A hybrid DNN model using novel integrated interface features for predicting protein-protein complexes binding...

Area of Science:

• Computational Biology
• Biophysics
• Structural Biology

Background:

• Accurate prediction of protein-protein binding affinity is vital for deciphering complex biological processes.
• Key interface features like inter-residue contacts (ICs) and buried surface area (BSA) are known determinants of binding.
• Existing models often analyze BSA and ICs separately, limiting prediction accuracy.

Purpose of the Study:

• To develop an integrated approach for predicting protein-protein binding affinity.
• To create novel interface features by combining BSA and ICs using kernel density estimation and Hadamard product.
• To build and validate a hybrid deep neural network (DNN) model for enhanced prediction.

Main Methods:

• Integration of BSA and ICs using kernel density estimation-based mutual information and Hadamard product to create dual-information interface features.
• Development of a hybrid deep neural network (DNN) model incorporating the novel feature set.
• Customization of a combined activation function for output layers to improve prediction performance.
• Validation using four-fold cross-validation and an external test set from the SKEMPI 2.0 database.

Main Results:

• The hybrid DNN model achieved a Pearson correlation coefficient (R) of 0.88 and a root mean square error (RMSE) of 1.301 kcal/mol during cross-validation.
• The model demonstrated good generalization capability, with R = 0.82 and RMSE = 1.21 kcal/mol on the external test set.
• The proposed method significantly outperformed existing approaches in predicting protein-protein binding affinity.

Conclusions:

• The integration strategy for BSA and ICs provides a powerful method for representing binding affinity.
• The developed hybrid DNN model offers superior predictive performance for protein-protein binding affinity.
• This approach offers a valuable tool for computational biology and drug discovery research.