Evo2

Evo2 is a 7-billion parameter genomic foundation model trained on 800 billion tokens to predict and design DNA, RNA, and protein sequences at single-nucleotide resolution.

Developed by the Arc Institute, Evo2 leverages a StripedHyena architecture to process biological sequences across massive scales (up to 1-megabase contexts). It outperforms previous models by treating the genome as a unified language, enabling zero-shot prediction of mutation effects and the generative design of complex regulatory elements. By training on the 800-billion-token OpenGenome dataset, Evo2 captures multi-scale biological structures from molecular folding to whole-genome organization.

https://github.com/ArcInstitute/evo2