Functional domain annotation by structural similarity

Abstract

The most widely used tools for electronic functional annotation rely on domain annotation. Pfam, one of the most used domain annotation tools, has a list of well-known protein domain sequences called seed sequences from which a domain's sequence signature is drawn and is used for annotating new domains in other proteins. From a biological point of view, structures are more conserved than sequences. Therefore, we expect to annotate more domains in phylogenetically remote proteins if we consider the structural similarity. However, it was not possible on a large scale until recently, as the structure of a limited number of proteins had been resolved experimentally, and computational approaches were not sufficiently accurate. Besides, there were no efficient tools for finding structurally similar proteins quickly. In this study, we took advantage of two state-of-the-art programs, Alphafold2, the pioneer program for predicting protein structures with an accuracy comparable to experiments, and Foldseek, the ultra-fast program for finding structurally similar proteins, to annotate new Pfam domains in the proteome of Trypanosoma brucei based on structural similarity to the Pfam Seed structures. We did this by training a model for estimating the confidence score based on alignment characteristics and greedily selecting the minimally overlapping most confident Pfam predictions. As a result, we predicted over 1100 new Pfam domains in T. brucei, which increases the number of annotated Pfam domains in T. brucei up to 20%. The scripts used in this work are freely available at https://github.com/Pooryamb/FDASS.