We proposed hicGAN, an open-sourced framework, for inferring high resolution Hi-C data from low resolution Hi-C data with generative adversarial networks (GANs)

This work has been presented in ISMB2019 conference in an oral talk during July 21-25, Switzerland.

hicGAN consists of two networks that compete with each other. G tries to generate super resolution samples that are highly similar to real high resolution samples while D tries to discriminate generated super resolution samples from real high resolution Hi-C samples.

• TensorFlow == 1.13.1
• TensorLayer == 1.9.1
• hickle == 2.1.0
• Java JDK == 1.8.0
• Juicer Tool

hicGAN can be downloaded by

git clone https://github.com/kimmo1019/hicGAN

Installation has been tested in a Linux/MacOS platform with python2.7.

We provide detailed step-by-step instructions for running hicGAN model for reproducing the results in the original paper and inferring high resolution Hi-C data of your own interst.

Step 1: Download raw aligned sequencing reads from Hi-C experiments

We preprocess Hi-C data from alighed sequencing reads (e.g. GSM1551550_HIC001_merged_nodups.txt.gz from Rao et al. 2014). One can directly download raw Hi-C data from GEO database or refer to our raw_data_download_script.sh script in the preprocess folder. You will generate raw Hi-C data under a PATH-to-hicGAN/data/CELL folder. Please note that the download may take long time.

Step 2: Generate Hi-C raw contacts for both high resolutio Hi-C data and down-sampled low resolution Hi-C data given a resolution

We use Juicer toolbox for preprocessing the raw Hi-C data. Ensure that Java and Juicer toolbox are installed in your system. One can generate Hi-C raw contacts for both high resolutio Hi-C data and down-sampled low resolution Hi-C data by running preprocess.sh script in the preprocess folder. Note that one can speed up the preprocessing using slurm by modify one line of preprocess.sh. See annotation in preprocess.sh.

bash preprocess.sh <CELL> <Resolution> <path/to/juicer_tools.jar>

For example, one can directly run bash preprocess.sh GM12878 10000 path/to/juicer_tools.jar to extract Hi-C raw contacts of GM12878 cell line with resolution 10k.

Step 3: Preprate the training and test data

Typically, Hi-C samples from chromosomes 1-17 will be kept for training and chromosomes 18-22 will be kept for testing in each cell type.

python data_split.py  <CELL>

For example, one can directly run python data_split.py GM12878 to generate train_data.hkl and test_data.hkl under the data/GM12878 data folder.

Step 4: Run hicGAN model

After preparing the training and test data, one can run the following commond to run hicGAN

python run_hicGAN.py <GPU_ID> <checkpoint> <graph> <CELL>

For example, one can run python run_hicGAN.py 0 checkpoint/GM12878 graph/GM12878 GM12878 Note that checkpoint is the folder to save model and graph is the folder for visualization with TensorBoard. The three folders will be created if not exist.

Step 5: Evaluate hicGAN model

After model training, one can evaluate the hicGAN by calculating MSR, PSNR and SSIM measurements, just run the following commond

python hicGAN_evaluate.py <GPU_ID> <MODEL_PATH> <CELL>

For example, one can run python hicGAN_evaluate.py 0 checkpoint GM12878 for model evaluation. The predicted data will be saved in data/<CELL> folder.

We finally provide a demo.ipynb to illustrate the above steps with a demo of Hi-C model.

We also provide a Results_reproduce to show how the results in our paper were produced.

Note that we also provide a pre-trained model of hicGAN which was trained in GM12878 cell line.

We provided instructions on implementing hicGAN model from raw aligned sequencing reads. One could directly run hicGAN model with custom data by constructing low resolution data and corresponding high resolution data in run_hicGAN.py with custom data by the following instructions.

Step 1: Modify one line in run_hicGAN.py

You can find lr_mats_train_full, hr_mats_train_full = hkl.load(...) in run_hicGAN.py. All you need to do is to generate lr_mats_train_full and hr_mats_train_full by yourself.

Note that hr_mats_train_full and lr_mats_train_full are high resolution Hi-C training samples and low resolution Hi-C training samples, respectively. The size of hr_mats_train_full and lr_mats_train_full are (nb_train,40,40,1) and (nb_train,40,40,1).

We extracted training examples in the original Hi-C matrices by cropping non-overlaping 40 by 40 squares (resolution: 10k bp) within 2M bp. See details in data_split.py if necessary.

Step 2: Modify one line in hicGAN_evaluate.py

You can find lr_mats_test,hr_mats_test, _ = hkl.load(...) in hicGAN_evaluate.py. All you need to do is to generate lr_mats_test and hr_mats_test by youself.

Then run the following command:

python hicGAN_evaluate.py <GPU_ID> <MODEL_PATH> <CELL>

The predicted data will be saved in data/<CELL> folder.

We also provided a script hicGAN_predict.py for which the ground truth of test data is unknown. One can run the following commond:

python hicGAN_predict.py <GPU_ID> <MODEL_PATH> <DATA_PATH> <SAVE_DIR>

[GPU_ID] : GPU ID (e.g. 0)

[MODEL_PATH]:  path for weights file for hicGAN_g(e.g. checkpoint/g_hicgan_best.npz)

[DATA_PATH]: data path for enhancing (e.g. lr_mat_test.npy)

[SAVE_DIR]: save directory for predicted data 

You need to generate your own test data in npy format and use DATA_PATH to load it. The predcited data will be saved in SAVE_DIR folder.

Feel free to contact [email protected] if you have any problem in implementing your own hicGAN model.

Liu Q, Lv H, Jiang R. hicGAN infers super resolution Hi-C data with generative adversarial networks[J]. Bioinformatics, 2019, 35(14): i99-i107.

@article{liu2019hicgan,
  title={hicGAN infers super resolution Hi-C data with generative adversarial networks},
  author={Liu, Qiao and Lv, Hairong and Jiang, Rui},
  journal={Bioinformatics},
  volume={35},
  number={14},
  pages={i99--i107},
  year={2019},
  publisher={Oxford University Press}
}

This project is licensed under the MIT License - see the LICENSE.md file for details