Loading…
Phylogeny-informed transformer that detects episodic diversifying selection in codon alignments, amortizing maximum-likelihood dN/dS tests.
No providers recorded yet. Browse all providers
A gene under episodic positive selection may reveal it on three branches out of seven hundred, at a handful of codons. Maximum-likelihood codon models such as MEME find those sites by re-fitting a substitution matrix site by site and branch by branch, so a proteome-wide scan over a dense mammalian tree costs tens of thousands of CPU core-hours — and comparative genomics resorts to subsampling taxa or collapsing the phylogeny into a whole-tree summary, the two moves most likely to dilute a transient burst away.
HyphAeon replaces that per-locus optimization with a single forward pass. Built at the Institute for Genomics and Evolutionary Medicine at Temple University — the group behind the HyPhy and Datamonkey toolkits — with five collaborating institutions, it reads a codon alignment alongside its phylogeny and emits, for every column, the likelihood-ratio statistic MEME would have produced. The tree is not metadata for a preprocessing step: pairwise patristic distances become 4D classical-MDS coordinates applied inside attention as rotary phase rotations (Tree-RoPE), and a continuous-time Markov distance-decay prior on the attention logits makes two species attend to each other in proportion to how recently they diverged.
Because those coordinates depend on divergence rather than a fixed branching graph, the model can run tree-free, estimating distances from the alignment itself; and because the backbone yields a permutation-invariant vector per column, one frozen checkpoint drives four further analyses without retraining.
The backbone, PhyloAxialTransformer, is a six-layer axial transformer with 12 attention heads and embedding dimension 384, attending strictly across the species axis; codon and amino-acid tokenizations occupy separate 192-dimensional tracks to insulate selection inference from synonymous rate variation. A learnable [ROOT] token at the MDS origin pools each column, and a 16-threshold rank-consistent ordinal regression head decodes selection intensity, a choice forced by likelihood-ratio statistics that are roughly 95% zero-inflated. The transformer holds 1.79M parameters and the site bottleneck plus ordinal head 98.8k, for a 1.91M-parameter backbone; the full multi-task suite totals 2.46M. Pretraining used 17,186 mammalian coding alignments from TOGA — 742 species, 9.77 × 10⁶ codons — each paired with a subtree pruned from the reference mammalian species tree.
Against MEME across episodic burst regimes it reaches ROC-AUC up to 0.942 (mean 0.659), precision-recall lift up to 25.8× (mean 7.7×) and rank concordance up to ρ = 0.983, running more than 1,000× faster per locus and more than 10,000× faster at proteome scale. On a replication set of 84 published alignments from nine studies — 43,302 codons, up to 476 taxa — the aggregate is ROC-AUC 0.914 and positive predictive value 50.6% at a 0.191% false-positive rate, with 25,696 seconds of HyPhy compute cut to 78.87 seconds on CPU.
The immediate use is proteome-wide selection scanning at a scale that previously needed a cluster allocation: screening a clade's gene complement for adaptive sites, triaging candidates before committing MEME or FEL runs to the survivors, and catching alignment errors before they reach publication as positive-selection hits. The epistasis and phenotype heads serve narrower questions — co-selection networks within a gene family, or sites linked to a convergent trait such as echolocation — while the viral checkpoint supports surveillance of transmission-chain sweeps.
HyphAeon belongs, with GPN-Star, to a line of work writing phylogeny into the architecture rather than treating aligned species as interchangeable rows — but where those models learn constraint from scratch, it amortizes one well-characterized statistical test, inheriting its assumptions and its interpretation. That is also the limit: it approximates MEME rather than independently testing for selection, and is evaluated against MEME, not ground truth. Precision degrades on large, diverse gene families, falling to 44.8% positive predictive value on a Siglec and C-type lectin panel, and an aggregate PR-AUC of 0.286 reflects how rare positive sites are. Alignments must be in frame, branch lengths meaningful, and non-coding or recombinant sequence is unassessed. Treated as its authors present it — a fast screen whose hits warrant confirmatory likelihood analysis — it turns selection scanning from a batch job into an interactive step. Code, weights and training scripts are MIT-licensed, and the work is a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.