Scientists have developed a groundbreaking AI model, named Evo, which is trained on thousands of bacterial genomes to predict and generate entirely new DNA sequences. Instead of just working at the protein level, Evo learns directly from nucleotide-level patterns, treating DNA much like a language model treats text.
By prompting Evo with fragments of known genes, researchers observed that it could reconstruct the missing portions very accurately — and even produce functional sequences when parts of the genome were removed. Some of Evo’s generated sequences turned out to encode proteins or RNAs that were previously unknown — a remarkable sign that the model isn’t just copying, but genuinely creating.
In one experiment, Evo produced novel anti-toxin genes when given a toxin prompt, and when these were tested in bacteria, a few actually succeeded in neutralizing the toxin. Even more impressively, Evo was able to generate CRISPR inhibitor genes that bear little to no similarity with any known proteins, yet still work in functional tests.
This development is particularly exciting because it brings us closer to designing biology from the ground up — at the level of the genome. While the current work is focused on bacteria, the conceptual leap could eventually enable more informed bioengineering, synthetic biology, and novel therapeutics.
