AI trained on bacterial genomes produces never-before-seen proteins

TL;DR

AI systems have advanced in predicting protein structure and function, but previous efforts focused on proteins, not the nucleic acid level where biological changes originate.
A new AI model, Evo, trained on bacterial genomes, learns to link nucleotide patterns to genomic context.
Evo can predict missing gene sequences and restore deleted genes within functional clusters.
When prompted with a novel bacterial toxin, Evo generated antitoxins, half of which rescued toxicity, and two fully restored bacterial growth.
These novel antitoxins showed minimal similarity to known ones and appeared assembled from multiple protein parts.
Evo also successfully generated DNA encoding functional RNA inhibitors and novel CRISPR inhibitors.
The AI model generated 120 billion base pairs of AI-generated DNA from 1.7 million genes.
The approach may not work with more complex genomes due to gene clustering differences and intricate gene structures.

Continue reading the original article