crush-hic 1.0.0

Compartmental Refinement for Ultraprecise Stratification in Hi-C — A/B chromatin compartment analysis tool

CRUSH — Compartmental Refinement for Ultraprecise Stratification in Hi-C

---

CRUSH (Compartmental Refinement for Ultraprecise Stratification within Hi-C) is a command-line tool that identifies fine-scale A/B chromatin compartments from Hi-C contact matrices. It has successfully identified compartments in Hi-C, Micro-C, and Single-Cell Hi-C data, and specializes in calling compartments at high resolutions with significantly lower read depth than other compartment calling tools.

Manuscript in preparation — JRowleyLab, PI: Jordan Rowley

---

Table of Contents

• How It Works
• Installation
• Quick Start
• Input Files
• Output Files
• Key Parameters
• Test Dataset
• Dependencies
• Citation
• Contact

---

How It Works

At its core, CRUSH asks a simple question for every genomic bin: does this bin interact more with A-type regions (iA) or B-type regions (iB)?

The algorithm walks from coarse resolutions down to your target resolution, using each level to refine A/B compartment assignments at the next finer level:

1. Eigenvector initialization — Computes principal components of the Hi-C contact matrix (or accepts a user-supplied eigenvector) to define initial A (iA) and B (iB) states.
2. CRUSH score calculation — At each resolution, calculates a Genome Interaction (GI) score per bin reflecting how much more it contacts iA regions versus iB regions.
3. Compartment reclassification — After each resolution pass, A/B bin assignments are updated based on the new scores, then used to seed the next finer resolution.
4. Resolution walking with midpoint shifting — A rolling-window alignment step adjusts finer-resolution scores against the coarser baseline, removing systematic biases between resolution levels.
5. Statistical filtering — Applies Benjamini–Hochberg FDR correction and outputs a q-value filtered bedGraph.

A compartments → positive CRUSH score (gene-rich, open chromatin, active transcription)
B compartments → negative CRUSH score (gene-poor, closed chromatin, transcriptionally silent)

Unlike eigenvector-based methods, you never need to flip CRUSH scores — A is always positive and B is always negative.

---

Installation

pip install CRUSH-hic

We recommend setting up a dedicated conda environment:

conda create -n crush_env python=3.10
conda activate crush_env
conda install -c bioconda bedtools
pip install CRUSH-hic hic-straw cooler numpy scipy pandas statsmodels tqdm

Dependencies

Tool	Purpose	Install
Python ≥ 3.8	Runtime	python.org
bedtools	Genomic intersections	conda install -c bioconda bedtools
mawk	Fast text processing	sudo apt install mawk / brew install mawk
hic-straw	Read .hic files	pip install hic-straw
cooler	Read .mcool files	pip install cooler
numpy / scipy / pandas	Numerical computing	pip install numpy scipy pandas
statsmodels	FDR correction	pip install statsmodels
tqdm	Progress bars	pip install tqdm

Verify installation

crush --help

---

Quick Start

With genome build shortcut (supported builds: hg19, hg38, mm10, mm9; res ≥ 500 bp)

crush \
  -i data.hic \
  -gb hg38 \
  -r 10000 \
  -c 8 \
  -o output_prefix_

With manual reference files (any genome, any resolution)

crush \
  -i data.hic \
  -g hg38.sizes \
  -a hg38_genes.bed \
  -b hg38.fa \
  -r 10000 \
  -c 8 \
  -o output_prefix_

Chromosome naming: CRUSH automatically detects and converts chromosome prefix mismatches between your Hi-C file and reference files (e.g., chr1 vs 1). If output is empty or unexpected, verify that your Hi-C file itself uses a consistent naming convention throughout.

---

Input Files

Always required

Flag	Description
-i	Hi-C file (.hic from Juicer or .mcool from cooler). Local path or HTTPS URL.
-r	Target resolution in base pairs (e.g., 10000 for 10 kb). Must exist in your Hi-C file.

Reference files — choose one of two paths

PATH A — genome build shortcut (res ≥ 500 bp only)

Flag

Description

-gb

Genome build shortcut. Supported builds: hg19, hg38, mm10, mm9. Auto-downloads chr.sizes, genes.bed, and Bbins.bed from JRowleyLab GitHub. Not available for res < 500 bp because the hosted Bbins.bed was pre-computed at 500 bp — for sub-500 bp analysis supply -g, -a, and -b (FASTA) manually so CRUSH can recompute Bbins at your exact resolution. Explicit -g/-a/-b flags override the auto-download for that specific file.

