HiNT-RSC

HiNT-RSC is a residual-processing module for HiNT-CNV, the copy-number module of HiNT (Wang et al., Genome Biology 2020). HiNT-CNV turns a Hi-C contact map into a one-dimensional coverage profile, normalizes it with a generalized additive model (GC content, mappability, restriction sites) and segments the residuals with BIC-seq2. HiNT-RSC works after that normalization step. It smooths the residuals with a mean filter (window w). It can also subtract a fraction of a control Hi-C residual, or build the profile from intra-chromosomal contacts only. The residuals then go through HiNT's per-chromosome min-shift transform and BIC-seq2 segmentation, as in HiNT-CNV. It adds no new statistical model or segmenter, and it is not a standalone CNV caller.

The accompanying paper finds that residual smoothing and the BIC-seq2 penalty λ jointly set segmentation granularity:

• On K562, at the default λ = 3, smoothing (w = 9) added matched calls, but lowering λ without smoothing reached a similar or better balance.
• The K562-selected setting did not transfer to PANC-1.
• The most consistent benefit of smoothing was a higher bin-level agreement with WGS copy-number profiles.

The recommendations are to tune λ, to exclude hg38 problem regions, and to smooth when a bin-level copy-number profile is the goal.

Status

Version	Content
v1.2.0 (this release)	The hint-rsc command-line tool for your own .hic data (src/hint_rsc/), with tests that check it against the paper.
v1.1.0	Code, small derived inputs and frozen results that reproduce every table, figure and number of the APBC 2026 paper (paper/).

Install

You need conda. The tested platform is macOS with conda packages for osx-64.

git clone https://github.com/changlabtw/HiNT-RSC.git && cd HiNT-RSC
conda env create -f env/environment.yml        # Python 3.11, R 4.5 (mgcv, DNAcopy), htslib, UCSC tools, hic-straw
conda env create -f env/bicseq2.yml            # BIC-seq2 v0.7.2 (kept separate: it pins its own R and Perl)
# Apple Silicon: prefix both commands with CONDA_SUBDIR=osx-64
conda activate hint-rsc
pip install --no-deps -e .                     # installs the hint-rsc command
export BICSEQ="$(conda run -n bicseq2 which NBICseq-seg.pl)"
hint-rsc check                                 # R with mgcv, BIC-seq2, Python modules

Using HiNT-RSC on your own data

The input is a .hic file. HiNT-RSC builds the 50-kb coverage profile and fits HiNT-CNV's GAM, then segments with BIC-seq2. Covariates are bundled for hg19 + MboI and hg38 + HindIII. Problem regions are bundled for hg38: centromeres, assembly gaps and the ENCODE blacklist v2.

# 1. HiNT-CNV baseline (no smoothing, lambda = 3); coverage is cached as out/coverage.tsv (a few minutes)
hint-rsc run --hic sample.hic --genome hg38 --enzyme HindIII -o out/

# 2. Tune the segmentation granularity: every window x lambda, with call counts (and agreement with a reference, if any)
hint-rsc sweep --coverage out/coverage.tsv --genome hg38 --enzyme HindIII --workers 4 -o out/sweep
#    optional: --reference wgs_segments.bed  (chrom, start, end, log2 ratio)

# 3. Final run with the chosen setting
hint-rsc run --coverage out/coverage.tsv --genome hg38 --enzyme HindIII --lambda 1 -o out/lambda1

run writes the following files to the output folder:

File	Content
segments.seg	BIC-seq2 segments (chrom, start, end, bins, observed, expected, log2 ratio, p-value)
cnv_calls.bed	segments with |log2 ratio| ≥ 0.3 (--threshold), labelled gain or loss
profile.tsv	bin-level copy-number profile (residual and log2 ratio per GAM-fitted bin), mean-filtered with --profile-window (default 9; 0 = none)
metrics.tsv	with --reference: segment- and bin-level Pearson/Spearman correlation, matched calls, precision and recall
run.json	all parameters and summary counts

The defaults follow the paper's recommendations:

• Tune λ. The default λ = 3 (HiNT-CNV's default) under-segmented K562. The best unsmoothed λ was 0.5–1.5 across lines, so explore that range with sweep.
• Exclude hg38 problem regions. This is on by default for hg38 (--problem-regions auto). For other genomes, give BED files or none.
• Smooth when a bin-level copy-number profile is the goal. For example, the profile can be the input to CNV-aware Hi-C normalization. Smoothing raised bin-level agreement with WGS, most for shallow libraries. Before segmentation, smoothing (-w/--window, --smoother) is off by default: once λ was tuned, it did not improve CNV calls.

Other options:

• --contacts intra uses intra-chromosomal contacts only. It is faster to compute, at some cost in segment-level agreement.
• --control-hic control.hic --control-fraction 0.3 subtracts a fraction of a control sample's residual.
• --smoother selects one of mean, gaussian, median, savgol, haar or sym4.

