First paragraph (55 words) Evo 2, an open-source AI system, has been trained on 8.8 trillion DNA base pairs from bacteria, archaea, and eukaryotes. Unlike its predecessor, this version identifies complex genome features like splice sites and regulatory sequences across all life domains.

The model’s architecture, StripedHyena 2, processes sequences in 8,000-base chunks, then analyzes million-base segments to detect large-scale patterns. Researchers emphasized its ability to recognize evolutionary constraints, enabling zero-shot predictions without task-specific training. OpenGenome2 dataset inclusion of viruses (excluding eukaryotic-infecting pathogens) aimed to mitigate misuse risks. 40 billion parameters in the full model version allow nuanced feature detection, surpassing specialized tools in eukaryotic genome analysis.