Jilong Zhang, Xiaohan Sun, Zhixiang Wu, Jingjie Su, Xi Zhang, Chunhua Li
Allostery is a fundamental mechanism for regulating protein functions. Accurately identifying the allosteric sites of proteins is crucial for understanding the allosteric process and advancing the development of allosteric drugs. Here, we develop AlloEF, an effective method for protein allosteric site prediction, which adopts a soft-voting classifier with LightGBM, Random Forest, and XGBoost combined. The multidimensional features are integrated, including the transfer entropy-based dynamic property and energetic frustration introduced by us, as well as the traditional sequence-, structure-, and network topology-based characteristics. Specifically, the Spatial Neighborhood Feature Aggregation (SNFA) strategy is proposed, which shows good effectiveness in capturing the local environmental information on a residue. The Boruta algorithm is applied to select an optimal feature set. To address the class imbalance in the training dataset, the SVM-SMOTE technique is utilized to generate synthetic minority class samples. The test shows that our model AlloEF achieves an F1 of 0.630 and an MCC of 0.609 on the independent test set, outperforming the existing methods. AlloEF can detect not only the allosteric sites in the canonical allosteric pockets but also the ones distributed beyond the pocket regions. Overall, this work establishes a new benchmark model for allosteric site prediction, helpful for strengthening our understanding of the protein allosteric mechanism and designing allosteric modulators.