Summary
Highlights
EVO 2 is a new AI that achieved 90% accuracy in predicting breast cancer from genetic mutations without prior medical training. It analyzed billions of DNA letters to distinguish harmless from deadly mutations, described as the largest biological AI model ever created and published in Nature in March 2026. Beyond cancer detection, EVO 2 can write functional DNA, including synthetic viruses to combat antibiotic-resistant bacteria and decipher the genome of extinct woolly mammoths, demonstrating a deep understanding of life's universal language.
Similar to language models like ChatGPT learning human language from text, EVO 2 learns the language of DNA. DNA is a code of four letters (A, T, C, G) that contains all instructions for building and operating living organisms. Researchers from the ARC Institute, Nvidia, Stanford, and Berkeley adapted large language model architecture to DNA, feeding EVO 2 9,300 billion DNA letters from 128,000 diverse species. This massive dataset and processing power allows EVO 2 to analyze up to a million DNA letters simultaneously, understanding global genetic contexts critical for biological function.
EVO 2 was trained without labels or explanations about genetic functions. It learned through evolution: functionally viable DNA sequences from living organisms were studied, while fatal mutations were naturally selected out over millions of years. This allows EVO 2 to deduce the purpose of sequences, identifying vital ones common across species and flagging unusual ones. It accurately predicts which single-letter DNA mutations would prevent protein production or disrupt cellular machinery. It even deduced different genetic code rules for specific unicellular organisms without explicit instruction, showcasing its deep understanding of biological grammar.
EVO 2 achieved over 90% accuracy in distinguishing benign from potentially deadly mutations in the BRCA1 gene, linked to breast and ovarian cancer, without any medical training. This performance is comparable to or better than specialized tools but without gene-specific training, opening doors for analyzing thousands of genes lacking clinical data. Other teams are already applying EVO 2 to predict genetic risks in Alzheimer's disease and evaluate animal species variations.
EVO 2 can generate DNA from scratch. It completed the human mitochondrial genome, producing 250 unique sequences of 16,000 letters, all functionally coherent. It also generated a complete bacterial genome of 580,000 functional DNA letters, with 70% of identified genes matching known protein families. Most notably, EVO 2 can design synthetic bacteriophages (viruses that infect bacteria, not humans) to combat antibiotic resistance, a major global health threat. In lab tests, 16 out of 285 AI-generated bacteriophages were viable, effectively killing target bacteria and rapidly overcoming bacterial resistance in E. coli strains.
The ability of EVO 2 to generate DNA raises concerns about the potential creation of deadly viruses. Researchers anticipated this by excluding human, animal, and plant pathogenic viruses from training data, making EVO 2 unable to produce coherent human virus DNA. However, the model's code and data are open-source, meaning someone with sufficient resources could theoretically retrain it with viral data. Despite these risks, the positive applications are immense, including faster mutation analysis in medicine, creating more resilient crops, and developing new treatments for resistant infections. EVO 2 is seen as a foundational technology, akin to an operating system kernel, that could revolutionize biology by enabling us not just to read, but to truly understand and write the code of life.