Accurate Identification of Protein Binding Sites for All Drug Modalities Using ALLSites.

Proteins interact with diverse molecular modalities, yet the incomplete identification of their binding sites has left the proteome-wide druggability largely underexplored. Although various computational methods have been developed for the prediction of protein binding sites, existing approaches are limited by their specificity to a single drug modality, dependence on high-quality structural data, or insufficient predictive accuracy. Here, a unified sequence-based framework, ALLSites, is constructed to identify proteome-wide binding sites across all drug modalities. Leveraging ESM-2 embeddings, ALLSites integrates a gated convolutional network with a transformer architecture to capture both global and local sequence features, effectively modeling residue interactions directly from sequence. This design bridges the gap between sequence-based and structure-based approaches, enabling ALLSites to achieve superior predictive performance across diverse drug modalities, including proteins, peptides, small molecules, carbohydrates, DNA, and RNA. It achieves state-of-the-art performance among sequence-based methods and matches the accuracy of the best structure-based tools. By enabling accurate and structure-free binding site prediction across all drug modalities, ALLSites is expected to expand the druggable proteome and provide a powerful resource for drug discovery.