Accurately predicting binding affinity of protein-protein complexes is significant for gaining deeper insights into complex biological mechanisms. Given that binding between proteins primarily occurs at the interface region, previous studies have demonstrated that the number of inter-residue contacts (ICs) and the buried surface area (BSA) are critical interface features. However, existing models generally used these two types of interface features separately, ignoring integrating them effectively to achieve high prediction accuracy. In this study, utilizing kernel density estimation-based mutual information and the Hadamard product, we proposed an effective approach that integrates BSA and ICs to construct the novel integrated interface features that embody dual structural information, and further derived our feature set. Subsequently, the proposed feature set was input Deep Neural Network (DNN), and a hybrid DNN model was further developed following our hybrid modeling strategy. To enhance its prediction performance, a combined activation function was customized for the output layers. Ultimately, using four-fold cross-validation, our hybrid DNN model achieved a Pearson correlation coefficient (R) of 0.88 and a root mean square error (RMSE) of 1.301 kcal/mol, and we verified its good generalization capability, achieving R = 0.82 and RMSE = 1.21 kcal/mol on the external test set derived from the SKEMPI 2.0 database. Compared with existing approaches, our method consistently exhibited superior predictive performance, validating the effectiveness of the proposed method for protein-protein binding affinity prediction. Moreover, the integration strategy for binding affinity representation and the hybrid modeling method may be helpful for related research.
扫码关注我们
求助内容:
应助结果提醒方式:
