GPTomics/bio-genome-assembly-hifi-assembly

Version Compatibility

Reference examples tested with: hifiasm 0.25.0+, yak 0.1+, verkko 2.3+, gfatools 0.5+, meryl 1.4+.

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

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

hifiasm output filenames AND the default purge level are version-dependent: confirm the .bp.*/.dip.* prefixes and -l defaults against hifiasm --help and the man page for the installed build before scripting around them. The k-mer-size convention also matters more than the binary version: hifiasm/yak trio uses k=31; verkko/Merqury hap-mers commonly use k=30 - mixing them silently produces garbage hap-mer matching. If a command errors, introspect the installed tool and adapt rather than retrying.

HiFi Assembly

"Assemble a diploid genome from HiFi reads" -> Build two cleanly phased haplotypes (not one mosaic primary) from accurate long reads, choosing the phasing mechanism from the sample type and available data, and validate phasing with k-mer QC the headline metrics cannot see.

• CLI: hifiasm -o prefix -t 32 reads.hifi.fq.gz (HiFi-only), --h1/--h2 (Hi-C), -1/-2 *.yak (trio); verkko --hifi ... --nano ... (T2T)

The Single Most Important Modern Insight -- A Primary Assembly Is a Mosaic Chimera, Not a Haplotype

A "primary assembly" is not a genome that exists in any cell. At every heterozygous block the assembler picks one allele, and which one it picks switches arbitrarily from block to block - so the primary is a stitched chimera: maternal here, paternal there, matching no gamete, no parent, no individual. It is a fine haploid representation for "roughly where is gene X" and a terrible substrate for anything allele-aware (phased variant calling, allele-specific expression, HLA/KIR typing, compound-heterozygote analysis). HiFi's combination of length AND ~Q30 accuracy made phased diploid assembly routine, so the field's deliverable shifted from one collapsed reference to two phased haplotypes (and onward to the pangenome: each HPRC node is a per-sample phased diploid assembly, Liao 2023 Nature 617:312).

The killer corollary: the thing that is wrong with a mosaic - that it is a haplotype mix - is exactly the thing none of the metrics people check can see. A switch error does not break a contig (N50 unchanged), delete a gene (BUSCO unchanged), or introduce a wrong base (QV unchanged - both alleles are real sequence, just assigned to the wrong haplotype). Switch errors and haplotype mosaicism are structurally invisible to N50, BUSCO, and even base-level QV. Only k-mer/trio QC - Merqury hap-mer blob plots and switch/hamming error against parental k-mers - can see them. "Reported N50 and BUSCO but no switch/hamming or hap-mer plot" means the phasing was never validated. Subtle trap: the HiFi-only .bp.hap1/hap2 are partially phased (locally phased, switch errors between blocks) - "hap1" in a filename does NOT certify global phasing; the phasing data supplied (trio or Hi-C) does.

Tool Taxonomy

| Tool | Citation | Role | When | |------|----------|------|------| | hifiasm | Cheng 2021 Nat Methods | phased string-graph HiFi assembler; built-in purging | the default for diploid HiFi; HiFi-only / Hi-C / trio / --ul | | hifiasm (Hi-C) | Cheng 2022 Nat Biotechnol | global phasing from proximity-ligation, no parents | diploid, Hi-C available, parents unavailable (the broad default) | | yak | (Li, hifiasm suite) | parental haplotype-specific k-mer DBs for trio | feeds hifiasm -1/-2; trio gold-standard phasing | | verkko | Rautiainen 2023 Nat Biotechnol | HiFi + ultralong-ONT graph assembler (MBG -> GraphAligner -> rukki) | T2T-grade; trio or Hi-C phasing; reference-quality | | verkko2 | Antipov 2025 Genome Res | adds proximity-ligation phasing into the De Bruijn graph | T2T with Hi-C; ~doubled T2T-scaffold yield | | HiCanu | Nurk 2020 Genome Res | HiFi Canu fork; segmental dups/satellites/allelic variants | legacy/SD-focused; superseded by hifiasm for routine diploid | | meryl/Merqury | Rhie 2020 Genome Biol | hap-mer DBs; k-mer QV, completeness, switch/hamming, blob plots | the only QC that sees phasing (-> assembly-qc) |

