Encoding protein structure with functions on graphs

Abstract

The application of machine learning and datamining to the analysis and prediction of protein structure is a research area with potentially high impact in both computer science and biology. Proteins structures are inherently complicated objects with a mixture of crisp and fuzzy properties. Therefore developing effective representations for them is a research problem in itself, while quantifying and predicting properties and structure is of immediate importance in structural biology. This paper focuses on developing a compact, effective, efficient and accurate representation of protein structure that is compatible with widely used machine learning tools like the SVM. Graphs based on Delaunay triangulation are used to represent the structure, and then functions are constructed from these graphs to develop constant-size representations of protein structure that are tightly bound to the amino acid sequence. The representations preserve sufficient information to be valuable for model vs. experimental structure classification and regression analysis of model quality.