Reconstructing protein interactions across time using phylogeny-aware graph neural networks

Abstract Motivation Genes which are involved in the same biological processes tend to co-evolve. Thus, metabolic pathways, protein complexes, and other kinds of protein-protein interactions can be inferred by looking for correlated patterns of gene retention and loss across the tree of life—a technique called phylogenetic profiling. Recent methodological developments on phylogenetic profiling have focused on scalability improvements to take advantage of the rapidly accumulating genomic data. However, state-of-the-art methods assume that the correlation resulting from co-evolving proteins is uniform across all species considered. This is reasonable for interactions already present at the root of the species considered, but less so for ones that emerge in more recent lineages. To address this challenge and take advantage of recent developments in deep learning methods, we introduce a phylogenetic profiling method which processes large gene co-phylogenies using neural networks. Results We show that post-processing conventional phylogenetic profiles using deep neural networks can improve predictions, but requires onerous training on specific phylogenies. Overcoming this limitation by taking the topology of the species tree as an input, Graph Neural Networks are shown to outperform all other methods when interaction detection is not centered on just one species of interest, while also predicting when interactions appeared and in which taxa they are present. Conclusion Graph Neural Networks constitute a promising new approach for phylogenetic profiling. Our work is a first foray into “dynamic phylogenetic profiling”—the reconstruction of pairwise protein interaction across time. Availability All of the code is available on the project Git at https://github.com/DessimozLab/HogProf/tree/master/pyprofiler/notebooks/Graphnet . Datasets used are hosted at http://humap2.proteincomplexes.org/download and https://string-db.org/cgi/download . Contact dmoi@unil.ch