Decision Tree by Scenario

| Scenario | Recommended | Why | |----------|-------------|-----| | Sample type/heterozygosity unknown | profile first: k-mer spectrum (GenomeScope) for genome size + heterozygosity | purge setting, n-hap, and whether to phase are all downstream of outbred-vs-inbred | | Outbred diploid, HiFi only, quick draft | hifiasm default | primary + partially-phased .bp.hap1/hap2; partial phasing is the cost of no linkage data | | Diploid, HiFi + Hi-C, no parents | hifiasm --h1/--h2 (Cheng 2022) | global phasing from the sample itself; the pragmatic default for the non-human bestiary | | Diploid, HiFi + both parents | hifiasm trio -1 pat.yak -2 mat.yak | gold standard - read-level phasing, no switch ambiguity where parents are informative | | Inbred / doubled-haploid / mole | hifiasm -l0 (purging OFF) | nothing to phase; default purging would DELETE real segmental duplications | | Unbalanced hap1/hap2 size after a run | re-run with --hom-cov set to the k-mer peak | the size imbalance is a mis-estimated coverage alarm, not a phasing result | | T2T-grade reference, have HiFi + ultralong ONT + trio/Hi-C | verkko (or hifiasm --ul for speed) | only graph + UL spanning reaches gapless centromeres; T2T is a project, not a flag | | No ultralong ONT but want a reference | hifiasm-HiFi-only is the sensible stop | neither tool reaches T2T without UL reads to span repeats | | One of 50 individuals in a diversity panel | phased diploid (primary may suffice) | do NOT chase T2T per-sample; match grade to question | | Reads not yet QC'd | -> long-read-sequencing/long-read-qc | HiFi length/QV/contamination cap assembly quality |

hifiasm Invocations (the fragile commands - run as written)

# HiFi-only (default): primary mosaic + PARTIALLY-phased hap1/hap2
hifiasm -o prefix -t 32 reads.hifi.fq.gz

# Hi-C phased (no parents) - Cheng 2022; global phasing
hifiasm -o prefix -t 32 --h1 hic_R1.fq.gz --h2 hic_R2.fq.gz reads.hifi.fq.gz

# Trio phased (gold standard) - build parental k-mer DBs first (k=31)
yak count -k31 -b37 -t16 -o pat.yak paternal_R1.fq.gz paternal_R2.fq.gz
yak count -k31 -b37 -t16 -o mat.yak maternal_R1.fq.gz maternal_R2.fq.gz
hifiasm -o prefix -t 32 -1 pat.yak -2 mat.yak reads.hifi.fq.gz

# Inbred / homozygous / mole: DISABLE purging or real duplications are deleted
hifiasm -o prefix -t 32 -l0 reads.hifi.fq.gz

# Unbalanced haplotype sizes: pin the homozygous-coverage peak read off the k-mer histogram
hifiasm -o prefix -t 32 --hom-cov 38 reads.hifi.fq.gz

# Ultralong ONT toward T2T (combine with Hi-C or trio for phasing)
hifiasm -o prefix -t 32 --ul ul_ont.fq.gz reads.hifi.fq.gz

Key flags: -l purge level (0 none, 1 light, 2/3 aggressive; default 3 in HiFi-only, 0 in trio); --h1/--h2 Hi-C R1/R2; -1/-2 paternal/maternal yak DBs; --ul ultralong ONT; --hom-cov INT force the homozygous-coverage peak; --n-hap INT ploidy (default 2); --primary emit primary + alternate (a_ctg) instead of dual hap1/hap2.

Output Filenames: the .bp. vs .dip. Convention (read the prefix, it encodes the mode)

The prefix is not cosmetic - it tells how the assembly was phased:

• .bp. ("balanced phasing") = HiFi-only OR Hi-C mode. prefix.bp.p_ctg.gfa (the mosaic primary), prefix.bp.hap1.p_ctg.gfa/prefix.bp.hap2.p_ctg.gfa (the two haplotypes - partially phased in HiFi-only, fully phased with Hi-C). Also prefix.bp.r_utg.gfa/p_utg.gfa (unitig graphs).
• .dip. ("diploid") = trio mode. prefix.dip.hap1.p_ctg.gfa (paternal/hap1), prefix.dip.hap2.p_ctg.gfa (maternal/hap2).
• --primary mode: prefix.p_ctg.gfa (primary) + prefix.a_ctg.gfa (alternate). The alternate is incomplete by construction (only heterozygous loci produce alt contigs).
• Reusable binaries prefix.ec.bin/ovlp.*.bin let a re-run with different -l/phasing skip error-correction.

People grep hap1.p_ctg and are confused when a trio run made dip.hap1.p_ctg and a Hi-C run made bp.hap1.p_ctg. Verify the actual emitted names against the installed version.

