Harnessing DNA Foundation Models for Cross-Species TFBS Prediction in Plants

Overview

This project fine-tunes large pretrained DNA foundation models to predict transcription factor binding sites (TFBSs) in plant genomes. We benchmark three foundation models:DNABERT-2, AgroNT, and HyenaDNA, against specialized methods like DeepBind and BERT-TFBS using DAP-seq data from Arabidopsis thaliana and Sisymbrium irio.

In addition, we include a motif-based approach employing MEME for motif discovery and FIMO for scanning potential TFBSs.

The project uses a unified pipeline, which covers three evaluation protocols:

1. Cross-chromosome: Leave-one-chromosome-out evaluation on A. thaliana (Sun2022).
2. Cross-dataset: Train on Malley2016 AREB/ABF2 dataset, test on Sun2022 AREB/ABF2 dataset.
3. Cross-species: Train on one species, test on the other (Sun2022).

The HyenaDNA model achieves near state-of-the-art accuracy while training in minutes, demonstrating both predictive power and computational efficiency.

Installation

To get started, follow these steps:

1. Clone the repository

   git clone https://github.com/Maryam-Haghani/TFBS.git
   cd TFBS

2. Create conda environment

This project provides two Conda environment files:

• Linux System :
  Use environment_linux.yml

  conda env create -f environment_linux.yml
  conda activate tfbs

• Cross-platform (works on Linux, macOS, Windows):
  Use environment_cross-platform.yml

  conda env create -f environment_cross-platform.yml
  conda activate tfbs

⚠️ If you are not on Linux with an NVIDIA GPU, do not use the Linux environment file.

3. Navigate to the code directory:

   cd ./TFBS/code

---

1. Data Preparation

The DAP-seq peak used in this project come from Ronan et al., 2016 and Sun et al., 2022.
Genome sequences and DAP-seq peak files are available on Zenodo (DOI: 10.5281/zenodo.17229680).

After downloading, place the files in the project directory as follows:

• ./TFBS/data/fastas/ → genome FASTA files
• ./TFBS/data/peak_files/ → DAP-seq peak files

We used the 01-generate_samples.py script to process raw data and generate positive and negative samples for transcription factor (TF) binding sites. It takes as input a FASTA file of chromosome sequences for species genome and a CSV file containing peak regions. Negative sample will be generated based on neg_type argument with default value of "dinuc_shuffle".

Usage

To run the script, use the following command:

python 01-generate_samples.py --fasta_file path/to/your.fasta --peak_file path/to/peaks.csv --output_file path/to/output.csv --neg_type shuffle --species SI/ATA --dataset Malley2016/Sun2022

Arguments

• --fasta_file: Path to the genome FASTA file.
• --peak_file: Path to the CSV file containing peak data.
• --species: Species type (currently supporting S. irio and A. thaliana).
• --dataset: Origin of peak files.
• --output_file: Path to save the output CSV file to contain positive and negative samples.
• --neg_type: Method to generate negative samples (e.g., "dinuc_shuffle", "shuffle", "random").
• --sliding_window: Window around the peak region. Sample lengths vary based on the original peak length.
• --fixed_length: If specified, generates samples of fixed length, ignoring the sliding window.

Example

For A. thaliana (ABF1-4) dataset:

python 01-generate_samples.py --fasta_file ../data/fastas/Arabidopsis_thaliana.TAIR10.dna_sm.toplevel.fa --peak_file ../data/peak_files/Sun2022-AtABFs_DAP-Seq_peaks.csv --species "At" --dataset Sun2022 --neg_type dinuc_shuffle --output_file ../data/samples/Sun2022-AtABFs_dinuc_shuffle_neg_stride_200.csv

To run this command, you should download the FASTA file for the A. thaliana genome and place it in the ./TFBS/data/fastas directory.

For S. irio (ABF1-4) dataset:

python 01-generate_samples.py --fasta_file ../data/fastas/Si_sequence --peak_file ../data/peak_files/Sun2022-SiABFs_DAP-Seq_peaks.csv --species "Si" --dataset Sun2022 --neg_type dinuc_shuffle --output_file ../data/samples/Sun2022-SiABFs_dinuc_shuffle_neg_stride_200.csv  

To run this command, you should download the FASTA file for the S. irio genome and place it in the ./TFBS/data/fastas directory.

Output

This will generate positive and negative samples based on the given negative type generation, for the given species using the provided FASTA file and peak data, saving the results to --output_file.

2. Data Split

Description

This script splits the data based on the provided configuration file.

