Accurate identification of protein allosteric sites is critical for understanding the long-range regulatory mechanism and promoting allosteric drug design. Existing methods rely heavily on pocket-level geometric features and implicitly assume that allosteric regulation is confined to pockets. In reality, allosteric regulation is mediated by energy propagation and cooperative interactions, which often extend beyond the pocket boundaries. Thus, existing methods not only lack physical interpretability but also fail to detect residues outside of the pockets. To address these limitations, we developed ZHMolEReP, a residue-level allosteric site prediction tool that integrates perturbation response scanning (PRS) with free energy response approximation, establishing a concise and physically grounded seven-feature representation. By combining mechanical descriptors with local structural features, the model captures allosteric signals by using a minimal set of residue-level features. On the ASBench benchmark data set, ZHMolEReP outperforms existing methods, achieving a recall of 0.7037, an AUC of 0.7858, and successfully identifying allosteric sites in 33 out of 40 proteins, representing increases of 141.8%, 35.4%, and 26.9% over baseline methods. The performance advantage is maintained on the independent AlloReverse test set. Ablation experiments and SHAP analysis reveal that the model predictions are influenced both by physically grounded features and data set-specific structures. Notably, some statistically dominant features strongly interact in ways that obscure the contribution of physically interpretable features, further highlighting the importance of guiding the model to learn truly physically meaningful patterns in allosteric regulation. In summary, ZHMolEReP provides a framework with physical interpretability and strong generalization ability for allosteric site identification, supporting both allosteric mechanism research and drug discovery.
Xing et al. (Fri,) studied this question.