The workflow to optimize genome sequences for particular environments. The GenSLM-2 model is optimized in the top loop, using previously trained fitness metrics and reward models. Outputs are validated via high-resolution methods such as simulations and high-fidelity surrogate models.
This project develops the first genome-scale foundation model capable of whole-chromosome analysis across all domains of life, leveraging DOE’s exascale computing and genomic datasets to advance AI-driven discovery in bioenergy, environmental science, and national security.
The emergence of genome-scale language models represents a transformative moment in computational biology, with AI-driven genomic discovery projected to generate $13.65 billion in market value by 2029 and to have profound implications for advancing DOE’s strategic missions in bioenergy, environmental remediation, and national security.
State-of-the-art DNA-based language models still suffer from fundamental limitations, including narrow taxonomic representation, restricted context windows, and computational bottlenecks that prevent effective modeling of environmental microbes essential for DOE applications and limit understanding of long-range regulatory interactions critical for engineering biological systems and chromosome-level modeling of evolutionary processes. This creates an unprecedented opportunity for DOE to leverage its unique strengths: diverse genomic datasets, exascale computing infrastructure, and a mission-driven focus on energy, environmental, and biosecurity applications.
This project delivers the first comprehensive genome foundation model capable of whole-chromosome analysis across all domains of life, establishes DOE as a global leader in genome-scale AI, and builds the computational infrastructure necessary for next-generation biological discovery while advancing DOE’s AI for Science initiative and demonstrating the transformative potential of exascale computing for scientific research.
