Arc Institute's Evo 2 designs mouse brain enhancers from 4-64 prompt examples, 2 of 9 hit layer 5 neurons
Two designs labeled layer 5 extratelencephalic neurons with 73% and 95% specificity in a bioRxiv preprint. Designs beyond enhancers are still computational only.
What happened
Arc Institute researchers prompted Evo 2, an AI model trained on DNA, with a few example sequences. It then wrote new enhancers, which are short DNA stretches that switch genes on in specific cell types. They tested nine in mouse brains.
Why it matters to you
Many useful DNA sequences have only a few dozen known examples, which made them hard for AI to learn. This method needs no large training set.
The takeaway
A genome AI can learn a design task from a few examples in its prompt, and early mouse tests worked.
What it means for you
Evo2-flash, an inference engine Arc says generates sequences up to 35x faster and runs on Google Colab GPUs, is on GitHub along with the regulatory design code.
Arc says the approach avoids large-scale screens and fine-tuning on thousands of labeled sequences, which could cut cost for enhancer design.