Partner determination from protein sequences using class information

Abstract

Mapping protein interaction networks and understanding how duplicated proteins acquire distinct functions require determining which paralogs from interacting protein families bind to each other. Paralogs often fall into functional or architectural classes whose interactions can be conserved across organisms. Here, we introduce CLAPP (CLAss Pooling for Paralog matching), which pools interaction scores across organisms by class to infer one class-to-class map that constrains matching within each organism. It accepts any pairwise score assessing the likelihood of interaction of two proteins. On nine bacterial family pairs, with test sets restricted so that the class map determines every pairing, CLAPP recovered the correct map, and hence every reference pair. Recovering the map required little information: in the two systems tested, one training epoch of a coevolution-based model sufficed, whereas an untrained model performed no better than chance; and in eight of nine systems, a conservation-only score lacking inter-protein couplings also gave perfect matching. On unrestricted test sets, where the class map no longer uniquely determines every pairing, accuracy depended on the quality of the class labels: on 320 chaperone pairs, CLAPP performed below chance with unsupervised labels but reached 94\% with phylogeny-informed ones; remaining errors involved paralogs which can interact promiscuously. Paralog matching therefore need not rely on strong coevolutionary signals within individual pairs: when interactions between classes are conserved and labels are reliable, CLAPP can use even weak or noisy scores to accurately assign interaction partners.