AI can now model and design the genetic code for all domains of life with Evo 2 | Arc Institute

Abstraction: Largest open-source AI biology model trained on 9.3 trillion nucleotides

Key points:

• Evo 2 is trained on 9.3 trillion nucleotides from 128,000+ whole genomes across all domains of life (bacteria, archaea, phage, eukaryotes including humans and plants)
• Uses StripedHyena 2 architecture enabling a 1-million-nucleotide context window; trained on 2,000+ NVIDIA H100 GPUs via NVIDIA DGX Cloud on AWS; Greg Brockman (OpenAI co-founder) contributed during sabbatical
• Achieves >90% accuracy predicting pathogenic vs. benign BRCA1 variants; can design synthetic genomes as long as simple bacterial genomes
• Released fully open-source (weights, training code, training data) — the largest-scale fully open AI biology model to date; integrated into NVIDIA BioNeMo framework
• Applications include cell-type-specific genetic element design for targeted gene therapy and identification of disease-causing mutations
• Human pathogen sequences excluded from training data as a biosafety measure; Goodfire mechanistic interpretability visualizer available

Connections: Arc Institute · Nvidia · Foundation Models · Generative Biology · Drug Discovery

Source: https://arcinstitute.org/news/blog/evo2