GPTomics/bio-hi-c-analysis-contact-pairs

Version Compatibility

Reference examples tested with: pairtools 1.1+, bwa 0.7.17+ (or bwa-mem2 2.2+), chromap 0.2+, samtools 1.19+, cooler 0.10+

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags

pairtools defaults have shifted across releases (e.g. parse --max-molecule-size is 750 bp in 1.1.x; dedup --backend defaults to scipy). parse and parse2 report DIFFERENT positions by default - confirm --report-position before mixing outputs. If a command errors, introspect with pairtools <subcommand> --help and adapt rather than retrying.

Hi-C Contact Pairs

"Turn my Hi-C reads into clean contacts and tell me if the library worked" -> Align mates independently through ligation junctions, classify and deduplicate pairs to a 5'-canonical .pairs file, then read the cis/trans and orientation statistics to decide library quality before any matrix is built.

• CLI: bwa mem -SP5M ref.fa R1.fq R2.fq | pairtools parse -c chrom.sizes | pairtools sort | pairtools dedup | pairtools stats

The Single Most Important Modern Insight -- The Read Count Is a Lie Until Pairs Are Classified; Library Quality Is Decided in pairtools, Not in the Aligner

A Hi-C library's usable signal is not "reads sequenced" - it is the uniquely-mapped, deduplicated, long-range cis contacts. Everything between FASTQ and the matrix exists to strip a specific artifact of proximity-ligation chemistry, and the diagnostic ratios from pairtools stats reveal whether the experiment succeeded before any compute is spent binning it. Three load-bearing consequences:

1. % long-range cis is the one-number quality metric; trans is the noise floor. True crosslink-ligation contacts are overwhelmingly cis and distance-decaying. Random ligation between two unrelated molecules in solution is as likely to be trans as cis-far, so trans% is a direct readout of the spurious-ligation floor. A good in-situ human library runs cis>=1kb ~50-65%, inter-chromosomal <10%. But these numbers are genome-size dependent - a yeast or bacterial genome legitimately has higher expected trans (more inter-chromosomal volume per cis distance). Never apply a human trans threshold to a microbe.

2. The ligation junction lives INSIDE the read, so a local/split aligner aligning mates independently is required. A single read often sequences through a ligation junction (locus A | locus B within one read). An end-to-end aligner soft-clips or mis-maps it and the contact is lost. bwa mem -SP5M aligns R1 and R2 as independent single-end reads (-SP skips mate rescue and pairing - proper-pair logic would destroy every long-range and trans contact) and marks the 5'-most chimeric segment primary (-5, the anchor for pairtools' 5' convention). The chimera fraction rises with read length, so on 150bp PE and on Micro-C long reads this is a large, real chunk of contacts.

3. Keep UU AND rescued UC; selecting only UU silently discards every rescued ligation. A naive select pair_type=="UU" throws away the chimeric reads pairtools successfully reconstructed (UC = combined-unique) - a meaningful fraction on long reads. The 4DN standard keeps UU and UC.

Aligner Taxonomy

| Aligner | Role | Hi-C invocation | When | |---------|------|-----------------|------| | bwa mem -SP5M | reference standard; local/split alignment reconstructs in-read junctions | bwa mem -SP5M -t N ref.fa R1 R2 | default; best inter-contig accuracy in benchmarks | | bwa-mem2 | drop-in faster reimplementation, identical output, same flags | bwa-mem2 mem -SP5M -t N ref.fa R1 R2 | when speed matters and the larger index fits RAM | | chromap --preset hic | ultrafast integrated align + dedup + pairs (4DN .pairs out) | chromap --preset hic -x idx -r ref.fa -1 R1 -2 R2 -o out.pairs | ~10x faster scans; trades fine walk-policy control for speed |

-SP5M, letter by letter: -S skip mate rescue; -P skip pairing (no proper-pair rescue); together -SP align mates as independent single-end reads. -5 mark the 5'-most split segment primary (anchors the 5' convention). -M is legacy compatibility only (secondary flag 256 vs supplementary 2048); pairtools handles either - never agonize over -M, never drop -SP5.

