Genome modelling and design across all domains of life with Evo 2 (FM) (www.nature.com)

Evo 2, a groundbreaking biological foundation model, has been introduced, trained on an extensive 9 trillion DNA base pairs sourced from a diverse genomic atlas encompassing all domains of life. Unlike its predecessors, Evo 2 can predict the functional impacts of genetic variations without requiring task-specific fine-tuning. It incorporates a novel architecture called StripedHyena 2, which enhances efficiency and throughput while learning long-range biological relationships within sequences. The model's architecture allows it to process genomic data with a context length of up to 1 million tokens, a significant leap in handling the vast complexity of eukaryotic genomes. The significance of Evo 2 lies in its ability to bridge the gap in understanding genetic complexity, facilitating improved predictions and designs across various biological systems. Its zero-shot prediction capability across prokaryotic and eukaryotic sequences positions it as a vital tool for researchers in the AI/ML community, enabling advances in tasks such as mutational effect prediction and the design of new biological systems. With its open-source release, including model parameters and training datasets, Evo 2 is set to accelerate the exploration of genetic variations and their implications, making a substantial impact on the future of computational biology and bioengineering.