Abstract
Recent massive increases in the number of sequences available in public databases challenges current experimental approaches to determining protein function. These methods are limited by both the large scale of these sequences databases and the diversity of protein functions. We present a deep learning Graph Convolutional Network (GCN) trained on sequence and structural data and evaluate it on ~40k proteins with known structures and functions from the Protein Data Bank (PDB). Our GCN predicts functions more accurately than Convolutional Neural Networks trained on sequence data alone and competing methods. Feature extraction via a language model removes the need for constructing multiple sequence alignments or feature engineering. Our model learns general structure-function relationships by robustly predicting functions of proteins with ≤ 30% sequence identity to the training set. Using class activation mapping, we can automatically identify structural regions at the residue-level that lead to each function prediction for every protein confidently predicted, advancing site-specific function prediction. De-noising inherent in the trained model allows an only minor drop in performance when structure predictions are used, including multiple de novo protocols. We use our method to annotate all proteins in the PDB, making several new confident function predictions spanning both fold and function trees.