Usage

To run the script, use the following command:

python 02-split_data.py --config_file [config_path]

Arguments

• --config_file: Path to the configuration YAML file, which contains settings for the data split (file paths, random seed, etc.).

Data split configurations are located in the /configs/data_split directory.

Cross-configs

• cross_chromosome_config.yml: Use test_id values between 1 and 5.
• cross_species_config.yml: Use test_id as either At or Si.
• cross_dataset_config.yml: Use provided dataset identifiers.

Example

If you have a configuration file located at ../configs/data_split/cross-species-config.yml, run:

python 02-split_data.py --config_file ../configs/data_split/cross-species-config.yml

Output

Data splits are stored in <config.split_dir>/<config.name>.

3. Training / Fine-tuning Models on a Split

Description

A two‐stage pipeline that:

1. Phase 1: Trains model on each training fold, evaluates validation set over several training parameters (such as batch size, learning rate, and weight decay), and picks the parameters with the highest average F1.
2. Phase 2: Retrains the model on the full training set on the best hyperparameters for multiple random seeds, predicts for the held‑out test set, and logs key metrics (MCC, F1, accuracy, AUROC, AUPRC) for each seed.

It integrates with Weights & Biases (wandb) for experiment tracking and visualization. To enable logging, you should specify the entity_name and token values under the wandb parameter in the training configuration file.

wandb:
  entity_name: "your_wandb_username_or_team"
  token: "your_wandb_api_token"

Usage

To run the script, use the following command:

python 03-train.py --train_config_file [train_config_path] --split_config_file [data_config_path]

Arguments

• --train_config_file: Path to the training configuration YAML file.
• --split_config_file: Path to the data split configuration YAML file.

Example

python 03-train.py --train_config_file ../configs/train/HeynaDNA-config.yml --split_config_file ../configs/data_split/cross-species-config.yml

Output

Trained models and logs are saved in <config.output_dir>/<split_config.name>/

prediction_results.csv will have prediction metrics for different seeds.

4. Prediction

Description

This script generates predictions from saved pre-trained models for a dataframe or a genome sequence.

Usage

To run the script, use the following command:

python 04-predict.py --config_file [config_path]

Arguments

• --config_file: The path to the configuration YAML file.

Example

python 04-predict.py --config_file ../configs/predict/HeynaDNA-config.yml

Output

A directory named <config.input_dir without extension> will be created inside the predictions folder, located within the parent directory of the saved model directory (config.saved_model_dir). Inside this directory, a CSV file containing predictions for each sequence will be saved, named after each model, corresponding to the models saved in the model directory. Additionally, a prediction_result.csv file will be generated, summarizing performance metrics for all models.

5. Motif-Based Prediction (Traditional Approach)

This script generates predictions from traditional motif-based approach.

A two‐stage MEME + FIMO pipeline that:

1. Phase 1: Runs MEME on each training fold, scans validation sequences with FIMO over several q‑value thresholds, and picks the threshold with the highest average F1.
2. Phase 2: Retrains MEME on the full training set for multiple random seeds, applies FIMO with the chosen q‑value to the held‑out test set, and logs key metrics (MCC, F1, accuracy, AUROC, AUPRC) for each seed.

Usage

To run the script, use the following command:

python 05-predict_motif.py --split_config_file [split_config_path] --motif_config_file [motif_config_path]

Arguments

• --split_config_file: The path to the data split configuration YAML file.
• --motif_config_file: The path to the motif configuration YAML file.

Example

python 05-predict_motif.py --motif_config_file ../configs/motif-based.yml --split_config_file ../configs/data_split/cross-species-config.yml

Output

prediction_results.csv will have prediction metrics for different seeds.

Requirements

• Python 3.9+
• PyTorch 1.12+
• Transformers 4.20+
• scikit-learn 1.0+
• pandas, numpy, matplotlib

See requirements.txt for exact versions.

---

Citation

If you use this code or data, please cite:

Haghani M., Dhulipalla K.V., Li S. “Harnessing DNA Foundation Models for Cross-Species Transcription Factor Binding Site Prediction in Plant Genomes.” bioRxiv (2025).

@article{Haghani2025TFBS,
  title   = {Harnessing DNA Foundation Models for Cross-Species Transcription Factor Binding Site Prediction in Plant Genomes},
  author  = {Haghani, Maryam and Dhulipalla, Krishna Vamsi and Li, Song},
  journal = {Proceedings of the 20th Machine Learning in Computational Biology, PMLR},
  year    = {2025},
  url     = {https://proceedings.mlr.press/v311/haghani25a.html}
}

---

Contact

For questions or issues, please open an issue on GitHub or contact haghani@vt.edu.