ChromDL is a hybrid neural network comprised of bidirectional gated recurrent Units (BiGRUs), separable and traditional convolutional neural networks (CNNs), and bidirectional long short-term memory units (BiLSTMs) for the prediction of Transcription Factor Binding Sites (TFBS), DNase-I hypersensitive sites (DHS), and Histone Modifications (HM) using only DNA sequences as input.
ChromDL was trained and run on the Biowulf HPC Cluster hosted by the National Institutes of Health. To do so, a .sif container file was created that included all required packages and can be found in the container definition file. The docker image was compiled from the Nvidia CUDA 11.0 toolkit including cuDNN version 8 for Ubuntu 18.04.
Included in this container are the following:
- Python (version 3.7.5)
- Tensorflow-GPU (version 2.4.0 from Google API Linux releases)
- H5py (version 2.10.0)
- Numpy (version 1.19.5)
- Scikit-learn (version 0.24.1)
- SciPy (version 1.6.0)
- Sklearn (version 0.0)
Also included in the container file for other file processing were:
- Pandas (version 1.2.1)
- Tensorflow Addons (version 0.12.1)
- Matplotlib (version 3.3.4)
ChromDL was trained for 100 epochs but stagnanted after 20 epochs, with each epoch taking ~16 hours on a single core NVIDIA Tesla K80 GPU processor with 24 GB RAM.
ChromDL can also be run out of the box locally or on a Google Colaboratory GPU, where ChromDL_test.py can be run using a standard GPU runtime, and ChromDL_train.py using a high-RAM premium GPU runtime.
python ChromDL_train.py
ChromDL_train.py will train the model for the 100 epochs using a batch size of 500 and save results to train_out/ChromDL_train_out.txt, and the weights from the training to train_out/ChromDL_learned_weights
Or, the user can input the number of epochs or change the batch size and run:
python ChromDL_train.py <batch size> <number of epochs>
ChromDL_train.py will train the model for the specified number of epochs using the inputted batch size and save results to train_out/ChromDL_train_out.txt, and the weights from the training to train_out/ChromDL_learned_weights
python ChromDL_test.py
ChromDL_test.py will run the fully trained model loaded from ChromDL_best_weights and calculate the performance metrics for auROC and auPRC, saving them to test_out/ChromDL_test_metrics.txt. Alternatively, if you wish to calculate training metrics on a user trained model, run the following:
python ChromDL_test.py <weight file>
Where the weight file will be the ChromDL_learned_weights from training.
There is also a ChromDL_test.ipynb interactive notebook that can be run on Google Colaboratory using the free GPU runtime and the testing dataset in ~8-10 minutes.
The ChIP-seq and DNase-seq data that this model was trained on can be found in DeepSEA's codebase, http://deepsea.princeton.edu/media/code/deepsea_train_bundle.v0.9.tar.gz.
The individual links for each of the 919 chromatin features labels can be found in data/data_sources/DeepSEA_data_sources.txt, and the other data sources used in the enhancer and raQTL experimentation and validation of the model can be found in data/data_sources/
The container.sif used as the docker image for training and testing ChromDL can be downloaded from: https://drive.google.com/uc?id=1LHVtgATsO4n4WYCFdw46goFgjibCrjip&export=download
The container.sif file must be instantiated with python3.6 to initialize correctly.
The fully trained ChromDL weight files can be downloaded from: https://drive.google.com/drive/folders/1vWxdDRbn3v4oit5D9RiMxiR-vopYnj6w?usp=sharing
For assistance or issues with ChromDL, please contact chrishil@umich.edu.