NetTIME

A multitask and base-pair resolution framework for predicting cell-type specific transcription factor binding sites.

Citation

If you use NetTIME in your research, please cite:

@article{yi2022nettime,
    author = {Yi, Ren and Cho, Kyunghyun and Bonneau, Richard},
    title = "{NetTIME: a multitask and base-pair resolution framework for improved transcription factor binding site prediction}",
    journal = {Bioinformatics},
    year = {2022},
    month = {08},
    issn = {1367-4803},
    doi = {10.1093/bioinformatics/btac569},
    url = {https://doi.org/10.1093/bioinformatics/btac569},
}

Installation

Run the following commands to clone the repository and install NetTIME:

git clone https://github.com/ryi06/NetTIME.git
cd NetTIME

# Create a Conda environment with Python 3.6 and install required packages
conda env create --file environment.yml
conda activate nettime

Please refer to Pytorch documentation for instructions on setting up CUDA.

Making predictions using a trained NetTIME model

We use NetTIME_predict.py to make predictions from a trained NetTIME model. Pretrained NetTIME models can be found here. We use an example prediction dataset to demonstrate how to make predictions using trained NetTIME models. Check out this tutorial on how to generate datasets like this.

Making binding probability predictions for JUN.K562 and JUNB.GM12878 from a trained NetTIME model can be achieved by running the following. See NetTIME_predict.py -h for all available arguments.

python NetTIME_predict.py \
--batch_size 2700 \
--num_workers 10 \
--dataset "data/datasets/prediction_example/predict.h5" \
--dtype "prediction" \
--index_file "data/embeddings/pretrained.pkl" \
--experiment_name "prediction_example" \
--model_config "pretrained/seqCT/seqCT.config" \
--best_ckpt "pretrained/seqCT/seqCT_433000.ckpt" \
--eval_metric "aupr" \
--no_target \
--predict_groups "JUN.K562" "JUNB.GM12878"

Binding probability predictions will be saved in experiments/prediction_example/prediction_predict. If you wish to further perform binary classification on the predicted binding probabilities using a pretrained conditional random field (CRF) classifier, run NetTIME_CRF_predict.py as follows:

python NetTIME_CRF_predict.py \
--batch_size 2700 \
--num_workers 10 \
--prediction_dir "experiments/prediction_example/prediction_predict" \
--experiment_name "prediction_example" \
--dtype "prediction" \
--model_config "pretrained/seqCT/seqCT_crf.config" \
--best_ckpt "pretrained/seqCT/seqCT_crf_250000.ckpt"

Binary binding classification will be saved in experiments/prediction_example/prediction_crf_predict.

Training a NetTIME model

A tutorial on how to train a NetTIME model using example training data can be found in training_example.md.

Dataset

Instructions for downloading datasets used to generate main results in the manuscript can be found here.