Please email kxiong@andrew.cmu.edu with any questions about installation or usage.
To quickly install SNIPER, clone SNIPER's repository and install the necessary requirements by running
pip install -r requirements.txt
in the shell. We recommend creating a separate python 3.6 environment. Installation should take under 15 minutes for a computer with broadband internet connection. For systems not running SNIPER on a GPU, SNIPER can be installed by running
pip install -r requirements-cpu.txt
All Python dependencies can be installed by running
pip install -r requirements.txt (see Installation)
Requires Java Runtime Engine (on Linux, default-jre should work).
Juicer tools is a utility developed by the Erez Lieberman-Aiden lab that can extract Hi-C data from .hic files. Follow the link provided to download juicer_tools.jar. Make a note of where juicer_tools.jar is stored on your file system.
SNIPER will call the directory where juicer_tools.jar is stored.
SNIPER was developed using a NVidia GeForce GTX 1080 Ti. SNIPER uses CUDA 9.0 and cuDNN v7.0.5 to run Keras on the tensorflow-gpu backend. SNIPER should work with recent versions of CUDA and cuDNN as well. Please email kxiong@andrew.cmu.edu with any questions regarding python and CUDA environments.
Because there are so many versions of NVidia GPUs, we cannot say for certain how long SNIPER will need to finish training. For reference, SNIPER takes approximately 15 seconds to train one epoch of its autoencoder on our 1080 Ti. For users without a dedicated NVidia GPU, SNIPER will still work (install python packages using requirements-cpu.txt instead of requirements.txt), but will take significantly longer to train. Using a 3.6 GHz 6-core/12-thread processor, one epoch took approximately 3 minutes.
SNIPER is separated into two modules - training and application. To train a new SNIPER model, run the following python command:
python sniper_train.py <input_hic_path> <target_hic_path> <annotation_path> [options]
input_hic_path is the file path to the .hic file of the downsampled training Hi-C matrix. target_hic_path is the path to the .hic file of the dense target Hi-C matrix. We have provided GM12878's ground truth annotations in .mat format in SNIPER's root directory. annotation_path is the path to a .mat file of the GM12878 annotations published by Rao et al. (2014). We have included a .mat file of their annotations in the data directory of this repository (data/labels.mat).
Training with .hic files from scratch can take up to 60 minutes (depending on the size of the .hic file) because Juicer needs to extract the inter-chromosomal data of each pair of odd and even chromosomes, convert the contact data into a matrix, and then trim the matrix. Once training is complete, sniper_train.py will output six models to the user's specified directory (see Command line options). These models are the autoencoders, encoders, and classifiers trained on odd and even chromosomes.
To apply SNIPER to another cell line, run the following python command:
python sniper.py <input_path> <output_path> <odd_encoder> <odd_clf> <even_encoder> <even_clf> [options]
input_path specifies the path to the input Hi-C matrix (.hic or .mat format). output_path specifies a .mat file of the output predictions. [odd/even]_encoder specifies the keras model of the autoencoder trained with odd or even chromosomes along the rows. [odd/even]_clf specifies the keras model of the classifier trained with odd or even chromosomes along the rows.
Application of SNIPER can likewise take up to 60 minutes to run because of Juicer's extraction process. sniper.py will output a .mat file whose keys odd_predictions and even_predictions refer to odd and even chromosome predictions respectively. In addition, SNIPER will output a .bed file in 100kb resolution with chromosomal coordinates of each prediction. The bed file is formatted similarly to the subcompartment predictions bed file in Rao et al. 2014 with color coding for easier visualization on the genome browser. To view the bed file on the genome browser, an additional header must be added as the first line of the file in the following format:
track name='<track_name>' description='<description>' itemRgb='On'
Pre-computed SNIPER models can be found here:
http://genome.compbio.cs.cmu.edu/~kxiong/data/sniper/models/
-c Specify a custom crop map directory that contains a crop map and crop indices. A crop map maps chromatin loci after the original matrix was trimmed to original chromatin locations before removing sparse loci, loci labeled as NA, and loci labeled as B4. Crop map format:
The row and column crop maps we used are provided in the crop_map directory.
Crop indices specify which rows and columns are sparse or labeled as NA or B4 in Rao et al's annotation set, which are removed from the input matrix. The general rule of thumb is to remove rows and columns where more than 30% of entries are zeros.
-dd Specifiy a directory to store output files:
-sm flag is set-jt Specify the path to juicer_tools.jar
-sm Turn this flag on to store .mat files of Hi-C matrices. Doing so will occupy approximately 3.2 GB of disk space but save a substantial amount of time if the pipeline abruptly terminates for some reason and has to be re-run.
-ar Turn this flag on to automatically remove .txt files output by Juicer Tool. Doing so will prevent clutter on the hard drive. Leaving this flag off will save time on subsequent training runs. We recommend turning this flag on if running multiple training instances on different cell types in the same directory.
-ow Turn this flag on to overwrite data existing on in the -dd directory. Recommended if running a new training instance on different cell types in a directory with existing data.
.hic files we used for training and .mat files of the inter-chromosomal Hi-C matrices can be found at the following links:
http://genome.compbio.cs.cmu.edu:8008/~kxiong/data/sniper/hic_files/
http://genome.compbio.cs.cmu.edu:8008/~kxiong/data/sniper/mat_files/
Of the included files, GM12878_combined.hic is the high-coverage Hi-C data used for training. GM12878_combined_<ds>.hic are the downsampled GM12878 Hi-C data where <ds> refers to the downsample level, i.e. 0.1 denotes 10% of the contacts present in GM12878_combined.hic.
SNIPER annotations for GM12878, K562, IMR90, HeLa, HUVEC, HMEC, HSPC, T cells, and HAP1 can be found at:
http://genome.compbio.cs.cmu.edu:8008/~kxiong/data/sniper/annotations/
under the folder annotations/hg19/ or annotations/hg38/.
ValueError: Error when checking input: expected dense_17_input to have shape (128,) but got array with shape (13393,)Please make sure to use *_encoder.h5 instead of *_autoencoder.h5 in your command line argument.
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.This is just a warning message and SNIPER should still work. The weights in the encoder/decoder models were derived from the autoencoder but were not explicitly compiled in keras during training.
If you use SNIPER in your work, please cite:
Xiong, K., Ma, J. Revealing Hi-C subcompartments by imputing inter-chromosomal chromatin interactions. Nat Commun 10, 5069 (2019).