Sequence modeling and design from molecular to genome scale with Evo

Area of Science:

• Genomics
• Bioinformatics
• Synthetic Biology

Background:

• The genome encodes DNA, RNA, and proteins essential for organism function.
• Foundation models have shown success in language and vision, but genomic applications are emerging.

Purpose of the Study:

• To introduce Evo, a long-context genomic foundation model.
• To explore scaling laws for DNA sequences.
• To demonstrate Evo's capabilities in prediction and generation across molecular and genomic scales.

Main Methods:

• Trained Evo on millions of prokaryotic and phage genomes.
• Developed a frontier architecture for long-context genomic data.
• Evaluated Evo's zero-shot prediction and sequence generation capabilities.

Main Results:

• Evo generalizes across DNA, RNA, and proteins, achieving competitive zero-shot function prediction.
• Evo successfully generated functional CRISPR-Cas and transposon systems (protein-RNA and protein-DNA codesign).
• Evo learned mutation effects on fitness and generated large-scale genomic sequences with realistic architecture.

Conclusions:

• Evo represents a significant advancement in genomic foundation models.
• The model demonstrates novel capabilities in predicting biological functions and designing genetic systems.
• Evo has the potential to advance our understanding and control of biological systems at multiple scales.