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

Fox accurately infers the relative frequencies of four cophylogenetic host–symbiont events from host and symbiont MSAs, achieving near-zero bias on held-out simulated data.

• Why it matters: Existing cophylogenetic tools often depend on fully resolved, correctly rooted host trees and reconciliation, so tree-inference errors propagate and runtime scales poorly for large systems.
• What they did: Fox uses a transformer with association-aware attention, trained via simulation-based inference on alignments from a coevolutionary process, then predicts event frequencies for new systems in a single forward pass.
• The result: Compared with AmoCoala, Fox cuts mean absolute error by 45–76% for cospeciations, host switches, and duplications, runs over three orders of magnitude faster, and transfers to the cockroach–Blattabacterium system.

AI summary of the abstract · may contain errors · not medical advice · read the paper.