Evo 2 designs enhancers from a few dozen examples, and six of nine label mouse brain neurons in place

Evo 2 can now generate regulatory sequences with a specified biological function from scratch using only a few examples placed in its prompt, without first fine-tuning on a large labeled dataset. Previously, making Evo 2 generate a given sequence class meant fine-tuning it; conditioning it to generate a phage, for instance, required training on 15,000 natural examples, while many useful natural and synthetic sequences exist in only tens of copies and were nearly impossible to train on. With same-class sequences such as 5' UTRs, IRESs, promoters and enhancers concatenated in the prompt, novel sequences gained higher predicted performance in each class as more examples were supplied; from fewer than 100 in-context examples, Evo 2 generated functional sequences across five regulatory modalities with 30-50% yield and under 60% sequence similarity.

The capability was not built into Evo 2 but appears to emerge from model scale and the natural genomic data it was trained on, and it shows internal signatures resembling in-context learning in language models. Natural genomes do not contain explicitly delineated regulatory elements, yet the model picked up a pattern not seen in its training data. Candidate induction heads were isolated among Evo 2's attention heads, and ablating one actually affected sequence design performance while ablating another random head in the same layer caused essentially no change; a categorical Jacobian showed high-influence nucleotides concentrated in a single example when few were given and spread across the prompt as more examples entered context.

Repurposing sequence-to-function models as reward models lets a reinforcement learning loop improve generated sequence quality without retraining Evo 2 or large-scale fine-tuning. In-context learning already produced sequences, but the team sought to improve both efficiency and quality. Reinforcement took under 20 minutes for easier tasks and still hours for harder ones; post-training on a human promoter dataset also boosted enhancer design performance, and post-training on two cell types while holding out a third raised predicted specificity in all three.

The designs were tested directly in the living mouse brain, and two enhancers matched or exceeded the specificity of the current most specific synthetic enhancer for layer 5 extratelencephalic neurons. Cell-type-specific enhancers that behave as intended in animals are extremely challenging to design, and no large-scale predictive model existed for scoring whether a DNA sequence might work in a mouse. AlphaGenome was independently fine-tuned as a reward and fine-tuned Enformer and Borzoi evaluated designs, together predicting chromatin accessibility across 107 neuron types in the human brain; nine designs packaged into AAVs and delivered by retroorbital injection gave seven with detectable expression and six with spatially localized labeling, two labeling layer 5 extratelencephalic neurons at 73% and 95% specificity.

The demonstrated scope is in-context generation of regulatory sequences with Evo 2 and in vivo testing of enhancer designs delivered by AAV into the mouse brain, targeting layer 5 extratelencephalic neurons. The team states it plans to extend the work to generate specific and strong enhancers for more cell types in the brain.

Designs beyond enhancers, including novel proteins, CRISPR-Cas9 loci, epitope-specific antibodies and whole bacteriophage genomes, are computational and have not yet undergone experimental testing. The team also notes that designs made for a human brain cell type working in the mouse brain was surprising, suggesting some evolutionary conservation in the brain's regulatory codes.