For another genome or enzyme, build the covariates once and pass them with --covariates:

bigWigToBedGraph k50.Umap.MultiTrackMappability.bw /dev/stdout | gzip > k50.Umap.bedGraph.gz
hint-rsc covariates --fasta hg38.fa.gz --genome hg38 --enzyme DpnII --mappability k50.Umap.bedGraph.gz -o hg38_DpnII.tsv

--genome also accepts a chromosome-sizes file (chrom, length). hint-rsc coverage and hint-rsc evaluate run the coverage and scoring steps on their own. See hint-rsc <command> --help.

Tests

pip install pytest
pytest tests/          # unit tests; with R (mgcv) and BIC-seq2 available, also the paper regression (about 2 min)

The regression tests run hint-rsc on the shipped PANC-1 and Caki2 inputs and compare every metric with paper/expected/. With HINT_RSC_TEST_HIC=/path/to/ENCFF545SGZ.hic, they also compare the coverage computed from the .hic file with the shipped coverage table. In addition, the following were checked byte for byte against the paper's outputs:

• PANC-1 and Caki2: coverage computed from the ENCODE .hic files, the GAM residuals, the BIC-seq2 input bins and the segmentations.
• K562: the intra-only GAM and segmentation, and the control correction when given the paper's covariate values.

Reproducing the paper

The paper's own scripts are kept unchanged in paper/.

conda activate hint-rsc
export BICSEQ="$(conda run -n bicseq2 which NBICseq-seg.pl)"
cd paper

# Tables from the frozen results (a few minutes, no segmentation)
bash workflow/40_k562_tables_figures.sh                    # Tables 1-6, Supplementary Tables S2-S15, S19, S20, Figure 3
EXPECTED_ONLY=1 bash workflow/50_independent_tables_figures.sh    # Table 7, Supplementary Tables S16-S18

# Full reproduction from the shipped inputs (about 2.5 h with 2 workers)
bash workflow/run_all.sh

run_all.sh compares every result with paper/expected/ and stops at the first difference. See paper/README.md for the workflows, the provenance of every input, and a map from each manuscript table, figure and number to the command and output file. No download is needed. workflow/00_download_data.sh fetches the raw public data only if you want to rebuild the shipped inputs.

Repository layout

src/hint_rsc/   the hint-rsc package (CLI, pipeline, bundled genomes, covariates, hg38 problem regions, R/GAM.R)
tests/          unit tests and the regression against paper/expected/
paper/
  scripts/      the paper's code (Python and R); every script is run from paper/
  workflow/     numbered shell workflows (00 download, 10-50 runs, run_all.sh)
  data/         small derived inputs (K562, PANC-1, Caki2, covariates, hg38 annotations, genome lengths)
  expected/     frozen result tables that the workflows compare against
  output/       created by the workflows (not tracked)
env/            conda environments (environment.yml, bicseq2.yml)

Citation

If you use HiNT-RSC, please cite:

Wei-Han Chen and Jia-Ming Chang. HiNT-RSC: Residual Smoothing and Segmentation Granularity in Hi-C-based Copy Number Variation Detection. APBC 2026 (manuscript under revision).

Please also cite HiNT:

Su Wang et al. HiNT: a computational method for detecting copy number variations and translocations from Hi-C data. Genome Biology 21, 73 (2020). https://doi.org/10.1186/s13059-020-01986-5

To cite the software itself, use its Zenodo DOI: 10.5281/zenodo.23190181 (all versions; v1.2.0 is 10.5281/zenodo.23190182). Machine-readable metadata are in CITATION.cff.

License and attribution

HiNT-RSC is released under the MIT License.

It builds on HiNT (parklab/HiNT), also MIT-licensed (Copyright (c) 2019 Su Wang and contributors). The permission notice is the same as in LICENSE. From HiNT:

• The functions getmappability, getGCpercent and getFragmentsNumber in paper/scripts/load.py are taken from the HiNT source code.
• The GAM model (paper/scripts/GAM.R, src/hint_rsc/R/GAM.R) and the residual min-shift transform (paper/scripts/binlib.py, src/hint_rsc/segment.py) follow HiNT-CNV.
• paper/data/k562/hint_MboI/ and paper/data/covariates/hg19_MboI_cov.txt (bundled as src/hint_rsc/resources/covariates/hg19_MboI.tsv.gz) are output and input files of a HiNT-CNV run.

Third-party data in paper/data/ and src/hint_rsc/resources/ keep their own terms. The DepMap extracts are from DepMap Public 24Q4 (Broad Institute, CC BY 4.0). The other sources (ENCODE, UCSC, ENCODE blacklist, GEO GSE63525) are listed in paper/README.md.

Earlier versions

This work is adapted from the 2022 master's thesis of Wei-Han Chen (National Chengchi University). The thesis notebooks and experiment folders, which are not part of the paper, are no longer on the main branch. They remain available in the git history and under the tag v1.0.0 (the repository as it stood when the paper was submitted).