Phylogeny-informed sparse mixture-of-experts genome model at single-nucleotide resolution, with a 1M-token context for variant scoring and design.
No providers recorded yet. Browse all providers
A maize genome is more than 85% repetitive sequence, and the enhancer that domesticated the plant sits roughly 60 kb upstream of the gene it controls. Most DNA language models never see that landscape: they are trained on genes and their immediate surroundings, on sampled reference segments, or on curated human intervals, and their context windows are far shorter than the distances over which regulation actually operates. The intergenic space is discarded as noise, along with the signal inside it.
DeepD is a genome foundation model built to read genomes as they occur. Developed by China Agricultural University with BioMap Research and collaborators, and released as a bioRxiv preprint in October 2026, it is a sparse mixture-of-experts transformer that operates on raw nucleotides over a context of up to one million tokens. Two design choices carry the entry: pretraining on unselected whole genomes, so coding sequence, introns, repeats and intergenic DNA are all present in their native proportions; and a species-informed router that conditions expert selection on a species embedding derived from an established phylogenetic tree, letting one model hold lineage-specific composition and codon usage rather than averaging across life.
The same pretrained checkpoint scores variants by likelihood change, exposes genomic annotations through its routing weights, and generates DNA. Where Evo 2 continues from a starting sequence and AlphaGenome predicts functional tracks for human and mouse intervals, DeepD aims at specification: promoters with a requested strength, and chromosome-scale sequences assembled under constraints.
DeepD was trained at two scales: DeepD-2B (1.8B parameters, 128 experts, 1 trillion tokens) and the flagship DeepD-60B (63B parameters, 576 experts, 24 layers, hidden size 2,048, 5.69 trillion tokens drawn from about 3.07 million genomes across eukaryotes, bacteria, archaea and viruses). Each block mixes 24 sliding-window attention heads with 8 full-attention heads under a sigmoid output gate; pretraining runs at 8K context and midtraining extends it stepwise to 1M, with validation perplexity improving at every step. Zero-shot, DeepD-60B reaches AUROC 0.934 on ClinVar SNVs and 0.965 on non-SNVs, and leads on ncRNA fitness (mean |ρ| 0.297 over seven datasets) ahead of the Evo series and RiNALMo; bacterial protein fitness reaches mean |ρ| 0.551, ahead of ProGen2. Classifiers on frozen representations lead general genome language models on promoter (0.860) and core-promoter (0.620) detection, while AlphaGenome, trained directly on functional genomic tracks, remains stronger on TF binding and splice sites.
The model is aimed at crop and microbial genome work as much as human genetics. Fine-tuned on maize STARR-seq activity labels, it generated 170-bp synthetic core promoters conditioned on high or low strength; dual-luciferase assays in maize protoplasts confirmed the intended direction for 76% of high designs and 83% of low ones. Using Context-Anchor-Body, it produced 29 full-length S. cerevisiae chromosome III sequences on the S288C backbone and megabase Arabidopsis segments that recover 99.2% of core genes. For variant interpretation it offers zero-shot scoring in species with no curated functional data, which is where catalogs of labeled variants do not exist.
DeepD is the first genome language model to combine sparse mixture-of-experts routing with explicit phylogenetic conditioning, and its results argue that whole-genome pretraining — not gene-enriched corpora — is what makes non-coding and repetitive DNA legible. The work remains a preprint awaiting peer review. Only analysis scripts, notebooks and the generated sequences are on GitHub; the model itself is reachable through a hosted inference API that returns embeddings, per-position logits and prompt-conditioned generation, with free keys for non-commercial academic use under CC BY-NC 4.0. The authors state that source code, weights and the full inference pipeline will be released in the same repository no later than publication, so independent reproduction of the benchmark numbers is not yet possible.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.