Coevolutionary Algorithm Maps Species-Specific Bacterial Protein Interactomes Without Machine Learning Data

A computational study published in PLOS Computational Biology introduces ContextMirror 2.0, an unsupervised sequence-based framework designed to map species-specific protein-protein interactomes across bacterial proteomes without relying on machine learning training sets or physical structural coordinates. Developed by researchers Miguel Fernandez-Martin, Camila Pontes, Victoria Ruiz-Serra, and Alfonso Valencia at the Barcelona Supercomputing Center and partner institutions, the method leverages phylogenetic coevolutionary histories to detect functional relationships between proteins at a whole-proteome scale. By evaluating how evolutionary divergence patterns mirror each other across thousands of diverse orthologs, the algorithm reconstructs known cellular networks and reveals previously uncharacterized lineage-specific protein communities. The research team demonstrated the utility of the framework by predicting the comprehensive interactomes of three bacterial species: Escherichia coli, Salmonella enterica, and Staphylococcus aureus. Across these organisms, the top one thousand predicted pairwise interactions achieved an empirical precision of approximately forty percent when evaluated against established experimental databases, outperforming random chance expectations by roughly eight-point-five fold.

Mapping complete interactomes remains an essential frontier for deciphering cellular physiology, disease mechanisms, and evolutionary adaptation, yet most protein associations in non-model organisms and understudied bacteria lack experimental verification. Existing computational prediction methods face considerable hurdles that restrict their applicability across diverse species. Supervised machine learning algorithms and deep neural networks depend on extensive curated datasets of known interactions, an architecture that hinders reliable extrapolation beyond a handful of well-studied reference genomes and frequently suffers from limited explainability. Meanwhile, structure-based computational approaches such as molecular docking and template modeling require high-resolution physical coordinates or intensive computational resources, making systematic proteome-wide predictions impractical. Evolutionary methods provide a compelling unsupervised alternative because they derive predictive power entirely from natural sequence variation across homologous protein families. However, earlier coevolutionary tools frequently struggled with elevated false-positive rates driven by phylogenetic background noise or failed to capture subtle interaction differences that separate closely related bacterial lineages.

To resolve these computational challenges, ContextMirror 2.0 establishes a multi-stage bioinformatic pipeline that redesigns previous ContextMirror algorithms for high-throughput scalability. The automated initial filtering stage retrieves orthologous sequences across the NCBI non-redundant database using standard BLAST reciprocal best hit searches, requiring at least seventy percent query length alignment, sixty percent query coverage, and thirty percent sequence identity. Crucially, the protocol excludes any protein pair that does not share orthologs across a minimum threshold of one thousand matching species, ensuring sufficient evolutionary depth for robust coevolutionary analysis. For all retained candidate pairs, the pipeline constructs paired multiple sequence alignments using MAFFT, incorporating an automated strategy that alternates between progressive and iterative alignment modes while accommodating ambiguous amino acid residues. Pairwise evolutionary distance matrices are subsequently calculated using the McLachlan amino acid residue homology matrix, from which MirrorTree Pearson linear correlations are derived. Tree-to-tree correlations passing a significance cutoff of ten to the power of negative five are assembled into a proteome-wide coevolutionary matrix.

The framework subsequently refines raw phylogenetic tree similarities into an optimized coevolution measure through two advanced analytical steps. In the second stage, ContextMirror 2.0 computes profile-profile Pearson correlations across the entire matrix, evaluating whether two candidate proteins display matching correlation patterns across all other proteins in the proteome. Because analyzing paired alignments encompassing more than one thousand species drastically increases degrees of freedom and inflates raw significance values, the developers applied a strict Bonferroni correction, setting the statistical threshold to ten to the power of negative eleven to suppress spurious correlations. In the third stage, partial correlation coefficients are computed to assess whether the evolutionary association between two proteins persists after controlling for the profiles of third proteins at a threshold of ten to the power of negative six. To counteract score inflation, the authors implemented a linear penalization function that reduces the interaction score for candidates influenced by fifty or fewer third proteins, effectively deprioritizing isolated statistical artifacts.

Applying ContextMirror 2.0 to the reference proteomes of Escherichia coli K12 MG1655, Salmonella enterica serovar Typhimurium LT2, and Staphylococcus aureus NCTC 8325 established robust pairwise predictive accuracy. In Escherichia coli, the pipeline scored 505,294 candidate interactions from millions of potential pairings, recovering approximately twenty percent of the ground-truth set recorded in the STRING database at an overall precision of roughly twenty percent. This recovery matches performance reported for residue-level direct coupling analysis methods while avoiding their intensive computational demands. Among the top one thousand ranked predictions for each species, precision reached approximately forty percent, representing an eight-point-five-fold enrichment over baseline chance rates of 4.9 percent in Escherichia coli, 4.0 percent in Salmonella enterica, and 3.9 percent in Staphylococcus aureus.

