tech
AI trained on bacterial genomes produces never-before-seen proteins
Genes with related functions cluster together, and the AI uses that.
TL;DR
- AI systems have shown success in predicting protein structure and designing functional proteins.
- Changes in biology occur at the nucleic acid level before impacting proteins.
- Training an AI system (Evo) on bacterial genomes can help predict novel functional proteins.
- Evo learns by predicting the next base in a DNA sequence, similar to large language models.
- The system can complete known gene sequences and restore deleted genes within functional clusters.
- Evo generated novel antitoxins that successfully rescued bacterial growth despite having minimal similarity to known antitoxins.
- The AI also generated functional RNAs and CRISPR inhibitors that were entirely new.
- Evo appears capable of outputting entirely new yet functional proteins without explicit consideration of protein structure.
- The approach might not be directly applicable to more complex genomes like vertebrates.
- This work brings the challenge of finding functional proteins down to the nucleic acid level.