Evo 2 is a massive-scale genomic foundation model released by the Arc Institute in the United States, in collaboration with Stanford University and NVIDIA. Similar to how large language models in the field of natural language processing learn human language systems to understand context and generate new text, this tool treats the blueprint of living organisms – deoxyribonucleic acid (DNA), ribonucleic acid (RNA), and protein sequences – as a single, vast language, analyzing and designing them. Unlike previous biological models that were limited to analyzing specific proteins or short gene sequences, Evo 2 is trained on over 9 trillion base sequence data points, encompassing all life forms on Earth, including bacteria and archaea. Architecturally, it adopts the StripedHyena 2 architecture, achieving computational efficiency that allows it to process vast genomic information of up to 1 million tokens at once, while maintaining precision at the single-base level. Existing language models in the biological field have a limitation in that the computational complexity of the attention mechanism increases quadratically with the sequence length, allowing them to handle only relatively short contexts, such as protein structure prediction or gene fragment analysis. This made it difficult to track the vast non-coding genomic regions that regulate life phenomena or the long-range interactions between complex enhancers and promoters. Evo 2 replaces the computational bottleneck of traditional transformer architectures with a hybrid architecture of innovative implicit state space models and convolutions, offering a unique advantage, similar to magnifying the entire sequence of life with an ultra-high-resolution lens. This drastically reduces computational load while providing an unparalleled range of analysis, allowing researchers to understand the impact of even the slightest mutations in individual base sequences on the overall physiological function of an organism. In actual biotechnology research, researchers can use Evo 2 to completely transform the design paradigm of synthetic biology. For example, when a researcher wants to design the entire genome of a microorganism with a specific beneficial metabolic pathway, by inputting the alignment status of tens of thousands of base pairs of regulatory sequences and coding regions, Evo 2 can generate the entire transcriptional regulatory unit, including promoters and enhancers optimized for the given context, all at once. Furthermore, in the field of clinical research, it can be used as a predictive model that accurately predicts the probability of a single nucleotide polymorphism occurring in the non-coding region causing a specific disease by inputting rare disease genomic data. This replaces the process that molecular biology researchers had to repeat through trial and error in laboratory experiments for a long time with high-speed screening in a digital environment, shortening the research period for drug development and gene therapy design.

Evo 2 is a massive-scale genomic foundation model released by the Arc Institute in the United States, in collaboration with Stanford University and NVIDIA. This tool, similar to how large language models in the field of natural language processing learn human language systems to understand context and generate new text, treats the blueprints of living organisms – deoxyribonucleic acid (DNA), ribonucleic acid (RNA), and protein sequences – as a single, vast language, analyzing and designing them. Unlike previous biological models that were limited to analyzing specific proteins or short gene sequences, Evo 2 analyzes bacteria.

🔬 Genome-Scale Prediction of Gene Function and Variants: To explore disease-causing non-coding variants, input a human genome DNA sequence (10kbp or greater) into Evo 2 to calculate the fitness and transcriptional changes of single-nucleotide mutations within seconds, and utilize it for in silico functional change analysis.