Inferring interaction partners from protein sequences using mutual information

Abstract

Specific protein-protein interactions are crucial in most cellular processes. They enable multiprotein complexes to assemble and to remain stable, and they allow signal transduction in various pathways. Functional interactions between proteins result in coevolution between the interacting partners, and thus in correlations between their sequences. Pairwise maximumentropy based models have enabled successful inference of pairs of amino-acid residues that are in contact in the three-dimensional structure of multi-protein complexes, starting from the correlations in the sequence data of known interaction partners. Recently, algorithms inspired by these methods have been developed to identify which proteins are specific interaction partners among the paralogous proteins of two families, starting from sequence data alone. Here, we demonstrate that a slightly higher performance for partner identification can be reached by an approximate maximization of the mutual information between the sequence alignments of the two protein families. This stands in contrast with structure prediction of proteins and of multiprotein complexes from sequence data, where pairwise maximum-entropy based global statistical models substantially improve performance compared to mutual information. Our findings entail that the statistical dependences allowing interaction partner prediction from sequence data are not restricted to the residue pairs that are in direct contact at the interface between the partner proteins. Author summarySpecific protein-protein interactions are at the heart of most intra-cellular processes. Mapping these interactions is thus crucial to a systems-level understanding of cells, and has broad applications to areas such as drug targeting. Systematic experimental identification of protein interaction partners is still challenging. However, a large and rapidly growing amount of sequence data is now available. Recently, algorithms have been proposed to identify which proteins interact from their sequences alone, thanks to the co-variation of the sequences of interacting proteins. These algorithms build upon inference methods that have been used with success to predict the three-dimensional structures of proteins and multi-protein complexes, and their focus is on the amino-acid residues that are in direct contact. Here, we propose a simpler method to identify which proteins interact among the paralogous proteins of two families, starting from their sequences alone. Our method relies on an approximate maximization of mutual information between the sequences of the two families, without specifically emphasizing the contacting residue pairs. We demonstrate that this method slightly outperforms the earlier one. This result highlights that partner prediction does not only rely on the identities and interactions of directly contacting amino-acids.

Coevolutive, Evolutive and Stochastic Information in Protein-Protein Interactions

“…SI). Advances in PPI prediction (14)(15)(16)(17)(18) are highly welcome in the contexts of paralog matching, hostpathogen PPI network prediction and interacting protein families prediction. Recent studies suggest strategies like maximizing the interfamily coevolutionary signal (14), iterative paralog matching based on sequence "energies" (15) and expectation-maximization (18), which have been capable of accurately matching paralogs for some study cases.…”

Phylogenetic correlations can suffice to infer protein partners from sequences

“…Indeed, methods based only 29 on sequence similarity and on phylogeny [26,29,30] also allow to predict protein-protein 30 interaction partners. In addition, a method based on mutual information (MI) was 31 recently shown to slightly outperform the DCA-based one [31]. MI includes all types of 32 statistical dependence between the sequences of interacting partners.…”

“…In the second 97 case, we employ the Iterative Pairing Algorithm (IPA) developed in Refs. [21,31] to 98 bootstrap the predictions starting from initial random within-species pairings. Below, we 99 present the various inference methods assuming that there is a training set.…”

“…The extension 100 to the training set-free case is then performed exactly as described in Refs. [21,31].…”

“…Pseudocount. When dealing with real protein sequences, pseudocounts are often 108 introduced to avoid mathematical issues such as divergences due to amino-acid pairs 109 that never appear, both with DCA [5,6,12,14] and with MI [31]. Introducing a 110 pseudocount weight Λ, which effectively corresponds to adding a fraction Λ of chains 111 with uniformly distributed states, the corrected one-body frequencies readf i (σ i ) = 112…”