Fox skips tree building for host-symbiont coevolution: reading paired alignments directly, it runs over three orders of magnitude faster than an ABC baseline

Coevolution event frequencies can now be read directly from a pair of host and symbiont multiple sequence alignments and their mapping, without first reconstructing a host tree and reconciling the symbiont tree against it. Earlier statistical methods, represented by Coala and AmoCoala, estimate the four event probabilities with approximate Bayesian computation but require a host tree, and sometimes a symbiont tree, as input, so tree-inference error and uncertainty enter the subsequent estimate. On the held-out test set from the training simulator AsymmeTree, Fox reached a mean absolute error of 0.016–0.050 and R² of 0.45–0.85, with absolute bias no larger than 0.010.

Moved to an independent simulator it was not trained on, Fox's error roughly doubles but still tracks the true frequencies and clearly beats the tree-based ABC baseline. On 100 datasets generated by treeducken, Fox's mean absolute error was 0.060, against 0.030 on the AsymmeTree test set; AmoCoala reached a mean absolute error of 0.131 on the 92 datasets it completed. On those 92 datasets Fox's error for cospeciation, host switching and duplication was 45–76% lower than AmoCoala's, with similar accuracy for losses; AmoCoala over-estimated host switching by 0.147 and under-estimated duplication by 0.142, and its estimates had correlation with the truth no larger than 0.20 in absolute value.

Fox turns each host-symbiont association into a token so that information from the two alignments is exchanged at the level of associations, letting the coevolutionary signal be read straight from the sequences. A shared-weight alignment encoder processes the host and symbiont alignments separately, after which attention runs over association tokens, biased by host distance, symbiont distance and whether two associations share a host. On the 100 held-out AsymmeTree test datasets, shuffling the host-symbiont links dropped the predicted cospeciation frequency in every dataset, from 0.63 to 0.23 on average, and on the cockroach data from 0.61 to 0.39.

The evaluation covers a held-out set from the training simulator AsymmeTree, 100 datasets from the independent simulator treeducken, and one empirical dataset of 52 cockroach-Blattabacterium pairs; the model was trained under the LG+Γ+invariable-sites sequence-evolution model, with at most 50 host leaves, at most 142 symbiont leaves and 250 alignment columns, and it still runs on inputs beyond that range while issuing a warning.

On treeducken data Fox's error is about twice that on the training simulator, and for losses R² is −0.16 with a correlation of 0.37, which the authors suggest may reflect treeducken producing more losses through a different mechanism. On the empirical data both Fox and AmoCoala rank cospeciation first, but Fox assigns 70.5% and AmoCoala 54.3%, the latter with a standard deviation of 0.25 across its 50 accepted parameter vectors. Training took about 300 GPU-hours on 4×H100 GPUs, a one-time cost that ABC methods do not require; accuracy is essentially stable when alignments are shortened from 250 to 150 columns, degrades at 100, and at 50 columns cospeciation, host switching and losses become unreliable.