World’s first 10-bln-parameter genome foundation model launched

The BGI-Research and the Zhejiang Lab on Thursday jointly released Genos, the world's first deployable general foundational model for genomes with 10 billion parameters, which supports ultra-long context analysis of up to 1 million base pairs and achieves accurate identification at single-base resolution.

The human genome consists of 3 billion base pairs. The "Human Genome Project" decoded the sequence, but the specific functions of the bases in the sequence still require precise identification and analysis.

Most existing models are trained based on 1-2 reference genomes, making it difficult to reflect the diversity of human genetic resources. However, Genos integrates multiple authoritative public resources such as the Human Pangenome Reference Consortium (HPRC) and the Human Genome Structural Variation Consortium (HGSVC), Science and Technology Daily reported.

For the first time, it uses 636 high-quality "telomere-to-telomere" human genomes from around the world as training data. These data cover different populations globally, helping gain a more comprehensive understanding of human genetic diversity.

In terms of algorithm architecture, Genos adopts a "mixture of experts" framework to accurately schedule strongly relevant "expert" algorithms for collaborative processing. While aggregating massive 10-billion-level parameters, it successfully reduces inference costs and resource consumption, making the model both powerful and easy to use.

Test results show that Genos achieves an accuracy rate of 92 percent in the task of interpreting pathogenic mutations directly for clinical applications; when combined with scientific foundational models, the accuracy rate reaches as high as 98.3 percent. Multiple comprehensive evaluation results also indicate that Genos surpasses the current best-performing models.

A staff member in charge of BGI-Research revealed that the Genos model has been fully open-sourced on machine learning and AI tool platforms such as HuggingFace and ModelScope, providing two versions with 1.2 billion and 10 billion parameters to meet different needs.