Decision Tree by Scenario

| Scenario | Recommended | Why | |----------|-------------|-----| | Standard in-situ/Omni-C, FASTQ -> matrix | bwa mem -SP5M -> parse (5unique) -> sort -> dedup -> stats | the 4DN/distiller default; restriction-agnostic | | Fastest scan, control not critical | chromap --preset hic | integrated align+dedup+pairs ~10x faster | | Multi-way contacts (Pore-C, MC-3C, Micro-C walks) | parse --walks-policy all or parse2 --expand | 5unique COLLAPSES concatemers to pairwise silently | | Micro-C / DNase Hi-C | NO fragment map; do NOT apply a 1kb min-distance cut | sub-1kb (nucleosome ladder) is signal, not artifact | | Arima / Hi-C 3.0 dual-enzyme, fragment-level | fragment file must encode BOTH motifs (restrict -f) | a single-enzyme digest file is silently wrong | | Repeat-heavy genome / stringent loop anchors | raise parse --min-mapq 30 | reclassifies borderline reads M, drops repeat false anchors | | Allele-specific / diploid folding | diploid ref + -XA suboptimal hits -> pairtools phase -> two coolers | needs the suboptimal-score gap to resolve haplotypes | | Decide whether to sequence deeper | complexity curve from dup model / preseq lc_extrap | a bare dup% without depth is meaningless | | Build the matrix from clean pairs | -> hic-data-io (cooler cload pairs) | binning happens after classification/dedup | | Annotate boundary/anchor coordinates | -> genome-intervals/bed-file-basics | pairs are 1-based, half-open conventions differ |

Align: Mates as Independent Single-End Reads

bwa index ref.fa                                            # or bwa-mem2 index ref.fa
bwa mem -SP5M -t 16 ref.fa R1.fq.gz R2.fq.gz | \            # -SP: independent SE; -5: 5' segment primary
    samtools view -b -@ 8 - > aligned.bam
# chromap fast path (integrated align + dedup + 4DN pairs, no separate pairtools needed):
# chromap -i -r ref.fa -o idx && \
# chromap --preset hic -x idx -r ref.fa -1 R1.fq.gz -2 R2.fq.gz -o sample.pairs

Parse, Sort, Dedup, Select: the pairtools Core

# Parse alignments into a 5'-canonical .pairs. min-mapq 1 (default) is permissive: only MAPQ-0 is "multi".
pairtools parse -c chrom.sizes --walks-policy 5unique --min-mapq 1 \
    --add-columns mapq --drop-sam aligned.bam | \
pairtools sort --nproc 8 | \                                # block-sort; flips to upper-triangular (5'-canonical)
pairtools dedup --max-mismatch 3 --mark-dups \              # within 3bp on both sides = duplicate; tag DD
    --output-stats sample.dedup.stats | \
pairtools select '(pair_type=="UU") or (pair_type=="UC")' \ # keep both-unique AND rescued chimeric
    -o sample.valid.pairs.gz

Dedup MUST run on a flipped, 5'-canonical file (sort does the flip); on a non-canonical file dedup under-collapses and dup% reads falsely low. --max-molecule-size (750 bp in 1.1.x) governs single-ligation chimera rescue; --max-inter-align-gap (20 bp) sets when a coverage gap becomes a null alignment.

Library QC: the Decision This Skill Owns

pairtools stats is the canonical readout. Read it as a funnel, not a single number.

pairtools stats --bytile-dups -o sample.stats.tsv sample.valid.pairs.gz
# Key fields: frac_dups; frac_cis; cis_1kb+/cis_20kb+; trans; pair_types; dist_freq orientation FF/FR/RF/RR.

• % long-range cis (cis>=1kb, often cis>=20kb) = signal quality. trans = noise floor (genome-size dependent).
• Orientation vs distance = fragment-map-free dangling-end/self-circle QC. Above ~1kb the four orientations FF/FR/RF/RR each converge to ~25% (random, the positive QC signal). A short-range FR (inward) spike = dangling ends / undigested / self-ligation; a short-range RF (outward) spike = self-circles / religation. The distance where orientations equalize is the minimum reliable contact distance - derive the min-distance cut from this plot, do not hardcode 1kb (Micro-C structure lives below 1kb).
• % duplicates = complexity proxy, but --bytile-dups separates OPTICAL dups (patterned NovaSeq flowcells, same tile, adjacent coordinates) from PCR dups. Only the PCR fraction reflects library complexity; reading total dup% as over-amplification wrongly condemns a good library. A bare dup% without the depth it was measured at is meaningless - use the complexity/yield curve (preseq lc_extrap) to decide whether deeper sequencing buys uniques or duplicates.

