Bio Long Read Sequencing Haplotype Phasing

Version Compatibility

Reference examples tested with: whatshap 2.3+, longphase 1.7+, samtools 1.19+, tabix/htslib 1.19+.

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

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

Behavior to record:

• whatshap phase --reference enables realignment mode (rescues indel phasing on error-prone long reads); omitting it falls back to lower-quality genotype-only phasing.
• Phasing the VCF and haplotagging the BAM are SEPARATE steps; downstream read-level tools need the BAM HP tag.
• --max-coverage 15 (WhatsHap) is a runtime downsampling cap, not a minimum-depth requirement.

If code throws an error, introspect the installed tool (whatshap phase --help, longphase --help) and adapt the example to the actual API rather than retrying.

Read-Backed Haplotype Phasing

"Phase my long-read variants" -> Reconstruct haplotypes directly from reads that span heterozygous sites, then haplotag the BAM so downstream tools can see the phase.

• CLI: whatshap phase -o phased.vcf.gz --reference ref.fa --indels variants.vcf.gz aln.bam then whatshap haplotag -o haplotagged.bam --reference ref.fa phased.vcf.gz aln.bam

This is read-backed (physical, panel-free) phasing of a single sample. Statistical/reference-panel phasing for imputation lives in phasing-imputation/haplotype-phasing; building phased haplotype contigs lives in genome-assembly/hifi-assembly.

The Single Most Important Modern Insight -- Phasing the VCF Is Useless Until the BAM Is Haplotagged, and Blocks Break Where No Read Spans Two Hets

Read-backed phasing is sample-intrinsic and panel-free (the haplotypes are exactly what this individual's reads physically witness), with two load-bearing consequences:

1. phase writes the VCF; haplotag writes the BAM - they are different products. whatshap phase / longphase phase set GT pipe (0|1) and a PS phase-set in the VCF; they do NOT touch the BAM. Every read-level downstream tool (modkit --partition-tag HP for allele-specific methylation, pb-CpG-tools --hap-tag HP, Severus for phased SVs, IGV color-by-HP, whatshap split) keys on the per-read HP tag that ONLY haplotag writes. A user who runs phase and stops has a phased VCF and an un-haplotagged BAM, and the downstream step silently produces only an ungrouped partition. Triage: samtools view haplotagged.bam | grep -m1 'HP:i:'.
2. Phase blocks break wherever no single read spans two adjacent hets. Block length is capped by read length x heterozygosity: a long homozygous run (or a coverage/mapping dropout) ends a block no matter how long the reads are. A genome is phased into MANY blocks, not one haplotype per chromosome, and between-block phase is arbitrary until a long-range method (trio, Hi-C) stitches them. So "the genome is phased" is meaningless without block N50 AND switch error together.

Tool Decision Tree

Switch-error accuracy is comparable across read-based tools (~0.1-0.4% on long reads); choose on speed, SV/mod co-phasing, platform, and pedigree.

Clair3 uses WhatsHap (or LongPhase) internally to phase its het SNPs and haplotag the BAM feeding its full-alignment model - this skill owns that phase->haplotag mechanism (see clair3-variants).

The Tags (the central distinction)

Unassigned reads carry NO HP tag (not HP:i:0). Do not confuse the VCF HP FORMAT tag (GATK style) with the BAM HP read tag.

Phasing Quality - Report Block N50 AND Switch Error

Long blocks with a high switch rate are worse, not better, than honest short blocks. Benchmark against a trio-/strand-seq-phased GIAB truth.

Core Commands

# WhatsHap: phase (VCF), then haplotag (BAM). --reference enables realignment for indels.
whatshap phase -o phased.vcf.gz --reference ref.fa --indels variants.vcf.gz aln.bam
tabix -p vcf phased.vcf.gz
whatshap haplotag -o haplotagged.bam --reference ref.fa \
    --output-haplotag-list htlist.tsv.gz phased.vcf.gz aln.bam
samtools index haplotagged.bam

# Quality
whatshap stats --gtf blocks.gtf phased.vcf.gz                       # block N50, count, fraction
whatshap compare --names truth,mine truth.vcf.gz phased.vcf.gz      # switch error, flip decomposition

# Trio (gold standard) - --ped takes a PED file, not mother/father/child args
whatshap phase -o trio.vcf.gz --reference ref.fa --ped family.ped joint.vcf.gz mother.bam father.bam child.bam

# LongPhase: faster whole-genome, co-phase SNP+indel+SV(+5mC) into long blocks
longphase phase -s snps.vcf --indels --sv-file svs.vcf -b aln.bam -r ref.fa -o phased -t 16 --ont
longphase haplotag -s phased.vcf --sv-file phased_SV.vcf -b aln.bam -r ref.fa -o haplotagged -t 16

# Downstream consumer example: allele-specific methylation
modkit pileup haplotagged.bam asm/ --ref ref.fa --cpg --combine-strands --partition-tag HP

Per-Method Failure Modes

Phased VCF but no HP tags downstream

Trigger: running phase and pointing a read-level tool at the original BAM. Mechanism: phase writes the VCF only; the BAM HP tag comes from haplotag. Symptom: modkit returns only an ungrouped partition; IGV shows one color; Severus reports no phased SVs - all with no error. Fix: run haplotag; verify samtools view ... | grep HP:i:.