Genos: a human-centric genomic foundation model.

Area of Science:

• Genomics
• Artificial Intelligence
• Bioinformatics

Background:

• Existing foundation models struggle with ultra-long human genomic sequences and population diversity.
• Current models lack human-specific representation and efficient clinical inference capabilities.

Purpose of the Study:

• Introduce Genos, a human-centric genomic foundation model engineered for million-basepair sequence modeling.
• Address limitations in current models regarding sequence length, population diversity, and clinical utility.

Main Methods:

• Developed Genos (Genos-1.2B/Genos-10B) utilizing a mixture of experts structure optimized for a 1-Mb context.
• Trained on high-quality human de novo assemblies from diverse global populations (Human Pangenome Reference Consortium, Human Genome Structural Variation Consortium).
• Implemented optimization strategies for training stability, computational efficiency, and cost reduction.

Main Results:

• Genos achieves single-nucleotide resolution analysis and simulates noncoding variation effects on RNA expression.
• Outperforms state-of-the-art models on critical human genomics benchmarks.
• Demonstrates robust omics-text cross-modal diagnostic capabilities.

Conclusions:

• Genos provides a blueprint for next-generation high-efficiency genomic foundation models.
• Model weights, inference code, and documentation are publicly available on GitHub and Hugging Face Hub.
• Establishes a new performance benchmark for genomic foundation models.