Automated protein function prediction using an ensemble of deep neural networks

Abstract

Proteins are polymers of amino acids produced in cells of living organisms to perform diverse bodily functions. Over the years, different schemes to describe protein functions in a standard way have emerged, and Gene Ontology (GO) remains the most popular among them. Despite how beneficial it is to know the exact function of the proteins, the experimental procedures to determine the protein function are very laborious and time-consuming. Therefore, to keep up with the rate at which new proteins are sequenced, numerous computational methods that use various protein features for Automated Function Prediction (AFP) have emerged over the years. Starting from the earliest statistical approaches, the field has evolved into models that use the latest deep-learning techniques. We have developed a model for AFP using recurrent and convolutional neural networks with protein sequences and protein-protein interaction data, which has the potential to achieve comparably good results for human datasets.