miRBind 2.0

Deep-learning models for predicting miRNA–mRNA interactions.

This repository ships two models:

• Pairwise binding-site model — a CNN that predicts whether a given miRNA binds a given target site (≈50 nt window). Use this to score candidate binding sites.
• Gene-level repression model — predicts the gene-level fold change a miRNA induces from a full 3'UTR sequence. Built on top of the binding-site model via transfer learning.

Installation

Clone the repo and install the dependencies (Python ≥ 3.9, PyTorch ≥ 1.9):

git clone https://github.com/BioGeMT/miRBind_2.0.git
cd miRBind_2.0
pip install -r code/pairwise_binding_site_model/requirements.txt

A GPU is recommended but not required — the models will fall back to CPU automatically.

Quick start: predicting miRNA binding sites

The trained binding-site model is included in models/pairwise_onehot_model_20260105_200141.pt.

1. Prepare your input

A TSV file with at minimum these columns:

column 0 (target/mRNA)	column 1 (miRNA)	label
TTTTTTTT...GACAGTGG	TGTGCAAATCTATGCAAAACTGA	0

The label column is required by the data loader but is ignored at inference. Set it to 0 if you don't have ground truth. A small example is provided in data/chimeric_datasets/sample_dataset/.

2. Run inference

cd code/pairwise_binding_site_model

python -m inference.predict \
    --model_path ../../models/pairwise_onehot_model_20260105_200141.pt \
    --input_file path/to/your_sites.tsv \
    --output_file predictions.tsv \
    --model_type pairwise_onehot \
    --batch_size 32

The output TSV is your input plus two columns:

• prediction_score — binding probability in [0, 1]
• predicted_class — 1 if prediction_score > 0.5, else 0

There is also a ready-to-edit wrapper script at analysis/pairwise_binding_site_model/inference.sh.

Quick start: predicting gene-level repression

See analysis/gene_level_model/README.md for the full walkthrough. Briefly:

# install gene-level model dependencies
pip install -r analysis/gene_level_model/requirements.txt

# download the training/eval data
bash analysis/gene_level_model/download_data.sh

# evaluate on a test set (or train your own — see analysis/gene_level_model/train.sh)
bash analysis/gene_level_model/evaluate.sh

The gene-level model takes a full 3'UTR sequence (up to several thousand nt) and a miRNA sequence and predicts a scalar fold change.

Explainability

The binding-site model supports SHAP-based attribution (via Captum's GradientShap). See code/pairwise_binding_site_model/README.md for the SHAP, clustering, and aggregation pipelines.

Downloading the public datasets

To reproduce the published results or train from scratch:

bash data/scripts/run_zenodo_downloader.sh

This pulls the AGO2 eCLIP Manakov 2022 train / test / leftout splits from Zenodo into data/chimeric_datasets/.

Repository layout

• code/ — model definitions, encoders, training and inference scripts.
• analysis/ — runnable wrapper scripts (train.sh, inference.sh, etc.) for each model.
• data/ — placeholder; populated by the download scripts above.
• models/ — trained model checkpoints.

Models leaderboard

We track model performance on the Manakov22 test and leftout datasets, ranked by Average Precision score (AP) AP(test) + AP(leftout).

Rank	Model	AP(test)	AP(leftout)	Model	Code	Date	Authors
1	Pairwise encoding with conservation (+2 channels)	85.93	82.26	model	code	2025-03-27	Dimos, David, Panos
2	Pairwise encoding CNN	84.97	83.08	model	code	2025-03-19	David, Panos
3	Retrained miRBind CNN (miRBench)	84.00	81.00	—	—	2025-03-19	Eva
4	TargetScanCNN	77.00	76.00	—	—	2025-03-19	TargetScan

Transcript repression predictions

Transcript repression predictions stored on Google Drive

Citation

If you use miRBind 2.0 in your work, please cite the corresponding manuscript: miRBind2 enables sequence-only prediction of miRNA binding and transcript repression.