Integrative Chromatin Domain Annotation through Graph Embedding of Hi-C data

This pipeline annotates genomic regions in a domain-scale based on both functional and structural features. Structural features are embedded in high-throughput chromosome conformation capture (Hi-C) data that captures 3D interactions between genomic regions, represented as a 2D matrix. This problem requires a framework that jointly clusters the genomic regions based on their functional features (node attributes) and physical interactions (neighborhood information) between them. We first embed the Hi-C graph information into structural features, and then jointly cluster functional and structural features together.

Requirements

• numpy
• pandas
• hmmlearn
• scipy
• GSL should be installed and added to $LD_LIBRARY_PATH.

Data generation

python src/HC_data_generation.py -c test/HC_data_generation_config.json

src/HC_data_generation.py generates processed data given raw input signals (.bigwig format) and Hi-C data (.hic format). The parameters are:

• cell_type: The cell type corresponding to data.
• signals_dir: The path of directory including [signal_name].bedgraph files. The .bigwig files from Roadmap can be converted to .bedgraph files using bigWigToBedGraph tool in utilities.
• signals_names: The list of input signals names. Note that the files in signals_dir should be in format [signal_name].bedgraph .
• hic_file: The Hi-C data file path in .hic format.
• temp_dir: The path of directory to save dump files such as binned genomic signals and extracted Hi-C interaction matrices files.
• chr_size_path: The path of genome chromosomes sizes file. Available in data/supp/hg19.chrom.sizes .
• valid_chroms: The list of chromosomes included in the experiment.
• valid_bins: The path of file including valid bins for a specific cell type in a specific resolution. Available in data/valid_bins . Those genomic bins having node features and neighbors are considered valid.
• juicer_path: The path of juicer tool to extract Hi-C interaction matrices. Available in utilities/juicer_tools_1.22.01.jar .
• resolution: The resolution of genomic regions.
• processed_dir: The path of directory to save processed files.

, which are passed in .json file. Example: test/HC_data_generation_config.json. This script makes genomic bin file (each row has 4 columns, chromosome name, start, end, and index corresponding to valid bins. Indices are between 0 and N-1, where N is total number of valid bins.), signals file (i-th row signals values for i-th bin), and interactions file (each row has 3 columns, index of source bin, index of target bin, and the interaction weight between two bins) in processed_dir.

Clustering

python src/HC.py -c test/HC_config.json

src/HC.py learns structural embeddings and combinatorial domain annotation given generated data using src/HC_data_generation.py . The parameters are:

• resolution: The resolution of genomic regions.
• num_labels: The number of domain types of a domain annotation.
• line_path: The path to line exe file. Available in utilities/LINE/linux/line .
• chr_size_path: The path of genome chromosomes sizes file. Available in data/supp/hg19.chrom.sizes .
• valid_chroms: The list of chromosomes included in the experiment.
• bins_file: The path of generated genomic bin file.
• signals_file: The path of generated signals file.
• interactions_file: The path of generated interactions file.
• processed_dir: The path of directory to save embeddings and annotations.

, which are passed in .json file. Example: test/HC_config.json. This script learns LINE embeddings and save it in processed_dir/embeddings, and learn combinatorial domain annotation and save it in processed_dir/annotations .

Annotations

Our annotations are available in data/processed_data/annotations .