Improving Molecular-Level Protein Docking and Interpreting System-Level Cancer Mechanisms through Machine Learning

Abstract

Protein-protein interactions (PPIs) are crucial to cellular function, yet researchers still have much to discover about their mechanisms. At the molecular level, two new computational tools are proposed in this study to facilitate protein docking: 1. A regression model for predicting both the direction and extent of protein conformational change, especially the extent in this study, provides a new approach for structural ensemble generation and conformational sampling. 2. A classifier for assessing whether a protein pair is suitable for rigidity docking provides a method for performing a sanity check before uniformly applying rigidity assumption in protein docking. At the intra-cellular system level, PPIs participate intensively in the propagation of mutational effects of cancer, which is well-known as a complex disease often derived from "driver genes" containing pathogenic mutations. Here I propose a new machine learning framework with biologically meaningful features to identify driver genes with the help of PPI network topology. Further interpretation of the machine learning model can help us understand cancer mechanisms by explaining the reason why cancer would prefer to attack those positions in network.