A random forest classifier for protein-protein docking models.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning

Background:

• Accurate 3D models of protein-protein complexes are crucial for understanding biological processes.
• Existing protein-protein docking software generates numerous models, making it challenging to identify the correct ones.

Purpose of the Study:

• To develop and validate a machine learning approach for selecting accurate 3D protein-protein docking models.
• To create a robust classifier that can effectively discriminate correct docking models from incorrect decoys.

Main Methods:

• Generated a large dataset of protein-protein docking models using HADDOCK, FTDock, and ZDOCK.
• Trained and optimized machine learning classifiers, including Random Forest, Support Vector Machine, and Perceptron, using 158 scoring functions.
• Developed a feature selection algorithm and optimized hyperparameters to create the COnservation Driven Expert System (CoDES).

Main Results:

• The Random Forest algorithm demonstrated superior performance compared to Support Vector Machine and Perceptron.
• The optimized Random Forest classifier, CoDES, achieved state-of-the-art accuracy in identifying correct 3D docking models.
• CoDES effectively discriminated correct from incorrect decoys across global parameters and top-ranked positions.

Conclusions:

• CoDES represents a significant advancement in the computational prediction of protein-protein complex structures.
• The developed machine learning approach enhances the reliability of protein-protein docking results.
• CoDES provides a powerful tool for researchers needing to identify accurate protein complex models from docking simulations.