How does Evo 2 design genomes?

An AI trained on life’s DNA across the tree of life

Evo 2 is a large biological foundation model that learns statistical patterns across DNA from vast and diverse species. The model was trained on genomic data spanning more than 100,000 species and on the order of trillions of base pairs. By ingesting that breadth of sequence information, it identifies recurring motifs, regulatory signals and higher‑order structure that connect sequence to function.

What it can do

• Recognize patterns in genome sequence that correlate with biological properties.
• Suggest or generate short DNA sequences that are consistent with learned patterns.
• Help researchers explore sequence edits, candidate regulatory elements, or design ideas that would be time‑consuming to search by experiment alone.

Why this is significant

Evo 2 brings the foundation‑model approach used in language and vision into genomics. Instead of treating each species or gene as a completely separate problem, the model leverages shared statistical structure across many forms of life to generalize and propose plausible sequences. That can accelerate hypothesis generation, prioritize laboratory experiments, and shorten design cycles in synthetic biology.

Caveats and next steps

• Generated sequences are suggestions, not guarantees: biological function depends on cellular context, epigenetics and many interactions the model may not capture.
• Safety and ethics require attention: models that can produce functional biological sequences raise biosafety and dual‑use concerns.
• Further validation is needed: researchers must test outputs empirically, and the community needs agreed norms for responsible development.

Evo 2 represents a powerful new tool for genome modelling and design, but its promise depends on careful validation, transparent limits, and governance as it moves from research to practical applications.