Sergei Pond @sergeilkp.bsky.social · 14/10/2026HyPhy is getting an upgrade to sleek, clean, modern C++-20 code and all kinds of other things. Much faster, cleaner, and more robust. Cool terminal output too! See github.com/veg/hyphy-3 (much more to come with VERY exciting new developments).github.comGitHub - veg/hyphy-3: HyPhy v3. We are not in 1997 anymoreHyPhy v3. We are not in 1997 anymore. Contribute to veg/hyphy-3 development by creating an account on GitHub. 121
Sergei Pond @sergeilkp.bsky.social · 11/09/2026Runs 100% in your browser (WebGPU/WASM, zero install), Python package: 🧪 App: primaeon.org 📊 Slides: veg.github.io/HyphAeon/ 📄 bioRxiv: www.biorxiv.org/content/10.6... 💻 Code: github.com/veg/HyphAeonprimaeon.orgPrimAeonDrop a codon alignment. PrimAeon runs HyphAeon, the neural surrogate for MEME: episodic selection per site, a gene-level verdict, epistatic sectors and a digital deep mutational scan, in your browser,... 031
Sergei Pond @sergeilkp.bsky.social · 11/09/20266. What comes next? Because it learns evolutionary "geometry" directly, it serves as a foundation model: we have promising results on molecular dating, recombination breakpoint detection, and automated phylogenetic placement (more to come) 100
Sergei Pond @sergeilkp.bsky.social · 11/09/20265. Temporal selection profiling. Screened ~3K @nextstrain SARS-CoV-2 genomes in minutes. It tracks selection velocity in real time, catching selective sweeps (D614G, JN.1) months before frequency peaks while suppressing neutral hitchhikers. 100
Sergei Pond @sergeilkp.bsky.social · 11/09/20263. Built-in error masking. Previously we needed BUSTED-E to stop sequencing errors from faking selection. HyphAeon masks them on the fly 4. (Crude but useful) Digital DMS (dDMS). Evaluates 19 AA variants at 0.24ms/mut; 5,000 TEM-1 mutations in 1.2s. 101
Sergei Pond @sergeilkp.bsky.social · 11/09/20261. FAST (1,000x–10,000x speedup). Screening OrthoMaM (15,868 genes, 190 species) takes ~2h on a laptop (~1s/gene). 2. Epistasis for free. Untrained on structure, it recovers co-evolving networks in seconds (e.g. HIV-1 PR modules vs weeks of Potts models). 110
Sergei Pond @sergeilkp.bsky.social · 11/09/2026I've spent ~30 years writing HyPhy to study selection & evolution. That era is coming to a close. HyPhy is passing the torch to HyphAeon: a ~2M parameter geometric transformer running in the browser that does things we never built it to do. @nclark.bsky.social @brcanalytics.bsky.social 🧵👇 11514