Protein contact prediction by joint evolutionary coupling analysis across multiple families

Protein contacts contain important information for protein structure and functional study, but contact prediction is very challenging especially for protein families without many sequence homologs. Recently evolutionary coupling (EC) analysis, which predicts contacts by analyzing residue co-evolution in a single target family, has made good progress due to better statistical and optimization techniques. Different from these single-family EC methods, this paper presents a joint multi-family EC analysis method that predicts contacts of one target family by jointly modeling residue co-evolution in itself and also (distantly) related families with divergent sequences but similar folds, and enforcing their co-evolution pattern consistency based upon their evolutionary distance. To implement this multi-family strategy, we use a set of correlated multivariate Gaussian distributions to model related families, the inverse covariance matrix of each distribution encoding the contact pattern of one family. Then we co-estimate the inverse covariance matrices subject to the constraint that they shall share similar patterns. Experiments show that joint multi-family EC analysis can reveal many more native contacts than single-family analysis even for a target family with 4000 sequence homologs, which makes many more protein families amenable to co-evolution-based structure and function prediction.