Efficient biological design via genomic in-context learning

Evo 2 shows in-context learning in genome language modeling, generating realistic metal-binding proteins, de novo antibodies, protein-RNA systems, and entire bacteriophage genomes using as few as four prompt examples.

• Why it matters: Biological sequence design often requires task-specific training, yet ICL could enable new designs from prompts alone, improving efficiency and flexibility without additional model updates.
• What they did: The study uses Evo 2 with no additional training, demonstrating PhiX174-like phage designs that surpass previously fine-tuned models on computational measures, plus MS2 and M13 designs via different phage prompts.
• The result: Iteratively updating prompts with newly generated phage genomes (genomic chain-of-thought) improves diversity while maintaining quality metrics, and an inference engine accelerates generation up to 35-fold with megabase context.