🧬 TSProm: Tissue-Specific Promoter DNA LLM

---

Overview

TSProm is a framework that fine-tunes DNA foundation models (e.g., DNABERT2) to decipher the tissue-specific regulatory grammar encoded in promoter DNA sequences.
By comparing models specialized for general vs. tissue-specific promoter logic, TSProm isolates sequence features and motifs that uniquely define tissue identity.

---

Abstract

Characterizing tissue-specific (TSp) gene expression is crucial for understanding development and disease. Traditional expression-based methods, however, often overlook the latent regulatory grammar embedded in non-coding DNA, particularly in distal promoter regions.

Here, we introduce TSProm, a framework that specializes a DNA foundation model (DNABERT2) to decipher the long-range regulatory logic of TSp promoters at the gene isoform level.

The contributions of this work are twofold:

1. Comparative Model Design — Two distinct models are trained:

   • Model A: for general promoter biology
   • Model B: for tissue-specific promoter regulation
     This comparative strategy isolates motifs that uniquely define tissue identity.

2. Integrated Explainable AI (xAI) — Combines attention-based motif discovery with SHAP-based interpretability for robust, cross-validated feature attribution.

Applying TSProm to human brain, liver, and testis promoters, we identify clinically relevant transcription factors (e.g., SP1, MYC, HES6) and validate their known roles in diseases such as gliomas and neuroblastomas.
Our analysis further reveals that C2H2 Zinc Finger proteins dominate the global landscape of tissue-specific gene regulation.

TSProm provides an interpretable computational framework for identifying regulatory elements that drive tissue-specific gene expression — offering new insights into normal and disease states.

Keywords: Explainable AI, Tissue Specificity, DNA Language Models, Interpretable Genomics

---

Repository Structure

TSProm/
├── files/
│   ├── Figure1.pdf
│   ├── DNABERT2_TSP_vs_TransTExRest.csv
│   ├── GENA-LM_TSP_vs_TransTExRest.csv
│   ├── ModelA_best.csv
│   ├── ModelB_best.csv
│   ├── comparison_Models_A_B.svg
│   ├── comparison_methods.svg
│   ├── jobs.json
│   └── jobs1.json
│
├── notebooks/
│   ├── Fig2.ipynb
│   ├── Fig3.ipynb
│   └── Fig4.ipynb
│   ├── runs/
│   │   ├── 1_attention.....ipynb
│   │   └── 2....ipynb
│   │   └── 3_TransSHap.ipynb
│   │   └── 2C_Biclustering.ipynb
│
├── run_scripts/
│   ├── 0_generate_data_run.sh
│   ├── 1_dnabert2_finetune.sh
│   ├── 1_GENA-LM_finetune.sh
│   ├── 2_run_predict.sh
│   └── 3_run_attention.sh
│
├── src/
│   ├── 0_generate_data/
│   │   ├── make_data.R
│   │   ├── make_modelA_negSet.sh
│   │   └── make_modelB_nullSeqs.R
│   │
│   ├── 1_finetune/
│   │   ├── DNABERT2_AttentionExtracted.py
│   │   └── FineTune_GENALM.py
│   │
│   ├── 2_predict/
│   │   ├── 1_predict.py
│   │   └── 1_runAll_predict.py
│   │
│   └── 3_attention/
│       ├── 1_raw_attention_extract-gpu.py
│       ├── 2A_save_meme.py
│       └── 3_SHAP.py
│
└── README.md

---

Quick Setup

Once you clone this repository, run the following commands to install all dependencies required for both Python and R environments.

# 1️⃣ Clone the repository
git clone https://github.com/pallavisurana1/TSProm.git
cd TSProm

# 2️⃣ Set up Python environment
python3 -m venv venv
source venv/bin/activate
pip install -r requirements_LLM.txt

# 3️⃣ Set up R environment
# (Installs CRAN + Bioconductor packages automatically)
Rscript install_R_requirements.R

# 4️⃣ Verify installations
python -c "import torch, transformers, pandas; print('✅ Python OK')"
Rscript -e "library(GenomicRanges); library(BSgenome.Hsapiens.UCSC.hg38); cat('✅ R OK\n')"

---

🚀 Typical Pipeline Flow

Step	Script	Description
0	0_generate_data_run.sh	Generates benchmark datasets and Model A/B inputs. Positive sets can be obtained from Human TransTEx groupings.
1	1_dnabert2_finetune.sh / 1_GENA-LM_finetune.sh	Fine-tunes DNABERT2 and GENA-LM models. This step runs Model A and B training for the selected tissue type.
2	2_run_predict.sh	Runs prediction across train/dev/test splits. Subsets only true predictions for downstream motif analysis.
3	3_run_attention.sh	Executes attention-based motif discovery and SHAP-based interpretability module using enriched JASPAR motifs and TF analyses.

---

Example Outputs

• Example CSV outputs and visualizations are provided under the files/ and notebooks/ directories. notebooks/runs directory has information about the attention extraction module runs from the fine-tuned models.
• You can regenerate figures from the paper using Fig2.ipynb, Fig3.ipynb, and Fig4.ipynb.

---

Model Checkpoints

Fine-tuned DNABERT2 weights for Model A and Model B (Brain, Liver, Testis) will be publicly released on Zenodo after journal acceptance.

For pre-release access, please email:
📧 pallavi.surana@stonybrook.edu
📧 ramana.davuluri@stonybrookmedicine.edu
Include a brief statement of intended use.

---

Citation

If you use TSProm in your research, please cite:

Surana P, Dutta P, Papineni N, Sathian R, Zhou Z, Liu H, Davuluri RV. TSProm: deep learning framework to predict tissue-specific regulatory logic. NAR Genom Bioinform. 2026 Jun 3;8(2):lqag050. doi: 10.1093/nargab/lqag050. PMID: 42244857; PMCID: PMC13233143.

---

License

This project is released under the MIT License.
See LICENSE for details.