Evaluating Machine Learning Techniques for Decoy Detection of Protein-Protein Interactions

Generating accurate computational models for protein-protein interfaces (PPIs) and determining the quality of these models remains a significant challenge. Over the past two decades, several methods have been developed to generate and score PPIs. There are two main approaches for PPI scoring: physics-based forcefields that include protein stereochemistry and van der Waals and electrostatic interactions, and knowledge-based scoring functions that based on experimentally determined PPI structures from the Protein Data Bank. With advances in machine learning, neural networks can also be used for PPI model generation and scoring.

In this work, we constructed a dataset of high-resolution x-ray crystal structures of protein heterodimers and generated PPI models for each experimental target using current computational protein docking methods: ZDOCK, HDOCK, and Rosetta. Each method applies its own scoring function to rank its models. To assess their accuracy, we scored all models against the three separate scoring functions, as well as published neural networks train in classifying PPI models.