Apply the QC-derived distance cut without a fragment map:

pairtools select '(chrom1!=chrom2) or (abs(pos2-pos1) > 1000)' \   # keep trans + cis beyond the orientation-equalization distance
    -o sample.filtered.pairs.gz sample.valid.pairs.gz

Restriction Fragments: Opt-In, Not Default

Modern pipelines SKIP fragment filtering on purpose - the distance cut + dedup + UU/UC filter + balancing absorb the residual, and a digest file is one genome-specific place to mis-specify the enzyme. pairtools restrict -f frags.bed is opt-in for: sub-kb / restriction-fragment-resolution maps, capture-Hi-C / 4C-style fragment analysis, and bench QC where the dangling-end fraction is the digestion-efficiency readout.

cooler digest --out frags.bed chrom.sizes ref.fa DpnII     # single-enzyme: DpnII ^GATC
pairtools restrict -f frags.bed -o restricted.pairs.gz parsed.pairs.gz

Arima dual-enzyme has FOUR junction motifs (GATCGATC, GANTGATC, GANTANTC, GATCANTC); a single-enzyme (DpnII-only) digest file silently mis-assigns fragments. Micro-C / DNase Hi-C have NO fragment map (MNase/DNase cut sequence-nonspecifically) - any tool requiring a restriction file cannot process them, which is exactly why the restriction-agnostic pairtools path became the field default.

Allele-Specific Contacts: pairtools phase -> Two Coolers

Goal: Resolve each contact to maternal vs paternal haplotype for allele-specific 3D folding.

Approach: Align to a diploid reference reporting suboptimal hits, so each read carries its best alignment on both homologs; pairtools phase reads the gap between the two best alignment scores to tag each side resolved-hap1 / resolved-hap2 / non-resolved / multi-mapper - the score gap is what separates a genuinely uninformative read from an actual repeat.

bcftools consensus -H 1 -f ref.fa phased.vcf.gz > hap1.fa   # build a diploid (two-homolog) reference
bcftools consensus -H 2 -f ref.fa phased.vcf.gz > hap2.fa
bwa mem -SP5M ref_diploid.fa R1.fq R2.fq | \                # report suboptimal hits so both homologs are kept
pairtools parse -c chrom.sizes --add-columns XB,AS,XS | \
pairtools phase --phase-suffixes _hap1 _hap2 --tag-mode XB | \
pairtools sort | pairtools dedup -o phased.pairs.gz
# Then split into hap1/hap2/trans pairs and cload each into a SEPARATE cooler.

Per-Method Failure Modes

Aligned Hi-C as a normal PE library

Trigger: plain bwa mem without -SP, or proper-pair logic. Mechanism: mate rescue forces the shotgun insert model on mates from different loci. Symptom: trans% collapses, compartments vanish, sparse map. Fix: bwa mem -SP5M, mates aligned independently.

Dropped -5

Trigger: copying a pre-2016 -SP command lacking -5. Mechanism: the 5'-most chimeric segment is not primary, so pairtools picks an inconsistent anchor. Symptom: degraded flip/dedup, smeared loops. Fix: always -SP5M (or -SP5).

walks-policy 5unique on a multi-way protocol

Trigger: Pore-C/MC-3C/Micro-C walks parsed with the default. Mechanism: 5unique reports only the two 5'-most alignments, collapsing concatemers to pairwise. Symptom: "we found few multi-contacts." Fix: --walks-policy all or parse2 --expand.

Mixing parse and parse2 outputs

Trigger: combining a parse2 (.pairs, default outer/junction-anchored) file with a parse (5'-anchored) file. Mechanism: the two report positions by different conventions. Symptom: coordinates shift by the alignment length; dedup under-collapses; loops smear. Fix: one parser/convention per project; prefer plain parse if walks are not needed.

Selecting only UU

Trigger: select pair_type=="UU". Mechanism: discards rescued chimeric (UC) pairs. Symptom: lower valid-pair yield, especially on long reads. Fix: keep (pair_type=="UU") or (pair_type=="UC").

Dedup on a non-canonical file

Trigger: dedup run before sort/flip, or on mixed parse/parse2 positions. Mechanism: duplicates are not in canonical coordinates. Symptom: dup% reads falsely low - library looks better than it is. Fix: sort (flips to 5'-canonical) before dedup.

Optical dups read as PCR dups

Trigger: total dup% on a patterned flowcell taken as library complexity. Mechanism: optical (same-tile) dups inflate apparent PCR rate. Symptom: a good library condemned as over-amplified. Fix: --bytile-dups / --output-bytile-stats; judge complexity on the PCR fraction only.

Fixed 1kb cut on Micro-C

Trigger: applying Hi-C's >1kb min-distance cut to Micro-C. Mechanism: Micro-C's nucleosome-ladder signal lives below 1kb. Symptom: the structure Micro-C exists to capture is erased. Fix: derive the cut from the orientation-vs-distance plot per library.