The authors conducted extensive statistical controls to confirm that the predictive power of ContextMirror 2.0 genuinely reflects evolutionary coevolution rather than genomic proximity or alignment depth. Quantifying alignment diversity through effective sequence counts demonstrated that alignment depth accounted for only thirteen to sixteen percent of the variance in interaction scores. Furthermore, logistic regression models confirmed that the interaction score retained strong predictive significance after controlling for both the count of matching species and intergenic distance. Because bacterial genes residing in shared operons are frequently co-transcribed, the researchers also examined whether operon architecture biased predictions. While intra-operon gene pairs expectedly earned high interaction scores, the ranked interactome was overwhelmingly dominated by inter-operon predictions. At an interaction score threshold of 1.8 or higher, the framework identified 2,202 experimentally validated interactions connecting distinct operons alongside 53 validated same-operon pairs, proving that the algorithm detects functional cooperation spanning independent transcriptional units across the bacterial chromosome.

To contextualize performance against leading artificial intelligence and structural modeling techniques, the researchers benchmarked ContextMirror 2.0 against D-Script and RoseTTAFold2-PPI across four experimentally defined bacterial complexes: the Nuo and ATP synthase assemblies in Escherichia coli and the Dpp and Liv transport systems in Salmonella enterica. When evaluated across the candidate space retained by the coevolutionary filter, ContextMirror 2.0 achieved superior discrimination between validated and unsupported pairs, delivering an area under the receiver operating characteristic curve of 0.94 on the Nuo complex compared to 0.75 for D-Script and 0.52 for RoseTTAFold2-PPI. In parallel, structural evaluation of sixty predicted dimers using AlphaFold-Multimer and the pDockQ2 metric revealed that high-scoring pairs were three times more likely to yield high-confidence structural complexes than low-scoring pairs. Crucially, many validated high-scoring interactions did not produce stable physical complexes, underscoring that ContextMirror 2.0 successfully detects functional networks and transient enzymatic partnerships that evade purely physical docking predictors.

Beyond pairwise interaction scoring, ContextMirror 2.0 translates proteome-scale predictions into multispecies networks to detect functional protein communities. By mapping homologous proteins through reciprocal best-hit relationships, the investigators quantified network conservation, finding that the two closely related Gram-negative bacteria Escherichia coli and Salmonella enterica shared 70.0 percent of their predicted interactions, whereas the phylogenetically distant Gram-positive Staphylococcus aureus shared only 40.7 percent and 48.3 percent, respectively. Applying the Louvain community-detection algorithm to the combined networks yielded tightly clustered functional modules where intra-community precision consistently exceeded pairwise precision, reaching 36.9 percent in Escherichia coli at strict score thresholds. This modular network analysis uncovered 610 high-confidence interactions shared by both Gram-negative species that currently lack annotation in ground-truth repositories, highlighting strong candidates for conserved biological functions.

Detailed examination of the resulting communities demonstrated the platform’s capacity to illuminate distinct species-specific physiological adaptations. For example, ContextMirror 2.0 detected the phn gene community responsible for converting methylphosphonate into phosphate and methane in Escherichia coli, while reflecting the structural divergence of this complex in Salmonella enterica. In Salmonella enterica, the algorithm identified an exclusive community containing three electron transfer flavoproteins—STM0855, STM0856, and STM0858—linked to the glutathione S-transferase YibF and the putative O-antigen modification enzyme RfbI, suggesting an evolutionary integration between energy metabolism, redox homeostasis, and lipopolysaccharide-mediated immune evasion. Similarly, analysis of the conserved yejABEF Microcin C transporter complex revealed a species-specific interaction in Salmonella enterica involving STM2551, a homolog of the nickel-cobalt efflux pump RcnA, pointing to lineage-specific transport specializations. Furthermore, analysis of the xyl operon recovered the presence of the xylose isomerase xylA across both species while capturing the absence of the xylFGH transporter machinery in Salmonella enterica.

Finally, the research team demonstrated how multispecies coevolutionary networks facilitate functional annotation of uncharacterized proteins in understudied pathogens like Staphylococcus aureus. Within the peptidoglycan synthesis pathway, the algorithm detected interactions between the uncharacterized protein WP_000160904.1 and cell wall enzymes MurA and MurF, corroborating independent records identifying the sequence as a multispecies MurG glycosyltransferase. In purine biosynthesis, ContextMirror 2.0 correctly assigned the Gram-positive-specific components PurL, PurQ, and PurS of the FGAM synthase complex exclusively to Staphylococcus aureus. Investigating fatty acid biosynthesis operons further uncovered several uncharacterized proteins interacting with FabF and FabG, including candidates corresponding to the fatty acid kinase FakA, the aminopeptidase PepP, the adhesin YpjA, and the ribosomal maturase RnmV. Although the requirement for at least one thousand shared orthologs limits application to protein families with deep evolutionary records, ContextMirror 2.0 establishes an interpretable, scalable foundation for deciphering cellular organization across diverse bacterial proteomes.

Subject of Research: Unsupervised coevolutionary modeling of species-specific bacterial protein-protein interactomes

Article Title: Sequence-based coevolutionary prediction of species-specific interactomes

Article References: Fernandez-Martin, M., Pontes, C., Ruiz-Serra, V., & Valencia, A. (2026). Sequence-based coevolutionary prediction of species-specific interactomes. PLOS Computational Biology, 22(11), e1014839. https://doi.org/10.1371/journal.pcbi.1014839

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014839

Keywords: Computational Biology, Protein Interactions, Coevolution, Interactome, ContextMirror, Bacterial Proteomes, Phylogenetics, Structural Biology, Sequence-based, coevolutionary, prediction, species-specific