GFA Is Not FASTA (extract S lines before anything downstream)

hifiasm and verkko emit assembly graphs (GFA), and downstream tools want FASTA. Contig sequences live in GFA S (segment) lines:

gfatools gfa2fa prefix.bp.hap1.p_ctg.gfa > hap1.fa          # preferred
awk '/^S/{print ">"$2"\n"$3}' prefix.bp.hap1.p_ctg.gfa > hap1.fa   # dependency-free fallback

The graph also carries bubbles/alternate paths the FASTA throws away - keep the GFA. Verkko works internally in homopolymer-compressed coordinates (its .gfa is HPC); its final assembly.fasta is in normal space.

verkko (T2T, when the goal genuinely needs it)

T2T is a project, not a flag. Reach for verkko only when T2T completeness is the actual goal AND the data exist (deep HiFi PLUS good ultralong ONT N50 PLUS trio or Hi-C). It is slower, heavier (hundreds of CPU-hours, high RAM, Snakemake-orchestrated), and more fragile than hifiasm. CHM13 - the first T2T human - was a hydatidiform mole precisely because a mole is effectively homozygous, decoupling repeat-resolution from phasing (Nurk 2022 Science 376:44).

# Trio-phased T2T: build parental + CHILD meryl DBs (k=30), derive hap-mers, then run.
# --hap-kmers needs HAP-MER DBs (haplotype-specific k-mers), not raw parental count DBs.
meryl count k=30 paternal.fq.gz output pat.meryl
meryl count k=30 maternal.fq.gz output mat.meryl
meryl count k=30 child.fq.gz    output child.meryl
$MERQURY/trio/hapmers.sh mat.meryl pat.meryl child.meryl   # hapmers.sh takes maternal first; emits mat/pat.hapmer.meryl
verkko -d asm_out --hifi hifi.fq.gz --nano ul_ont.fq.gz \
  --hap-kmers pat.hapmer.meryl mat.hapmer.meryl trio       # paternal first -> haplotype1=paternal (matches hifiasm -1 pat)

# Hi-C-phased (no parents)
verkko -d asm_out --hifi hifi.fq.gz --nano ul_ont.fq.gz --hic1 hic_R1.fq.gz --hic2 hic_R2.fq.gz

--hifi accurate reads; --nano ultralong ONT (the spanning data, strongly recommended); --hap-kmers <hap1> <hap2> trio (argument ORDER sets which becomes haplotype1/haplotype2 - it is not inferred from biology) or --hic1/--hic2 for phasing. Outputs assembly.fasta, assembly.haplotype1.fasta, assembly.haplotype2.fasta, assembly.homopolymer-compressed.gfa. verkko-vs-hifiasm(--ul): verkko closes more chromosomes automatically at higher cost; hifiasm --ul brings much of the spanning benefit far cheaper but needs more manual finishing. Neither reaches T2T without ultralong reads.

Per-Method Failure Modes

Primary assembly used as if it were a haplotype

Trigger: running the primary .bp.p_ctg/p_ctg into an allele-aware analysis. Mechanism: the primary is a per-block haplotype mosaic. Symptom: garbage phased variant / ASE / HLA results; "best" metrics. Fix: use trio/Hi-C-phased hap1/hap2; never the primary for allele-aware work.

Partial phasing mistaken for full phasing

Trigger: treating HiFi-only .bp.hap1/hap2 as trio-grade haplotypes. Mechanism: no linkage data -> locally phased with inter-block switch errors. Symptom: phasing breaks at block boundaries; high switch rate vs trio. Fix: add Hi-C (--h1/--h2) or trio (-1/-2) for global phasing; validate with hap-mers.

Default purging on an inbred/homozygous sample

Trigger: running hifiasm at default -l3 on an inbred line, doubled-haploid, or mole. Mechanism: real segmental duplications/paralogs look like duplicate haplotigs in a homozygous genome and get purged. Symptom: assembly shrinks below true genome size; real duplications collapsed. Fix: -l0 (purging off) for low-heterozygosity samples.

Unbalanced hap1/hap2 size read as biology

Trigger: one haplotype much larger than the other. Mechanism: hifiasm mis-estimated the homozygous-coverage peak (bimodal coverage, odd ploidy, contamination), so it over-/under-purged. Symptom: size ratio far from 1. Fix: read the peak off the k-mer/coverage histogram (GenomeScope) and set --hom-cov.

Phasing reported without phasing QC

Trigger: N50 + BUSCO + QV reported, no switch/hamming or hap-mer blob plot. Mechanism: those metrics are structurally blind to mosaicism. Symptom: a "great" assembly that is a phasing disaster. Fix: Merqury hap-mer blob plot + switch/hamming vs trio/Hi-C hap-mers (-> assembly-qc). No hap-mers (no trio/Hi-C) = phasing unvalidated by construction.

verkko recommended reflexively / without ultralong reads

Trigger: reaching for verkko (or --ul) without ultralong ONT or without a T2T goal. Mechanism: UL N50 - not HiFi coverage - is the binding T2T constraint; 60x HiFi cannot span a 3 Mb satellite array. Symptom: months of compute, no T2T gain. Fix: confirm UL N50 (a meaningful fraction >100 kb) and a real T2T need first; else hifiasm-HiFi-only.

