Identifying Functional Protein Positions Using Neural Networks

Abstract

Characterizing proteins, which mediate a wide array of cellular processes by bind-ing various ligands, is a major aim of computational biology. While proteins maycontain hundreds of amino acids, often only a few are typically involved in inter-actions with biologically relevant ligands. The most direct approach to determinewhich amino acid residues within a protein are involved in binding is through ex-perimental methods, but only relatively few proteins have been captured in com-plex with a relevant ligand. To bridge this gap, we train a bidirectional Long ShortTerm Memory (BiLSTM) model to predict the binding properties of each aminoacid position from sequence-based features for five ligand groups: DNA, RNA,protein, ion, and metabolite. To increase power, we extend our set of true labels be-yond the limited experimental data by using protein domain-based inferred bind-ing scores. We then evaluate our model by measuring performance on a held-outtest set, and compare performance to a baseline XGBoost model, as well as an ex-isting method. In both these comparisons, our model performs at least as well orbetter for all ligand groups. Because they reflect the binding potential of individ-ual amino acid sites, our predictions can also provide insight into both healthy anddiseased protein function.