End-to-end Hi-C analysis workflow from FASTQ to compartments, TADs, and loops, with the decision of WHICH features the sequencing depth can support. Covers pairtools read-pair processing and library QC, cooler matrices, ICE balancing and distance-decay expected, A/B compartments, TAD boundaries, loop calling, and the routing of HiChIP/PLAC-seq/Capture Hi-C to protein-directed loop callers. Use when processing Hi-C data end to end, deciding a resolution for a given depth, or choosing between bulk-Hi-C and protein-directed loop calling.
End-to-end Hi-C analysis workflow from FASTQ to compartments, TADs, and loops, with the decision of WHICH features the sequencing depth can support. Covers pairtools read-pair processing and library QC, cooler matrices, ICE balancing and distance-decay expected, A/B compartments, TAD boundaries, loop calling, and the routing of HiChIP/PLAC-seq/Capture Hi-C to protein-directed loop callers. Use when processing Hi-C data end to end, deciding a resolution for a given depth, or choosing between bulk-Hi-C and protein-directed loop calling.
[{"after_pairs":"Long-range cis (>=20kb) fraction, not just %valid; trans is genome-size-dependent"},{"after_balance":"balance=True returns finite weights; masked bins are NaN by design"},{"after_analysis":"Eigenvector sign phased by GC; feature scale matches the resolution"}]
Before using code patterns, verify installed versions match. If versions differ:
Python: pip show <package> then help(module.function) to check signatures
CLI: <tool> --version then <tool> --help to confirm flags
If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.
Hi-C Pipeline
"Analyze my Hi-C data from FASTQ to 3D genome features" -> Process read pairs and judge library quality, build and balance a cooler, then call ONLY the features the depth can support: compartments are cheap, TADs need moderate depth, de-novo loops need billions of contacts.
Complete workflow for Hi-C chromosome conformation capture analysis.
The Decision That Frames the Whole Pipeline -- depth dictates the feature
Contacts scale with the SQUARE of the bin count, so the affordable resolution is set by depth, not ambition. Read hi-c-analysis/matrix-operations for the budget; the rule of thumb is ~1000 contacts/bin. Compartments (100kb-1Mb) come from almost any library; TAD boundaries (10-40kb) need a moderate map; de-novo loop calling (5-10kb dots) needed ~5 billion contacts in Rao 2014. On a shallow map, do NOT de-novo call loops -- run aggregate peak analysis (APA) on known CTCF/cohesin anchors instead (hi-c-analysis/loop-calling).
Protein-directed assays branch here: HiChIP, PLAC-seq, and Capture Hi-C have non-uniform peak-anchored coverage, so generic dots/HiCCUPS use the wrong null. Route them to hi-c-analysis/hichip-plac-loops (FitHiChIP/MAPS/CHiCAGO), NOT to step 6 below.
Workflow Overview
Hi-C FASTQ files
|
v
[1. Alignment & Pairs] --> bwa-mem2 -SP5M + pairtools (parse/sort/dedup/split)
| QC: long-range cis fraction is the one-number readout
v
[2. Matrix Generation] --> cooler cload + zoomify (sum RAW, re-balance per resolution)
|
v
[3. Balancing] --------> ICE (cooler balance); REQUIRED before any analysis
|
v
[4. Compartments 100kb] -> eigs_cis, sign-phased by GC (E1 is a choice, not an output)
|
v
[5. TADs 10kb] ---------> insulation score across a window sweep (boundaries, not domains)
|
v
[6. Loops 10kb] --------> cooltools dots IF deep; else APA on known anchors
|
v
Hi-C features (compartments / boundaries / loops)
Step 1: Alignment and Pair Processing
Goal: Turn raw Hi-C FASTQ into a deduplicated, classified list and judge whether the library worked.
.pairs
Approach: Align the two mates independently with bwa-mem2 -SP5M (proper pairing would destroy long-range contacts), then parse, sort, deduplicate, and split with pairtools, reading the long-range cis fraction as the go/no-go.
# Pass BOTH mates to ONE bwa-mem2 call. -SP5M: -S/-P make bwa treat the mates as single-end and# skip proper-pair rescue (so long-range contacts survive), while both sides are still emitted for# pairtools to form the pair; -5 reports the 5'-most alignment of split reads, -M flags secondaries.
bwa-mem2 mem -SP5M -t 16 reference.fa reads_R1.fastq.gz reads_R2.fastq.gz | \
pairtools parse --min-mapq 40 --walks-policy 5unique \
--max-inter-align-gap 30 --nproc-in 8 --nproc-out 8 \
--chroms-path reference.genome | \
pairtools sort --nproc 16 --tmpdir ./tmp | \
pairtools dedup --nproc-in 8 --nproc-out 8 \
--mark-dups --output-stats stats.txt | \
pairtools split --nproc-in 8 --output-pairs sample.pairs.gz
QC Checkpoint: read pairtools stats as the go/no-go. The one-number readout is the LONG-RANGE cis fraction (>=20kb), not bare %valid: short-range cis is inflated by dangling ends and self-circles. Trans fraction is a noise floor but its acceptable value is genome-size-dependent (a human <10% threshold is meaningless for a microbe). High duplicate rate = low library complexity (not rescuable by sequencing deeper). See hi-c-analysis/contact-pairs for the orientation-balance QC and the Micro-C/Arima variants.
