Fox: Predicting rates of host-symbiont co-evolution events with association-aware attention

Abstract

Motivation: Hosts and their symbionts share a history of associations in which lineages cospeciate, switch hosts, duplicate, or are lost. Quantifying the relative frequencies of these four events is a central goal of cophylogenetics, but existing tools typically require a fully resolved, correctly rooted host tree, often alongside a symbiont tree, propagating tree-inference errors into the reconciliation and scaling poorly to large systems. Results: We introduce Fox, a transformer neural network architecture that infers the relative frequencies of the four cophylogenetic events directly from a pair of host and symbiont multiple sequence alignments (MSAs) and their sequence mapping, with no tree-reconstruction or reconciliation step. Fox is trained under a simulation-based inference scheme: it learns event frequencies from alignments generated by a coevolutionary process, paying its computational cost once at training time and processing any new system in a single forward pass. On held-out data from its training simulator, Fox recovers event frequencies accurately and with near-zero bias. On additional synthetic data from an independent cophylogeny simulator, its error roughly doubled, but is still very low allowing to track the true frequencies. Most importantly, when compared to the Approximate Bayesian Computation (ABC)-based method AmoCoala on these data, it reduces mean absolute error by 45-76% for cospeciations, host switches and duplications and shows similar error for losses, while running more than three orders of magnitude faster. Applied to the empirical cockroach-Blattabacterium system, Fox recovers the expected cospeciation-dominated profile, showing that behaviour learned on simulated data transfers to real sequences. These results confirm that the coevolutionary signal usually extracted from trees can be read directly, accurately, and fast in the alignments themselves