Protein Language Models and Structure-Based Machine Learning for Prediction of Allosteric Binding Sites in Protein Kinases: An Explainable AI Framework Grounded in Energy Landscape-Encoded Frustration

Reliable identification of allosteric binding sites remains a major bottleneck in structure-based drug discovery, particularly in protein kinase families where such sites are often structurally cryptic, evolutionarily non-conserved, and sparsely populated. In this work, we present a systematic analysis of binding site prediction across a rigorously curated dataset of human kinase-ligand complexes, encompassing 453 kinases and spanning five inhibitor classes: Type I, Type I.5, and Type II (orthosteric ATP-competitive) and Type III/IV (non-ATP allosteric) modulators. We employed the pretrained protein language model (PLM) ESM2-650M model that was fine-tuned for prediction of protein-ligand binding sites by replacing the original masked language modeling head with a token-level classification head that acts as a projection layer that maps the high-dimensional latent representation of each residue to a scalar probability score for a given protein residue to be part of the binding site. We employed this fine-tuned sequence-based PLM and structure-based detection approach P2Rank for identification of orthosteric and allosteric binding sites in protein kinases. Our analysis reveals a stark performance divergence: while both methods achieve high precision-recall (AUPR = 0.64-0.76) on orthosteric sites, PLM performance collapses on allosteric sites (AUPR = 0.06), despite retaining moderate ranking ability (AUROC = 0.70). This deficit persists even after strict control for sequence similarity, structural redundancy, and extreme class imbalance (allosteric residues constitute