Imported from bg-szy/TOP-SKILLS (skills/awesome-skills/bio-comparative-genomics-pangenome-analysis/SKILL.md). Install upstream with npx skills add bg-szy/TOP-SKILLS --skill bio-comparative-genomics-pangenome-analysis. Copyright stays with the author.

Version Compatibility

Reference examples tested with: Panaroo 1.5.1+ (Tonkin-Hill 2020 Genome Biol 21:180), PPanGGOLiN 2.2.0+ (Gautreau 2020 PLoS Comp Biol 16:e1007732), PEPPAN 1.0.5+ (Zhou 2020 GR 30:1667), GET_HOMOLOGUES 25102023+, anvi'o 8.0+ (Eren 2021 Nat Microbiol 6:3), Minigraph-Cactus (Hickey 2024 Nat Biotech 42:663; bundled with Cactus 2.5+), PGGB 0.7.5+ (Garrison 2024 Nat Methods 21:2008), vg 1.59.0+ (Sirén J et al 2024 Nat Methods 21:2017), PanGenie 3.1.0+ (Ebler 2022 Nat Genet 54:518), PGR-TK 0.3.6+ (Chin 2023 Nat Methods 20:1213; cschin/pgr-tk; repo archived April 2026 transitioning to PANGEA), PANGEA (in development by DGI / Diploid Genomics as PGR-TK's successor for pangenome graph exploration + analysis -- check https://github.com/cschin/pgr-tk for current repository pointer), Bakta 1.10.4+ (annotation for input), Roary 3.13.0+ (DEPRECATED; use Panaroo), Scoary 1.6.16+, pyseer 1.3.11+, BUSCO 5.7+, FastTree 2.1.11+, RAxML-NG 1.2+. Python 3.10+ required for Panaroo / PPanGGOLiN.

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

• CLI: panaroo --version; ppanggolin --version; peppan --help; cactus-pangenome --help; pggb --version; vg version
• Python: pip show panaroo ppanggolin

If code throws Bakta annotation incompatible, GFA file inconsistent, vg index version mismatch, the bacterial pangenome ecosystem expects consistent annotation; re-annotate all input genomes with the same tool and version before pangenome analysis.

Pangenome Analysis

"What genes are universal vs accessory across this set of genomes?" -> The pangenome is the union of all genes across a sampled group; the Tettelin partition (Tettelin 2005 PNAS 102:13950) splits it into core (universal), shell (in many but not all), cloud (rare), and species-specific (private) genes. The fundamental dichotomy is bacterial pangenome (clusters genes into orthogroups; Panaroo / PPanGGOLiN / PEPPAN for compact genomes) vs eukaryotic pangenome (graph-based; Minigraph-Cactus / PGGB / vg for haplotype-resolved sequences). The choice depends on what's being represented: bacterial pangenome captures gene-content variation in a species/genus; eukaryotic pangenome graph captures haplotype-level structural and sequence variation. Roary (Page 2015) is now deprecated in favor of Panaroo, which handles annotation-error noise that previously inflated bacterial pangenomes substantially.

• CLI: panaroo -i annotated_gffs/ -o panaroo_out --clean-mode strict --remove-invalid-genes -- bacterial pangenome
• CLI: ppanggolin workflow --fasta fasta_list.tsv --output ppanggolin_out -- partitioned bacterial pangenome
• CLI: peppan -i genomes.gff -o peppan_out -- genus-scale bacterial pangenome
• CLI: cactus-pangenome jobStore seqFile.txt --reference name --vcf --gfa -- Minigraph-Cactus pangenome
• CLI: pggb -i genomes.fa.gz -n 90 -t 32 -o pggb_out -- PGGB pangenome graph
• CLI: vg autoindex --workflow giraffe -r ref.fa -v variants.vcf.gz -- vg pangenome indexing

Algorithmic Taxonomy

Methodology evolves; verify the current Panaroo and PPanGGOLiN manuals + the 2024-2025 microbial pangenome reviews. The HPRC draft pangenome (Liao 2023 Nature 617:312) sets the modern eukaryotic pangenome standard; for bacterial work, Panaroo + PPanGGOLiN is the standard combination.

Decision Tree by Experimental Scenario

Per-Tool Failure Modes

Annotation heterogeneity inflating accessory genome

Trigger: Running Panaroo on Prokka- vs Bakta- vs RefSeq- annotated genomes mixed.

