DeepD enables efficient genome modeling and controllable long-context sequence generation across domains of life

Genome foundation models have the potential to revolutionize biological research, yet modeling intricate genomes and creating controllable, genome-scale sequences pose significant challenges. DeepD, a novel phylogeny-informed sparse mixture-of-experts model, addresses these issues by operating at the single-nucleotide level across a 1-million-token context, all while minimizing active computation.

Trained on a diverse range of whole-genome sequences from every domain of life, DeepD learns both genic and intergenic sequence features, enabling broad sequence-to-function predictions across various species. This model excels in predicting human protein fitness and the effects of disease variants, such as those found in BRCA1 and BRCA2, with a particular emphasis on regulatory regions like promoters and transcription-factor binding sites.

DeepD recognizes exon-intron boundaries, transposable elements, and distal cis-regulatory elements through expert routing. In addition to its predictive capabilities, DeepD can generate experimentally validated core promoters with specified transcriptional strengths. The innovative Context-Anchor-Body strategy allows DeepD to generate complete yeast chromosome sequences and megabase-scale Arabidopsis segments, preserving gene structure and order while introducing natural variations within certain ranges.

In summary, DeepD represents an efficient genome foundation model that seamlessly combines accurate sequence-to-function prediction with controllable, genome-scale generation, driving forward the predictive and generative design of complex genomes.