PrimAeon: selection analysis in seconds, in your browser
The HyphAeon team
PrimAeon is live: you can now paste a codon alignment into a browser tab and watch a selection report stream back in seconds. No account, no install, no upload — the inference runs entirely on your device.
Why this is fast enough to matter
PrimAeon is powered by HyphAeon, a phylogenetic foundation model that acts as a neural surrogate for HyPhy's likelihood methods. The headline number from the preprint is speed: site-level episodic selection inference comparable to MEME at 100×–1,100× the speed of maximum-likelihood estimation. Analyses that used to justify a coffee break — or a cluster job — now finish before the page has finished scrolling.
That speed changes what the analysis is for. A selection scan stops being a deliberate, budgeted step and becomes something you can do speculatively: screen every gene in an alignment set, sanity-check a hypothesis before committing to a full study, or triage an outbreak sequence the moment it arrives.
What you get
One alignment in, one report out:
- Site-level selection — the neural MEME surrogate
- Co-selection networks — which sites evolve together
- Digital deep mutational scanning — predicted effect of every substitution
- Phenotype association — sites linked to a trait you describe (a curated preset, a foreground list, or a table)
Because these results are generated by a neural surrogate model, every report is
flagged is_surrogate and includes a direct path to verify the findings using
exact maximum-likelihood methods on Datamonkey.
A tree is optional. Supply one with branch lengths and it is used as given; without one, PrimAeon works tree-free from pairwise TN93 distances.
Private by construction
Because the models ship as ONNX graphs running on ONNX Runtime's WebAssembly backend — the analysis pipeline itself ported to JavaScript — your sequences never leave the browser tab. For sensitive data — pre-publication genomes, clinical cohorts, surveillance data under embargo — that removes the upload question entirely rather than merely answering it.
The other half: ChronAeon
The same foundation model powers ChronAeon, the "radar" to HyphAeon's
"microscope": molecular clock dating, R₀/Rₜ phylodynamics, phylogeography, and
streaming genomic triage — analyses that reproduce landmark benchmarks (the
1931 HIV-1 group M origin date, H5N1 dispersal routes, 2009 H1N1 growth rates)
in seconds rather than MCMC hours. ChronAeon is Python today
(pip install chronaeon); see the docs.
PrimAeon itself also has a dated route — /time
fits a molecular clock and runs temporal selection in the browser, BEAST XML
included.
HyphAeon is also wrapped for Galaxy — installed on usegalaxy.eu, with ChronAeon wrappers in progress — useful when you want web-based batch runs, or to customize an analysis beyond what PrimAeon's one-report interface offers.
Try it, then dig in
- primaeon.org — run a scan now
- Choosing between PrimAeon, HyphAeon, ChronAeon, and HyPhy — our Learn guide to the ecosystem
- HyphAeon on GitHub and the interactive overview
- bioRxiv preprint · model weights on Hugging Face
HyphAeon is supported in part by BRC Analytics and the NIH/NIAID.