PATH B — manual reference files (any genome, any resolution)

Flag	Description
-g	Chromosome sizes file — two tab-separated columns: chr_name and size (bp). No header.
-a	BED file (≥ 3 columns) for A-compartment initialization. Gene annotations work well. ChIP-seq peaks for an active histone mark (e.g., H3K27ac) also work.
-b	Genome FASTA or pre-computed Bbins BED for B-compartment initialization. With FASTA, CRUSH generates Bbins at 500 bp (res ≥ 500 bp) or at the input resolution (res < 500 bp). With BED, the file is used directly as B-compartment seeds.

Optional

Flag	Description
-e	Pre-computed eigenvector bedGraph (4 columns: chr, start, end, value). Positive = A, Negative = B. Skips automatic eigenvector calculation.

---

Output Files

CRUSH produces four output files, each prefixed with whatever you supply via -o:

File	Description
{prefix}CRUSHparamters.txt	Record of all parameters used. Keep this for reproducibility.
{prefix}mergedCrush_{res}.bedgraph	Main output. CRUSH scores for every bin. Positive = A compartment, Negative = B compartment. Unlike eigenvectors, scores never need to be flipped.
{prefix}mergedqvalue_{res}.bedgraph	Estimated q-value (BH-corrected) for each bin's score.
{prefix}mergedCrush_{res}_qfiltered_reprocess.bedgraph	CRUSH scores filtered to bins passing the q-value threshold. Note: this filter can be overly stringent — excellent results are often obtained from the unfiltered mergedCrush file.

All bedGraph files include a UCSC track header for direct loading into genome browsers (IGV, UCSC, WashU).

While running, CRUSH creates a temporary working directory named CRUSHtmp_[randomnumber] in your current directory. This is removed automatically when the run completes. To keep it (e.g., for debugging), use -C 0. You can also name it yourself with -f.

---

Key Parameters

Flag	Default	Description
-c	1	Number of CPU threads. Set to number of chromosomes or available cores, whichever is smaller.
-gb	(none)	Genome build shortcut (hg19, hg38, mm10, mm9). Auto-downloads reference files. res ≥ 500 bp only.
-o	(none)	Output file prefix.
-N	NONE	Normalization: NONE, VC, VC_SQRT, KR, SCALE.
-m	2500000	Coarsest resolution to start walking from.
-Z	100000	Resolution for eigenvector calculation (100 kb recommended).
-w	5	Sliding window size (kb) for score averaging. Set to 1 to disable. Set to 0 for legacy auto-calculation from sequencing depth.
-q	0.05	Q-value threshold for filtered output. Set to 0 to disable filtering.
-s	0	Enable boundary smoothing (1 = on).
-A	0	Adjust score distribution. Do not use when comparing samples.
-C	1	Clean up temp files after run (0 = keep).
-v	0	Verbose output (1 = on).

For the complete parameter reference, see the User Manual.

---

Test Dataset

A small test dataset covering chromosomes 17–19 of hg19 is provided in examples/TestData/:

File	Description
hg19_c17_18_19_1kb.hic.gz	Hi-C contact file
hg19_c17_18_19_genes.bed.gz	Gene annotations for A-state initialization
hg19_c17_18.fa.gz	Genome FASTA for GC-based B-state initialization
hg19_c17_18.fa.fai	FASTA index
hg19_c17_18_19.sizes.gz	Chromosome sizes
Eigen_100kb_c17_18_19.bedgraph.gz	Pre-computed eigenvector (optional -e input)
Bbins_hg19_c17_18_19.bed.gz	Pre-computed B-bins (alternative to FASTA for -b)

Run the test

# Decompress
gunzip examples/TestData/*.gz

# Run with FASTA-based B initialization
crush \
  -i examples/TestData/hg19_c17_18_19_1kb.hic \
  -g examples/TestData/hg19_c17_18_19.sizes \
  -a examples/TestData/hg19_c17_18_19_genes.bed \
  -b examples/TestData/hg19_c17_18.fa \
  -r 10000 \
  -c 4 \
  -o test_

Expected output: test_mergedCrush_10000.bedgraph, test_mergedqvalue_10000.bedgraph, and test_mergedCrush_10000_qfiltered_reprocess.bedgraph.

Load test_mergedCrush_10000.bedgraph into IGV or the UCSC browser to verify the A/B compartment pattern on chr17–19.

---

Citation

Manuscript in preparation. If you use CRUSH in your research, please check back for the citation or contact us directly.

---

Contact

JRowleyLab | PI: Jordan Rowley
For questions, bug reports, or feature requests, please open a GitHub Issue.