Quantitative Thresholds

| Threshold | Source | Rationale | |-----------|--------|-----------| | cis>=1kb ~50-65% of nodup pairs (good in-situ human) | Dovetail/Arima QC guidance (~approx) | long-range cis is the signal; library- and genome-size dependent | | inter-chromosomal (trans) <10% (clean), 20-30% acceptable | in-situ Hi-C practice | trans is the spurious-ligation floor; NEVER apply to small genomes | | FF/FR/RF/RR -> ~25% each above ~1kb | random strand combination at true contacts | convergence is the fragment-map-free positive QC signal | | parse --min-mapq 1 default, raise to 30 for stringency | pairtools default | 1 drops only MAPQ-0; 30 removes repeat-driven false anchors | | dedup --max-mismatch 3 bp | pairtools default | tolerates mapping wobble; 0 over-splits, larger over-collapses complexity | | parse --max-molecule-size 750 bp (1.1.x) | pairtools default | bound on single-ligation chimera rescue; revisit for unusual size selection | | min-distance cut ~1kb (Hi-C), derive from orientation plot | orientation-equalization distance | the cut is protocol-specific; Micro-C signal is sub-1kb |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | trans% high, no compartments | aligned without -SP (proper-pair logic) | re-align bwa mem -SP5M | | Few multi-way contacts on Pore-C/Micro-C | default --walks-policy 5unique collapsed walks | --walks-policy all / parse2 --expand | | Loops smeared, dedup under-collapses | mixed parse/parse2 position conventions | one parser per project; sort before dedup | | Valid-pair yield lower than expected | select kept only UU | keep UU and UC | | dup% suspiciously low | dedup ran before sort/flip | sort to 5'-canonical first | | dup% high on NovaSeq, complexity looks bad | optical dups counted as PCR | --bytile-dups; judge on PCR fraction | | Micro-C structure disappears after filtering | fixed 1kb min-distance cut | derive cut from orientation-vs-distance | | Fragment assignment wrong on Arima data | single-enzyme digest file | encode all four Arima junction motifs | | phase resolves nothing | aligned to a haploid reference | diploid ref + suboptimal (-XA) alignments |

References

• Open2C, Abdennur N, Fudenberg G, Flyamer IM, Galitsyna AA, Goloborodko A, Imakaev M, Venev SV. 2024. Pairtools: from sequencing data to chromosome contacts. PLoS Comput Biol 20(5):e1012164.
• Li H. 2013. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. arXiv:1303.3997.
• Zhang H, Song L, Wang X, et al. 2021. Fast alignment and preprocessing of chromatin profiles with Chromap. Nat Commun 12:6566.
• Durand NC, Shamim MS, Machol I, et al. 2016. Juicer provides a one-click system for analyzing loop-resolution Hi-C experiments. Cell Syst 3:95-98.
• Servant N, Varoquaux N, Lajoie BR, et al. 2015. HiC-Pro: an optimized and flexible pipeline for Hi-C data processing. Genome Biol 16:259.
• Akgol Oksuz B, Yang L, Abraham S, et al. 2021. Systematic evaluation of chromosome conformation capture assays. Nat Methods 18:1046-1055.
• Krietenstein N, Abraham S, Venev SV, et al. 2020. Ultrastructural details of mammalian chromosome architecture (Micro-C). Mol Cell 78:554-565.
• Ramani V, Cusanovich DA, Hause RJ, et al. 2016. Mapping 3D genome architecture through in situ DNase Hi-C. Nat Protoc 11:2104-2121.
• Abdennur N, Mirny LA. 2020. Cooler: scalable storage for Hi-C data and other genomically labeled arrays. Bioinformatics 36:311-316.
• Daley T, Smith AD. 2013. Predicting the molecular complexity of sequencing libraries (preseq). Nat Methods 10:325-327.

Related Skills

• hic-data-io - Bins the deduplicated valid pairs into a .cool/.mcool matrix
• matrix-operations - Balancing and O/E that the binned pairs feed into
• hic-visualization - Render contact maps from the resulting cooler
• read-alignment/bwa-alignment - Aligner upstream; this skill adds the Hi-C -SP5M idiom
• alignment-files/duplicate-handling - General duplicate-marking context for the pairtools dedup step
• genome-intervals/bed-file-basics - Coordinate/digest BED handling for restriction fragments and anchors
• genome-assembly/scaffolding - Same Hi-C reads used to order contigs into chromosomes