Qian, B., Liang, C., Qin, C., Liu, C., Zhang, C., Xu, C., Li, D., Xue, G. et al.
Developed OneGenome-Rice (OGR), a 1.25-billion parameter genomic foundation model pre-trained on 422 rice genomes, and demonstrated its potential as a foundational computational infrastructure for functional genomics and precision breeding.
A lack of genomic foundation models for plants, particularly staple crops, hinders the transition of genomics into a predictive intelligence discipline.
Built upon a Mixture of Experts (MoE) transformer architecture with 1.25 billion parameters and pre-trained on a dataset of 422 high-quality cultivated and wild rice genomes. It supports an ultra-long context window of up to 1 million base pairs at single-nucleotide resolution.
Significantly outperforms existing state-of-the-art models across 11 benchmark categories (e.g., motif identification, sweep detection). Demonstrated utility in downstream applications like indica-japonica subspecies introgression analysis, identification of agronomy trait-associated functional loci, and prediction of gene expression from DNA sequences. The model, weights, training code, and benchmark suite have been fully released, establishing OGR as a promising foundation for rice functional genomics and precision breeding.