Mechanism: Different annotation tools predict different gene boundaries; the same gene is annotated slightly differently across tools, appearing as separate orthogroups. Roary's tolerance of these differences inflated the bacterial accessory genome substantially -- nearly an order of magnitude on the clonal M. tuberculosis benchmark (Tonkin-Hill 2020 GB 21:180). Panaroo's graph-based correction reduces this but cannot eliminate it.

Symptom: Per-strain "accessory" gene count is inflated relative to known biology; comparison to a single-pipeline reference reveals substantial spurious gene-content differences.

Fix: Re-annotate ALL genomes with one pipeline (currently Bakta 1.10.4+ for bacteria; Bakta is GenBank-compliant and faster than Prokka). Use panaroo --clean-mode strict --remove-invalid-genes to apply graph cleaning. Document annotation pipeline + version in methods.

Tettelin partition class boundary artifacts

Trigger: Reporting "core genome" vs "accessory" as percent-of-strains thresholds (e.g. 99% = core).

Mechanism: Tettelin 2005 used 100%-presence = core; pragmatic studies use 95%-99%. The class boundary is arbitrary; small variation in the threshold dramatically changes core/accessory ratio.

Symptom: Core genome size varies 10-30% depending on whether threshold is 95% or 99%.

Fix: Report core/shell/cloud at multiple thresholds; PPanGGOLiN's HMM partition is more principled but still has tunable parameters. Standard reporting: core = present in >=95% (or >=99%); shell = 15-95%; cloud = < 15%. Document threshold.

Roary's annotation-error noise (DEPRECATED tool)

Trigger: Using Roary in new analyses.

Mechanism: Roary tolerates annotation errors (low-identity matches; protein-vs-DNA matches), producing thousands of artifactual accessory genes. Panaroo's introduction (Tonkin-Hill 2020) demonstrated this by re-analyzing 413 Mtb genomes and finding Panaroo's accessory genome was nearly an order of magnitude smaller than the inflated tools' (Roary included).

Symptom: Roary output has many "lineage-specific genes" with poor evidence (single-strain hits, short proteins, no functional annotation).

Fix: Migrate to Panaroo. Panaroo can read Roary's input format; for legacy projects, re-run with Panaroo and compare. Panaroo's --clean-mode strict enforces strict graph-based quality control.

Pangenome graph reference bias (eukaryote)

Trigger: Using Minigraph-Cactus with a single reference; calling variants against the "reference" path.

Mechanism: Minigraph-Cactus is reference-anchored; the chosen reference appears throughout the graph as a privileged path. Variants are called relative to the reference path; non-reference haplotypes are under-represented in the variant calling.

Symptom: Variant call density on non-reference haplotypes is lower than on reference; allele frequencies skewed toward reference.

Fix: Use PGGB (reference-free) for less reference-biased analysis. Alternatively, treat the reference choice as a methodological parameter and document. For HPRC, multiple references can be used and results pooled.

Heaps law misapplied

Trigger: Concluding "open pangenome" from a single Heaps-law fit on insufficient data.

Mechanism: Heaps law parameter alpha distinguishes open (alpha < 1) from closed (alpha > 1) pangenome; estimation requires sampling many genomes. Few genomes give unstable estimates.

Symptom: Heaps-law alpha varies by > 0.2 across resampling; conclusions about pangenome openness flip.

Fix: Require >= 50 genomes (preferably > 100) for Heaps-law estimation; report 95% CI from resampling. Tettelin 2005 demonstrated open Streptococcus agalactiae; Vernikos 2015 reviews open vs closed across taxa.

PGGB memory exhaustion at many genomes

Trigger: Running PGGB with > 30 large eukaryotic genomes on a single node.

Mechanism: PGGB's wfmash all-vs-all step has O(N^2) memory pattern; large eukaryotic genomes (> 1 Gb) make memory prohibitive for > 30 input genomes.

Symptom: PGGB OOMs at the wfmash stage; cluster job killed by OOM-killer.

Fix: Use Minigraph-Cactus for > 30 genomes (it scales better); or split PGGB into chromosomes/regions. PGGB recommendation is <= 20 large genomes per run.

vg index version mismatch breaking giraffe

Trigger: Pre-built vg index used with a different vg version for read alignment.

Mechanism: vg index format evolved; pre-built indexes from one version may not be compatible with another.

Symptom: vg giraffe fails with "index version" error.

Fix: Rebuild vg index with current version; or pin vg version for an analysis. Future vg releases promise backward compatibility but verify.

PanGenie genotype false positives in repetitive regions

Trigger: PanGenie on highly repetitive regions (centromeres, segmental duplications).

