Biomaths (Mathematics) – EPSRC CMPH Seminar
Vladimir Gligorijevic1 Tomasz Kosciolek2 Richard Bonneau1
1 Center for Computational Biology, Flatiron Institute, Simons Foundation, New York, USA
2 Małopolska Centre of Biotechnology, Jagiellonian University, Kraków, Poland
The human gut microbiome is estimated to harbor over 3 million unique protein-coding gene families. Only a fraction of them are experimentally annotated and therefore require computational predictions. Community-wide benchmarking efforts such as CAFA show that homology-based function annotation approaches are lacking and require more sophisticated approaches. In this talk, we will introduce deepFRI (deep learning-based functional residue identification), our recently proposed method based on Graph Convolutional Networks . We use protein families, within a deep learning structure prediction framework, to predict protein 3D structures. We show that deepFRI achieves state-of-the-art accuracy in predicting gene ontology terms of proteins with predicted 3D structural models. We are now in a position to functionally annotate microbial genomes and metagenomes with higher coverage and accuracy. We may also start addressing microbe-microbe and host-microbiome protein-protein interactions to determine the mechanisms of microbiota-induced immune response. An added value of deepFRI is its ability to produce region-specific structure predictions through class activation mapping.
 Gligorijevic, V., Renfrew, P.D., Kosciolek, T., Leman, J.K., Cho, K., Vatanen, T., Berenberg, D., Taylor, B.C., Fisk, I.M., Xavier, R.J. and Knight, R., 2019. Structure-Based Function Prediction using Graph Convolutional Networks. bioRxiv, p.786236.