deepvregulome 0.1.1

DNABERT-based framework for predicting the functional impact of regulatory variants

DeepVRegulome

DeepVRegulome is an end‑to‑end framework for predicting the functional impact of small somatic variants in non‑coding regulatory regions (splice sites and transcription‑factor‑binding sites) using fine‑tuned DNABERT models.

---

✨ Key Features

• ✅ DNABERT-based classifiers for:
  • Splice sites (acceptor, donor)
  • ~700 TFBS models
• ✅ Region-aware scoring of somatic variants using Δp and log₂ odds
• ✅ Batch processing with multiprocessing and BED/VCF support
• ✅ Interactive Streamlit dashboard with:
  • Variant tables, plots, and survival analysis
  • Attention score visualizations

---

📁 Repository Structure

DeepVRegulome/
├── .devcontainer/
├── .streamlit/
├── data/
│   └── Brain/
├── figures/                         # Exported visualizations (e.g. attention maps)
│   └── attention/
│       ├── CTCFL/
│       └── ZNF384/
├── notebooks/                      # Jupyter notebooks for key pipeline steps
│   ├── 01_parse_and_merge_vcfs.ipynb            # Merge and parse VCFs
│   ├── 02_tfbs_intersection.ipynb               # Intersect VCF with TFBS BEDs
│   ├── 03_dnabert_input_generation.ipynb        # Generate sequences for DNABERT
│   ├── 04_scoring_candidate_variants.ipynb      # Compute Δp / logOR & rank variants
│   └── 05_tfbs_attention_motif_visualization.ipynb  # Plot attention scores & motifs
├── scripts/                       # Shell scripts for batch inference
│   ├── run_prediction_tfbs.sh                 # Predict with TFBS models
│   └── run_prediction_splice_acceptor.sh      # Predict with acceptor models
├── src/
│   └── deepvregulome/             # Core Python modules
│       ├── __init__.py
│       ├── dnabert_data_generation.py         # Wild/mutated seq generation
│       ├── intersect.py                       # BED/VCF overlap engine
│       ├── vcf_loader.py                      # VCF parsing utilities
│       └── config.yaml                        # Centralized path config
├── streamlit_app/
│   └── app_variant_clinical_dashboard.py      # Live clinical dashboard
├── LICENSE
├── README.md
├── requirements.txt
└── .gitignore

🧪 Installation

git clone https://github.com/DavuluriLab//DeepVRegulome.git
cd DeepVRegulome
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

⚙️ Typical Pipeline Flow

Step	Description	Location
1️⃣	Parse + merge somatic VCFs	01_parse_and_merge_vcfs.ipynb
2️⃣	Intersect variants with TFBS BEDs	02_tfbs_intersection.ipynb
3️⃣	Generate ref/mutated k-mers for DNABERT	03_dnabert_input_generation.ipynb
4️⃣	Predict with DNABERT models	scripts/run_prediction_tfbs.sh
5️⃣	Compute Δp, find candidate variants	04_scoring_candidate_variants.ipynb
6️⃣	Visualize attention scores and motifs	05_tfbs_attention_motif_visualization.ipynb
7️⃣	Browse results interactively	streamlit_app/app_variant_clinical_dashboard.py

📊 Example Outputs

• Candidate variant count by TFBS
• DNABERT attention heatmaps
• High-impact motif shifts due to mutations
• Kaplan–Meier plots for clinical stratification

See figures/attention/ for examples like CTCFL.

🌐 Live Demo

An interactive instance of the DeepVRegulome dashboard is hosted here: ➡️ https://deepvregulome.streamlit.app/ The deployed app allows you to explore model performance metrics and variant-effect predictions without the need to install any software locally.

🧬 Model Checkpoints

Full DNABERT fine-tuned weights (acceptor, donor, and 700 TFBS models) will be deposited in Zenodo and made publicly available immediately upon journal acceptance. In the meantime, researchers may request access by emailing [e-mail] and [e-mail] with a brief statement of intended use.

Citation

If you use DeepVRegulome in your research, please cite:

@misc{dutta2025deepvregulomednabertbaseddeeplearningframework,
      title={DeepVRegulome: DNABERT-based deep-learning framework for predicting the functional impact of short genomic variants on the human regulome}, 
      author={Pratik Dutta and Matthew Obusan and Rekha Sathian and Max Chao and Pallavi Surana and Nimisha Papineni and Yanrong Ji and Zhihan Zhou and Han Liu and Alisa Yurovsky and Ramana V Davuluri},
      year={2025},
      eprint={2511.09026},
      archivePrefix={arXiv},
      primaryClass={q-bio.GN},
      url={https://arxiv.org/abs/2511.09026}, 
}

🧬 Model Checkpoints

462 fine-tuned DNABERT models (458 TFs + 4 histone marks) are available on HuggingFace: ➡️ duttaprat/DeepVRegulome

pip install deepvregulome

from deepvregulome import DVR
dvr = DVR(genome="hg38.fa")
result = dvr.score_variant("chr1", 3456782, "A", "C", models=["CTCFL", "SP1"])

After editing both files on GitHub, rebuild and re-upload to PyPI:

cd /vast/home/pdutta/Github/dvr_package
git pull   # get the README changes
python -m build
~/.local/bin/twine upload dist/* --skip-existing

The --skip-existing flag is important — it won't fail on the already-uploaded files and will only upload the new version if you bump the version number. Actually, for a new release you'll need to change version = "0.1.1" in pyproject.toml first, since PyPI won't accept the same version twice.

MIT. See LICENSE for details.