Mechanism: Pangenome-graph-based genotyping requires unique paths in the graph; highly repetitive regions create graph-spaghetti paths that are difficult to genotype reliably.

Symptom: PanGenie calls many heterozygous SVs in known-repetitive regions; quality scores low.

Fix: Restrict PanGenie to non-repetitive regions; combine with traditional read-based SV callers (DELLY, Manta) for repeat regions. The HPRC paper documents this limitation (Liao 2023).

Bacterial pangenome with frequent gene-content recombination

Trigger: Building a pangenome of a species with extensive recombination (e.g. Neisseria, Streptococcus pneumoniae).

Mechanism: Frequent recombination breaks the "vertical inheritance" assumption underlying ortholog clustering; the same gene appears in many phylogenetic positions across strains, complicating orthology and inflating accessory genome.

Symptom: Phylogenetic trees from core genome are unstable; per-gene trees show extensive incongruence; pangenome accessory genome appears artificially large.

Fix: Use ClonalFrameML (Didelot 2015 PLoS Comp Biol 11:e1004041) to identify recombinant regions; mask them before pangenome analysis. Restrict core genome analysis to non-recombinant regions.

Annotation density variation across genomes

Trigger: Mixing well-annotated reference genomes with newly assembled, draft-annotation genomes.

Mechanism: Draft annotations miss small genes, pseudogenes, and lineage-specific genes; well-annotated genomes have these. Comparing them inflates "accessory" in draft genomes.

Symptom: Draft genomes have 200-500 fewer accessory genes than expected; per-genome BUSCO completeness > annotation completeness.

Fix: Re-annotate all genomes consistently with Bakta + Prodigal; document BUSCO completeness for each. Exclude genomes with > 5% lower BUSCO than median.

Quantitative Thresholds

Panaroo Bacterial Pangenome Workflow

Goal: Construct a high-quality bacterial pangenome with annotation-error correction.

Approach: Annotate genomes consistently with Bakta -> run Panaroo strict mode -> partition with PPanGGOLiN.

# 1. Annotate all genomes with Bakta (consistent annotation)
mkdir -p annotated
for fa in genomes/*.fa; do
    name=$(basename $fa .fa)
    bakta --db /path/to/bakta-db --threads 16 \
        --output annotated/${name} --prefix $name \
        --genus Escherichia --species coli \
        $fa
done

# 2. Run Panaroo
panaroo -i annotated/*.gff -o panaroo_out -t 16 \
    --clean-mode strict --remove-invalid-genes

# 3. Extract pangenome matrix
# panaroo_out/gene_presence_absence.csv      strains x genes matrix
# panaroo_out/core_gene_alignment.aln        core gene MSA for phylogeny
# panaroo_out/pan_genome_reference.fa        consensus pangenome sequence

# 4. Tettelin partition (custom)
python tettelin_partition.py \
    --presence panaroo_out/gene_presence_absence.csv \
    --core-threshold 0.99 --shell-threshold 0.15 \
    --output panaroo_out/tettelin_classification.tsv

# 5. PPanGGOLiN HMM partition (alternative)
# PPanGGOLiN expects a TSV index: `genome_name<TAB>path/to.gff3` per row
for f in annotated/*.gff3; do
    printf "%s\t%s\n" "$(basename "$f" .gff3)" "$(realpath "$f")"
done > gff_list.tsv
ppanggolin all --anno gff_list.tsv -o ppanggolin_out --threads 16
# Output: ppanggolin_out/pangenome.h5 (HDF5 with HMM-partitioned genes)

'''Tettelin core/shell/cloud partition from Panaroo gene presence/absence matrix.'''
import pandas as pd


def tettelin_partition(presence_matrix, core_threshold=0.99,
                       shell_threshold=0.15):
    '''Returns DataFrame[gene_name] -> Tettelin class.'''
    # presence_matrix: rows = genes, cols = strains, 0/1 entries
    n_strains = presence_matrix.shape[1]
    fraction = presence_matrix.sum(axis=1) / n_strains
    classes = pd.cut(fraction,
                     bins=[-0.01, shell_threshold, core_threshold, 1.01],
                     labels=['cloud', 'shell', 'core'])
    return pd.DataFrame({'fraction': fraction, 'class': classes})


def heaps_law(presence_matrix, n_iters=100):
    '''Estimate Heaps law alpha from genome sampling order.'''
    import numpy as np
    n_strains = presence_matrix.shape[1]
    pan_sizes = []
    for _ in range(n_iters):
        order = np.random.permutation(n_strains)
        pan = set()
        sizes = []
        for i in order:
            genes_in_i = presence_matrix.iloc[:, i] == 1
            pan.update(genes_in_i.index[genes_in_i].tolist())
            sizes.append(len(pan))
        pan_sizes.append(sizes)
    pan_array = np.array(pan_sizes)  # n_iters x n_strains
    n_sampled = np.arange(1, n_strains + 1)
    # Fit log-log
    mean_pan = pan_array.mean(axis=0)
    log_n = np.log(n_sampled)
    log_pan = np.log(mean_pan)
    alpha = np.polyfit(log_n, log_pan, 1)[0]
    return alpha