Goal: ICE-balance the matrix so every bin has equal marginal coverage, without letting empty/artifact bins corrupt the result.
Approach: Mask low-coverage and blacklist bins BEFORE balancing, then balance per resolution. ICE assumes equal visibility per bin, so an unmasked empty or repeat/blacklist bin is iteratively up-weighted into a bright stripe artifact; mad_max filters bins whose coverage is mad_max MADs below the median, and a blacklist/--blacklist (or pre-masking bad bins) removes known artifacts. Balancing is REQUIRED before any analysis, but it does NOT make two maps comparable across conditions — that needs depth-matching + distance-stratified normalization (hi-c-analysis/hic-differential).
import cooler
import cooltools
# Mask before balancing: mad_max drops low-coverage bins that would otherwise become stripes.# Balance EVERY resolution the downstream steps analyze -- weights are resolution-specific, and an# unbalanced cooler has no 'weight' column, so cooltools (eigs_cis, insulation, dots) fails on it.# Steps 4-6 below use 100kb (compartments) and 10kb (loops and insulation).for res in (10000, 25000, 100000):
clr = cooler.Cooler(f'sample.mcool::/resolutions/{res}')
cooler.balance_cooler(clr, store=True, mad_max=5, ignore_diags=2, min_nnz=10) # masked bins are NaN by design# CLI equivalent (add --blacklist regions.bed to remove known-artifact bins first):# for res in 10000 25000 100000; do cooler balance --mad-max 5 --ignore-diags 2 --min-nnz 10 sample.mcool::/resolutions/${res}; done
Step 4: Compartment Analysis
Goal: Assign each genomic bin to the active (A) or inactive (B) compartment with a non-arbitrary sign.
Approach: At a coarse 100kb resolution, compute the cis eigenvector and orient it with a GC phasing track so positive E1 is the active compartment (the sign is arbitrary without it).
import cooler
import cooltools
import bioframe
import numpy as np
# Compartments are coarse-scale: 100kb, balanced matrix
clr = cooler.Cooler('sample.mcool::/resolutions/100000')
# Phasing track is NOT optional: the eigenvector sign is arbitrary. A GC track# (matching the cooler binning exactly) orients positive E1 to the active (A) compartment.
view_df = bioframe.make_viewframe(clr.chromsizes)
gc = bioframe.frac_gc(clr.bins()[:][['chrom', 'start', 'end']], bioframe.load_fasta('reference.fa'))
eig_values, eig_vectors = cooltools.eigs_cis(clr, gc, view_df=view_df, n_eigs=3)
compartments = eig_vectors[['chrom', 'start', 'end', 'E1']].copy()
# Masked bins have E1 = NaN; NaN > 0 is False, so guard or a bare np.where mislabels them all 'B'.
compartments['compartment'] = np.where(compartments['E1'].isna(), None, np.where(compartments['E1'] > 0, 'A', 'B'))
compartments.to_csv('compartments.tsv', sep='\t', index=False)
Step 5: TAD Detection
Goal: Locate domain boundaries at the sub-Mb scale.
Approach: Compute the insulation score across a window sweep at 10kb and read the is_boundary/boundary_strength columns the function returns directly (report boundaries, not a fixed domain partition).
import cooltools
# Load matrix at TAD resolution
clr = cooler.Cooler('sample.mcool::/resolutions/10000')
# Insulation across a window sweep; the function already returns boundary columns# (is_boundary_<W>, boundary_strength_<W>) -- there is no separate find_boundaries call.
ins = cooltools.insulation(clr, window_bp=[100000, 200000, 500000])
# Boundaries at the 200kb window; keep the continuous strength (comparable across samples).# is_boundary is NaN for bad/low-mappability bins; fillna(False) before masking or pandas raises.
boundaries = ins[ins['is_boundary_200000'].fillna(False).astype(bool)][['chrom', 'start', 'end', 'boundary_strength_200000']]
boundaries.to_csv('tad_boundaries.tsv', sep='\t', index=False)
# Alternative: use HiCExplorer# hicFindTADs -m sample.cool --outPrefix tads --correctForMultipleTesting fdr
Step 6: Loop Calling
Goal: Detect focal CTCF/enhancer-promoter contacts, but only when the map is deep enough to support de-novo calling.
