Large genome model: Open source AI trained on trillions of bases

In late 2025, the spotlight was on an innovative AI system named Evo, designed to analyze extensive collections of bacterial genomes. This groundbreaking system showcased its ability to accurately predict the next gene in a sequence or even propose entirely new proteins. The success of Evo stemmed from the relatively simple organization of bacterial genomes, where related genes are often found clustered together. However, the complexity of more advanced organisms posed a challenge for the initial model. Recognizing this, the creators of Evo took on the task of developing Evo 2, an open-source AI that now encompasses genomic data from all three life domains: bacteria, archaea, and eukaryotes. After training on trillions of DNA base pairs, Evo 2 has successfully developed internal representations that capture key elements of complex genomes, including intricate regulatory DNA and splice sites that are often elusive to human researchers. Bacterial genomes operate under clear and efficient principles, where genes responsible for specific functions are neatly arranged without interruptions in their coding sequence. In contrast, eukaryotic genomes exhibit a more convoluted structure. Here, coding regions are often interrupted by non-coding introns, and regulatory sequences can be dispersed across vast stretches of DNA. Evo 2's advanced capabilities allow it to navigate this complexity, identifying critical features that may be overlooked by traditional methods. The AI not only highlights the edges of introns but also pinpoints binding sites for regulatory proteins, showcasing the potential for significant advancements in genetic research and understanding the function of what is often labeled as 'junk DNA'—a term that encompasses inactive viruses and degraded genes. This leap forward could pave the way for breakthroughs in genetics and biotechnology, making Evo 2 a vital tool for researchers worldwide.