfaces

Abstract Membrane-exposed ligand binding sites in membrane proteins are emerging as important modulatory targets in drug discovery, however, their prediction remains challenging due to limited structural data. Here, we present DeepPocket-Mem, a deep learning-based approach that applies transfer learning to a convolutional neural network originally developed for soluble protein binding sites. A curated dataset comprising 605 protein–ligand complexes was used to fine-tune the components responsible for ranking candidate binding sites and predicting their shapes, followed by validation using measures of binding-site localization and spatial overlap. While transfer learning of the classification module alone did not improve discrimination between binding and non-binding sites, fine-tuning of the segmentation module enhanced pocket localization and increased the number of correctly identified sites among top-ranked predictions, as reflected by improved success rates based on distance center-to-center (DCC) and discretized volume overlap (DVO). Further improvements were achieved by restricting training to drug-like ligands with confirmed biological activity, underscoring the importance of biologically relevant data for model performance. Overall, these results demonstrate the potential of transfer learning for binding site prediction under data-limited conditions and establish DeepPocket-Mem as a useful tool for identifying membrane-embedded binding sites. Concurrently, they highlight the need for expanded structural datasets to further improve prediction accuracy.