Approach: Compute a distance-matched expected, then run cooltools dots on a deep map; on a shallow library, skip de-novo calling and run APA on known anchors instead.
import cooltools
# Load high-resolution matrix
clr = cooler.Cooler('sample.mcool::/resolutions/10000')
# De-novo dot calling is only honest on a DEEP map (Rao 2014 used ~5B contacts).# On a shallow library, skip this and run APA on known anchors (see loop-calling).
expected = cooltools.expected_cis(clr)
loops = cooltools.dots(clr, expected, max_loci_separation=2000000, nproc=4)
loops.to_csv('loops.tsv', sep='\t', index=False)
# Alternative caller (template matching): chromosight# chromosight detect --pattern loops sample.mcool::/resolutions/10000 loops# For HiChIP/PLAC-seq/Capture Hi-C do NOT use dots -> hi-c-analysis/hichip-plac-loops
Step 7: Visualization
import matplotlib.pyplot as plt
from matplotlib.colors import LogNorm
import cooltools.lib.plotting # registers the 'fall' cmap; if unavailable in your stack, use 'afmhot_r'# A square balanced map on a log scale; importing cooltools.lib.plotting registers 'fall'.# Show O/E with a symmetric diverging cmap to see compartments/loops (see hic-visualization).
mat = clr.matrix(balance=True).fetch('chr1:50000000-60000000')
fig, ax = plt.subplots(figsize=(8, 8))
ax.matshow(mat, norm=LogNorm(vmax=mat[mat > 0].max() * 0.5), cmap='fall')
plt.savefig('hic_matrix.pdf')
# Triangle/track-stacked browser views: use pyGenomeTracks or HiCExplorer hicPlotTADs# (data-visualization/genome-tracks), not a hand-rolled rotation.
Long-range contacts missing / map looks like short-range only
Mates aligned as a proper pair instead of independently
Align with bwa-mem2 mem -SP5M (each end separately, no proper-pair rescue)
Library "passed" %valid but is unusable
Judged on bare %valid; short-range cis is inflated by dangling ends/self-circles
Read the long-range cis (>=20kb) fraction as the go/no-go
Bright stripes/plaid artifacts after balancing
Empty/repeat/blacklist bins not masked before ICE
Mask with mad_max/--blacklist/min_nnz before balancing
A/B compartments flipped between samples
Eigenvector sign is arbitrary without phasing
Phase E1 by a GC track matching the cooler binning exactly
De-novo loops look sparse/noisy
Called dots on a shallow map
Only de-novo call on deep maps (~billions of contacts, Rao 2014); else APA on known anchors
Cross-condition differences dominated by depth
Compared balanced maps directly
Depth-match + distance-stratified normalize first (hi-c-analysis/hic-differential)
HiChIP/PLAC "loops" full of false positives
Generic dots/HiCCUPS null on peak-anchored coverage
Route to FitHiChIP/MAPS/CHiCAGO (hi-c-analysis/hichip-plac-loops)
References
Rao SSP, Huntley MH, Durand NC, et al (2014) A 3D map of the human genome at kilobase resolution reveals principles of chromatin looping. Cell 159:1665-1680. DOI 10.1016/j.cell.2014.11.021. (depth-vs-resolution; ~5B contacts for kilobase loops.)
Imakaev M, Fudenberg G, McCord RP, et al (2012) Iterative correction of Hi-C data reveals hallmarks of chromosome organization. Nature Methods 9:999-1003. DOI 10.1038/nmeth.2148. (ICE balancing.)
Abdennur N, Mirny LA (2020) Cooler: scalable storage for Hi-C data and other genomically labeled arrays. Bioinformatics 36:311-316. DOI 10.1093/bioinformatics/btz540.
Open2C, Abdennur N, Fudenberg G, et al (2024) Cooltools: enabling high-resolution Hi-C analysis in Python. PLOS Computational Biology 20:e1012067. DOI 10.1371/journal.pcbi.1012067.
Open2C, Abdennur N, Fudenberg G, et al (2024) Pairtools: from sequencing data to chromosome contacts. PLOS Computational Biology 20:e1012164. DOI 10.1371/journal.pcbi.1012164.
Related Skills
hi-c-analysis/contact-pairs - Read-pair processing and the library-QC decision
hi-c-analysis/hic-data-io - Cooler file operations and format conversion
hi-c-analysis/matrix-operations - ICE balancing, expected/P(s), and the resolution-vs-depth budget
hi-c-analysis/compartment-analysis - Sign-phased A/B compartments and saddle strength
hi-c-analysis/tad-detection - Insulation-score boundaries across a window sweep
hi-c-analysis/loop-calling - Dot calling and APA validation