Evo 2, Open-Source AI Model for Generative Genomics, Validated and Published in Nature
Evo 2, a large open-source genomic foundation model designed to analyze and generate DNA sequences has been reported in Nature with experimental validation. The work demonstrates that AI-generated DNA sequences can alter chromatin accessibility in living cells, suggesting that generative genomic models may eventually be used to design regulatory DNA elements with predictable biological effects.
In experiments, sequences designed by the model were synthesized, inserted into mouse and human cells, and then tested with standard chromatin accessibility assays that map how “open” or “closed” different DNA regions are; the AI-generated sequences produced measurable changes in accessibility patterns.
Because chromatin accessibility is widely used as a proxy for regulatory activity and gene control, this result suggests that such models could eventually be used as tools to design regulatory DNA elements with more predictable effects on gene activity, moving from using AI to merely analyze genetic data toward using it to propose testable, functional edits to the genome’s regulatory code.
Evo 2 is a 40-billion-parameter genomic foundation model developed to predict the functional impact of genetic variants and generate DNA sequences. The system operates at single-nucleotide resolution and can analyze sequence contexts up to 1 million base pairs in length.
The project involved collaboration between Arc, Stanford University, NVIDIA, and additional research contributors.
The model covers both coding and noncoding regions of the genome and is trained to operate across organisms across all domains of life, ranging from bacteria to humans using a single set of model weights, without additional fine-tuning.
Both Evo 1 & Evo 2 are available on GitHub and HuggingFace.
Evo models capabilities
Since its open-source release one year ago, external research groups have applied Evo 2 to several biological questions, including prediction of Alzheimer’s disease–associated variant effects, classification of genetic variants in domesticated animal species, and modeling aspects of 3D genome structure.
Evo 2 can also perform in-context learning, meaning it can infer patterns and make predictions directly from examples in the prompt, a capability previously associated mainly with large language models.
Additionally, a functional bacteriophage designed by Evo 2 was synthesized and experimentally validated, showing that AI-generated DNA sequences can produce a working biological organism. The designed virus was able to infect and suppress its intended bacterial host while leaving unrelated bacteria unaffected, a property known as host specificity that is important for potential phage therapies against antibiotic-resistant infections.
Now, future work with Evo 2 will test whether Evo-generated DNA can be inserted at specific locations in the genome using genome engineering tools such as programmable bridge recombinases, which can rearrange very large DNA segments in human cells. The researchers also plan to expand “lab-in-the-loop” experiments, where AI designs sequences and laboratory results are used to refine the model.
Topic: AI in Bio