Minigraph-Cactus for Eukaryotic Pangenome

Goal: Build pangenome graph from haplotype-resolved assemblies.

Approach: Provide reference + haplotypes -> Minigraph-Cactus produces GFA + VCF + GBZ for downstream genotyping.

# Prepare seqFile (Cactus convention)
cat > pangenome_seqs.txt << 'EOF'
GRCh38      GRCh38.fa
HG002.hap1  HG002.hap1.fa
HG002.hap2  HG002.hap2.fa
HG003.hap1  HG003.hap1.fa
HG003.hap2  HG003.hap2.fa
EOF

cactus-pangenome jobStore_path pangenome_seqs.txt \
    --outDir hprc_pangenome \
    --outName hprc_pangenome \
    --reference GRCh38 \
    --vcf \
    --gfa \
    --gbz \
    --indexCores 32 \
    --mapCores 32

# Outputs:
#   hprc_pangenome/hprc_pangenome.full.hal      Full Cactus HAL
#   hprc_pangenome/hprc_pangenome.gfa.gz        Graph Fragment Assembly format
#   hprc_pangenome/hprc_pangenome.vcf.gz        Short variants relative to GRCh38
#   hprc_pangenome/hprc_pangenome.gbz           GBZ compressed graph
#   hprc_pangenome/hprc_pangenome.giraffe.gbz   Giraffe-indexed graph for mapping

vg Pangenome Genotyping

Goal: Genotype short reads against a pre-built pangenome graph.

Approach: Pre-built pangenome -> vg autoindex -> vg giraffe (fast mapping) -> vg call variants.

# Pre-build index
vg autoindex --workflow giraffe --threads 16 \
    --ref-graph hprc_pangenome.gfa.gz \
    --output hprc_index

# Map reads
vg giraffe \
    --gbz-name hprc_index.giraffe.gbz \
    --dist-name hprc_index.dist \
    --minimizer-name hprc_index.min \
    --fastq-in sample.R1.fq.gz \
    --fastq-in sample.R2.fq.gz \
    --output-format GAM \
    --threads 16 \
    > sample.gam

# Pack alignment information
vg pack -x hprc_index.giraffe.gbz -g sample.gam -o sample.pack

# Call variants
vg call hprc_index.giraffe.gbz -k sample.pack -a > sample.vcf

Pan-GWAS with Scoary

Goal: Identify gene presence/absence associated with a phenotype.

Approach: Panaroo presence/absence matrix + phenotype file -> Scoary -> phenotype-gene associations.

# Run Scoary
scoary \
    --gene-presence-absence panaroo_out/gene_presence_absence.csv \
    --traits phenotypes.tsv \
    --output scoary_out \
    --threads 16 \
    --upgma-tree

# Output:
#   scoary_out/*_results.csv   per-trait gene associations

Reconciliation: When Methods Disagree

Operational rule for publication: Bacterial pangenome uses Panaroo + PPanGGOLiN cross-validation; eukaryotic pangenome uses Minigraph-Cactus (HPRC scale) or PGGB (reference-free). Document annotation pipeline + version; report core/shell/cloud at multiple thresholds; verify Heaps-law on > 100 genomes for openness claims.

Cohort Gotchas

• Endosymbiont genomes (Buchnera, Wolbachia, mitochondria): core genome is dominant; accessory is minimal; pangenome analysis less informative
• Hypothetical proteins: unknown function genes dominate "accessory" in non-model species; functional analysis needed
• Phage genes: prophages contribute to accessory; mask prophages for "core species genes"
• Mobile genetic elements: plasmids/IS elements appear in accessory; tag with mobileOG-db
• Highly recombinogenic species (Neisseria, S. pneumoniae): core genome unstable; restrict to non-recombinant regions
• Plasmids: chromosome vs plasmid distinction matters for pangenome; classify with PlasmidFinder
• Polyploid eukaryotic pangenome: subgenomes must be assigned first (see [[whole-genome-duplication]])
• Pangenome graph for SV detection: PanGenie + graph aligners; reference-anchored variant callers miss SVs

