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

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Allosteric regulation is key to protein function and drug discovery.
• Current computational methods struggle with transient, cryptic, or non-pocket allosteric sites.
• Accurate identification of allosteric sites is critical for targeted therapeutics.

Purpose of the Study:

• To develop a novel computational framework, AlloFusion, for accurate prediction of allosteric sites.
• To integrate multimodal features for improved identification of allosteric site-forming residues (AFRs).
• To address limitations of existing methods in detecting challenging allosteric sites.

Main Methods:

• Developed AlloFusion, a residue-level multimodal prediction framework.
• Integrated protein language model embeddings, residue biochemical properties, and evolutionary profiles (position-specific scoring matrices).
• Classified allosteric site-forming residues (AFRs) and nonallosteric residues (FRs) for site localization.

Main Results:

• AlloFusion demonstrated superior performance on the ASD2023 dataset, outperforming mainstream methods in specificity, precision, F1-score, and AUC.
• Achieved high accuracy on the independent D24 test set, correctly predicting 23 out of 24 allosteric sites.
• Significantly improved prediction accuracy for allosteric sites, including those difficult to detect.

Conclusions:

• AlloFusion is a promising and accurate method for allosteric site prediction.
• The multimodal approach effectively captures diverse features for robust allosteric site identification.
• AlloFusion offers a valuable tool for advancing drug discovery through precise allosteric site localization.