The human gut microbiome is estimated to have over 2 million unique protein-coding genes. Only a fraction of them is experimentally annotated and therefore require computational predictions. Community-wide experiments such as CAFA have shown that homology-based function annotation approaches are limited and require more sophisticated approaches. With the maturing of de novo structure prediction methods and the rise of deep learning techniques, it now becomes possible to generate high-throughput structure and function predictions for microbial proteins.

In this talk, I will introduce deepFRI (deep functional residue identification), our recently proposed method based on Graph Convolutional Networks. deepFRI is able to accurately predict GO terms from protein sequences and predicted 3D structures. The structural component of the method not only improves predictions, but also brings residue-level saliency mapping. The mapping provides insight into putative functional sites allowing for biological interpretation, hypothesis generation or the design of targeted validation experiments.