Anticipated Reviewer Pushback

Common Errors

Tool Installation Notes

# Bacterial pangenome
conda install -c bioconda panaroo ppanggolin peppan get_homologues anvio

# Eukaryotic pangenome
conda install -c bioconda cactus pggb vg pangenie

# PGR-TK (repeat-rich / clinical gene focus)
conda install -c bioconda pgr-tk
# Or: cargo install pgrtk; or Docker quay.io/cschin/pgr-tk
# Note: cschin/pgr-tk archived April 2026; transitioning to PANGEA (developed by DGI / Diploid Genomics; check upstream repo for pointer)

# Annotation
conda install -c bioconda bakta prokka

# Pan-GWAS
conda install -c bioconda scoary pyseer

# Recombination
conda install -c bioconda clonalframeml

# Mobile elements
git clone https://github.com/clb21565/mobileOG-db

# QC
conda install -c bioconda busco compleasm

For HPRC-scale eukaryotic pangenome, use cluster with >= 500 GB RAM and HPC scheduler integration via Toil (see [[whole-genome-alignment]]).

References

• Tettelin H et al 2005 PNAS 102:13950 (core/accessory pangenome framework)
• Tonkin-Hill G et al 2020 Genome Biol 21:180 (Panaroo)
• Gautreau G et al 2020 PLoS Comp Biol 16:e1007732 (PPanGGOLiN)
• Zhou Z et al 2020 Genome Res 30:1667 (PEPPAN)
• Page AJ et al 2015 Bioinformatics 31:3691 (Roary; DEPRECATED)
• Contreras-Moreira B & Vinuesa P 2013 Appl Environ Microbiol 79:7696 (GET_HOMOLOGUES)
• Eren AM et al 2021 Nat Microbiol 6:3 (anvi'o pangenomics)
• Hickey G et al 2024 Nat Biotech 42:663 (Minigraph-Cactus)
• Garrison E et al 2024 Nat Methods 21:2008 (PGGB)
• Sirén J et al 2024 Nat Methods 21:2017 (vg pangenome update)
• Ebler J et al 2022 Nat Genet 54:518 (PanGenie)
• Liao W-W et al 2023 Nature 617:312 (HPRC draft pangenome)
• Brynildsrud O et al 2016 Genome Biol 17:238 (Scoary)
• Lees JA et al 2018 Bioinformatics 34:4310 (pyseer)
• Didelot X & Wilson DJ 2015 PLoS Comp Biol 11:e1004041 (ClonalFrameML)
• Schwengers O et al 2021 Microb Genom 7 (Bakta)
• Vernikos GS et al 2015 Curr Opin Microbiol 23:148 (pangenome openness review)
• Brown CL et al 2022 Appl Environ Microbiol 88:e0099122 (mobileOG-db)
• Mikheenko A et al 2018 Bioinformatics 34:i142 (QUAST-LG; long-read assembly evaluation -- earlier "pangenome review" attribution was incorrect; QUAST-LG benchmarks large-scale assembly QC).
• Chin C-S et al 2023 Nat Methods 20:1213 (PGR-TK; MAP graphs + principal bundle decomposition for repeat-rich / clinical genes; MHC, DAZ1-4, OPN1LW/OPN1MW examples)
• cschin/pgr-tk GitHub (repo; archived April 2026, transitioning to PANGEA developed by DGI / Diploid Genomics; consult upstream README for PANGEA pointer)

Related Skills

• comparative-genomics/whole-genome-alignment - Minigraph-Cactus builds on Cactus; PGGB underlies eukaryotic pangenome
• comparative-genomics/ortholog-inference - Pangenome clusters are bacterial orthologs at species/genus level
• comparative-genomics/hgt-detection - Accessory genes often HGT-derived; mobile-element annotation cross-references
• comparative-genomics/gene-family-evolution - CAFE5 modeling on Panaroo presence/absence matrix
• genome-annotation/prokaryotic-annotation - Bakta annotation is pangenome input
• genome-annotation/repeat-annotation - Repeat masking before eukaryotic pangenome
• variant-calling/structural-variant-calling - Pangenome graph SV calling complements read-based callers
• variant-calling/joint-calling - vg + graph variant calling integrates with traditional VCF
• metagenomics/amr-detection - AMR genes often in bacterial accessory
• metagenomics/strain-tracking - Strain-specific accessory genes for tracking
• population-genetics/association-testing - Pan-GWAS for phenotype-gene-content association