DeepD: Phylogeny-Informed Genome Foundation Model and Controllable Long-Context Sequence Design

Overview

DeepD is a sparse mixture-of-experts (MoE) genome foundation model for efficient long-context DNA modeling and generation. It operates at single-nucleotide resolution with a 1-million-token context window, while activating only a subset of experts for substantially reduced computation.

Pretrained on 5.69 Tb unselected whole-genome sequences across all domains of life, DeepD learns genic and intergenic sequence features and supports broad sequence-to-function prediction across species. DeepD performs well in human protein fitness and disease variants effect prediction, such as BRCA1 and BRCA2, with particular strength in regulatory regions, such as promoter and transcription-factor binding sites.

The Context–Anchor–Body strategy of DeepD enables generation of coherent yeast chromosome sequences and megabase-scale Arabidopsis segments, that can hardly be distinguished from sequences of the natural yeast population. Together, DeepD is an efficient genome foundation model that unifies accurate sequence-to-function prediction with controllable genome-scale generation, advancing predictive and generative design of complex genomes.

Repository contents

deepd/
├── apiexample/          inference gateway client, CLI, and parameter reference
├── notebook/            zero-shot, generation, interpretability, synteny, and fine-tune demos
├── figure/              overview and architecture figures
└── requirements.txt

Installation

git clone https://github.com/biomap-research/DeepD.git
cd deepd
pip install -r requirements.txt

Set an API token before calling the gateway:

export INFERENCE_API_TOKEN="your_api_token"

You can also place the token in the git-ignored API-Key.txt at the repository root, or pass --token on each command.

Inference API

All DeepD features go through the apiexample gateway client (apiexample/cli.py). Each invocation submits a job, polls until it finishes, writes {task_id}.json to --output-dir, and prints a short task-type summary.

Task	Returns
embedding	nucleotide-resolution sequence representation
logits	per-position logits, mean loss, and perplexity
generate	DNA sequence conditioned on a prompt

cd apiexample

python cli.py --prompt ATGCATGC --task-type embedding \
  --species "Homo sapiens" --output-dir results

The API usage guide covers configuration, worked examples, and the full parameter list.

Notebooks

Demo	Notebook	What it shows
Zero-shot	notebook/zero_shot/coding_auroc.ipynb	ClinVar coding SNVs scored from mutant vs. wild-type log-likelihood
Generation	notebook/generation/sequence_generation.ipynb	Species-conditioned generation and GC-content check
Interpretability	notebook/interpretability/DeepD_HBB_best_layer_OVR_demo.ipynb	HBB interval, best-layer one-vs-rest tracks — see notebook/interpretability/README.md
Synteny	notebook/synteny/synteny_workflow.ipynb	AIYeast00 vs. S288C chromosome III — see notebook/synteny/README.md
Fine-tune	notebook/fine_tune/gue_h3_embedding_classification.ipynb	GUE H3 classification on DeepD embeddings — see notebook/fine_tune/README.md

Several notebooks download large runtime assets on first run. Use the README in each notebook directory for setup.