Allofusion: Allosteric Site Prediction Based on Language Models and Multi-Feature Fusion

Allosteric regulation plays a crucial role in modulating protein function and has emerged as a promising strategy in drug discovery. However, current computational methods often rely on static structures or single-modality features, limiting their ability to identify allosteric sites that are transient, cryptic, or located outside conventional pockets. Here, we propose AlloFusion, a residue-level multimodal prediction framework for accurate allosteric site prediction. AlloFusion integrates the embedding representations from a pretrained protein language model, biochemical properties of residues, and evolutionary profiles derived from position-specific scoring matrices. By leveraging these diverse features, AlloFusion effectively classifies allosteric site-forming residues (AFRs) and nonallosteric residues (FRs) in protein sequences while localizing allosteric sites. On the ASD2023 data set, AlloFusion outperforms mainstream methods in specificity, precision, F1-score, and AUC. Furthermore, on the independent test set D24, AlloFusion correctly predicts 23 out of 24 allosteric sites, significantly improving prediction accuracy. The comprehensive results demonstrate that AlloFusion is a promising method for allosteric site prediction. The source code of AlloFusion is available at https://github.com/hjb-001/AlloFusion.