Quantitative Thresholds

| Threshold | Source | Rationale | |-----------|--------|-----------| | HiFi coverage >=13x per haplotype; ~30-40x diploid sweet spot | hifiasm FAQ + community | below ~13x/hap het bubbles fragment; past ~40x diminishing returns and worse false-dup | | HiFi >40x is not free quality | community convention | extra coverage can confuse the --hom-cov estimate and inflate false duplications | | Ultralong ONT N50: the longer the better, meaningful fraction >100 kb (T2T targets >100 kb-1 Mb) | T2T practice | UL value is entirely about spanning centromeres/satellites; short "long" reads add little | | -l purge: default 3 (HiFi-only), 0 (trio), set 0 for inbred | hifiasm man page | aggressive purging cleans outbred het-dups but deletes real sequence in homozygous samples | | k=31 (yak/hifiasm trio) vs k=30 (verkko/Merqury hap-mers) | tool conventions | mixing k silently corrupts hap-mer matching - match each tool | | Assembly QV: HiFi diploid routinely Q40-Q50; T2T reference ~Q60-Q70+ | Merqury practice | base QV is a SEPARATE axis from phasing accuracy (switch/hamming) | | Do NOT reflexively polish HiFi | over-polishing risk | HiFi is already ~Q30+ at read level; polishing an accurate assembly often lowers QV (-> assembly-polishing) |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | Downstream tool rejects hifiasm output | GFA, not FASTA | extract S lines (gfatools gfa2fa or awk) | | Assembly below expected genome size, real duplications gone | default purging on an inbred/mole sample | re-run with -l0 | | hap1 and hap2 wildly different sizes | mis-estimated homozygous-coverage peak | set --hom-cov from the k-mer histogram | | "hap1" file but phasing breaks across blocks | HiFi-only partial phasing | add Hi-C/trio for global phasing | | Allele-aware analysis (ASE/phased variants) looks wrong | used the primary mosaic | use phased hap1/hap2, not primary | | QV drops after polishing the HiFi assembly | over-polishing an already-accurate assembly | stop; HiFi rarely needs short-read polish | | verkko run never reaches T2T | no/short ultralong ONT | UL N50 is the binding constraint; add long UL or stop at hifiasm | | Trio run produced dip.*, scripts expected bp.* | mode encodes the prefix | match filename prefix to phasing mode |

References

• Cheng H, Concepcion GT, Feng X, Zhang H, Li H. 2021. Haplotype-resolved de novo assembly using phased assembly graphs with hifiasm. Nat Methods 18:170-175.
• Cheng H, Jarvis ED, Fedrigo O, et al. 2022. Haplotype-resolved assembly of diploid genomes without parental data. Nat Biotechnol 40:1332-1335.
• Nurk S, Walenz BP, Rhie A, et al. 2020. HiCanu: accurate assembly of segmental duplications, satellites, and allelic variants from high-fidelity long reads. Genome Res 30:1291-1305.
• Koren S, Rhie A, Walenz BP, et al. 2018. De novo assembly of haplotype-resolved genomes with trio binning. Nat Biotechnol 36:1174-1182.
• Rautiainen M, Nurk S, Walenz BP, et al. 2023. Telomere-to-telomere assembly of diploid chromosomes with Verkko. Nat Biotechnol 41:1474-1482.
• Antipov D, Rautiainen M, Nurk S, et al. 2025. Verkko2 integrates proximity-ligation data with long-read de Bruijn graphs. Genome Res 35:1583-1594.
• Nurk S, Koren S, Rhie A, et al. 2022. The complete sequence of a human genome (T2T-CHM13). Science 376:44-53.
• Liao WW, Asri M, Ebler J, et al. 2023. A draft human pangenome reference. Nature 617:312-324.
• Rhie A, Walenz BP, Koren S, Phillippy AM. 2020. Merqury: reference-free quality, completeness, and phasing assessment for genome assemblies. Genome Biol 21:245.

Related Skills

• genome-profiling - Decide outbred/inbred/ploidy and the homozygous-coverage peak before setting purge level and --hom-cov
• assembly-qc - Merqury hap-mer blob plots and switch/hamming are the only QC that sees phasing
• assembly-polishing - HiFi is already accurate; deciding whether to polish at all
• scaffolding - Hi-C used here for phasing; chromosome-scale scaffolding (YaHS/SALSA2) is the adjacent step
• contamination-detection - Screen contigs for foreign sequence after assembly
• long-read-sequencing/long-read-qc - HiFi read length/QV/contamination before assembly
• workflows/genome-assembly-pipeline - End-to-end profile -> assemble -> phase -> QC -> scaffold