henhaobio
Sign in

Genome biologyGenomicsPeer-reviewed journal article

Producing Polished Prokaryotic Pangenomes with the Panaroo Pipeline

Gerry Q. Tonkin‐HilliD, Neil MacAlasdairiD, Christopher RuisiD, Aaron WeimanniD, Gal HoreshiD, John A. LeesiD, Rebecca A. GladstoneiD, Stephanie W. LoiD, Christopher A. BeaudoiniD, R. Andrés FlotoiD, Simon D. W. FrostiD, Jukka CoranderiD, Stephen D. BentleyiD, Julian ParkhilliD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Population-level comparisons of prokaryotic genomes must take into account the substantial differences in gene content, resulting from frequent horizontal gene transfer, gene duplication and gene loss. However, the automated annotation of prokaryotic genomes is imperfect, and errors due to fragmented assemblies, contamination, diverse gene families and mis-assemblies accumulate over the population, leading to profound consequences when analysing the set of all genes found in a species. Here we introduce Panaroo, a graph based pangenome clustering tool that is able to account for many of the sources of error introduced during the annotation of prokaryotic genome assemblies. We verified our approach through extensive simulations of de novo assemblies using the infinitely many genes model and by analysing a number of publicly available large bacterial genome datasets. Using a highly clonal Mycobacterium tuberculosis dataset as a negative control case, we show that failing to account for annotation errors can lead to pangenome estimates that are dominated by error. We additionally demonstrate the utility of the improved graphical output provided by Panaroo by performing a pan-genome wide association study in Neisseria gonorrhoeae and by analysing gene gain and loss rates across 51 of the major global pneumococcal sequence clusters. Panaroo is freely available under an open source MIT licence at https://github.com/gtonkinhill/panaroo.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (28 Jan 2020) · CC BY 4.0

Panaroo's graph-based pangenome clustering uses gene adjacency across genomes to correct annotation errors, shrinking spurious accessory genomes and enlarging core genomes in simulated and real bacteria.

As bacterial genome collections grow, errors from fragmented assemblies, contamination, mis-annotation and divergent gene families accumulate, inflating accessory and shrinking core genome estimates, yet many existing pangenome tools had not been rigorously checked on simulated data. Panaroo shares information across genomes in a population graph to correct these errors, and its graph output supports structural-variant association and gene gain/loss analyses.

Key results

  1. ShownOn a highly clonal M. tuberculosis outbreak used as a negative control, competing tools reported large accessory genomes driven mainly by annotation errors, whereas Panaroo returned the most core and fewest accessory genes.

    413 Prokka-annotated Mtb assemblies, maximum pairwise SNP distance 9; PanX, PIRATE, COGsoft and Roary reported accessory genomes of 2584 to 3670 genes, nearly ten-fold Panaroo's; 59% of the between-method difference traced to assembly-fragmented genes and 10% to genes called in only a minority of near-identical isolates. · Corrected analysis of a Mycobacterium…; Figure legends

  2. ShownIn simulated E. coli pangenomes built from realistic short-read assemblies, Panaroo kept error counts low under contamination and heavy fragmentation, where the other tools' errors rose sharply.

    Five simple and two complex scenarios, three replicates of 100 genomes each, simulated from ASM584v2; simple runs: PanX and Panaroo had fewest erroneous clusters; with S. epidermidis contamination all tools but Panaroo and COGsoft inflated the accessory genome; with fragmentation PanX, COGsoft, PIRATE and Roary showed nearly an order of magnitude more errors. · Superior Performance on Simulated Populations; Simulation and Comparison with Previous Methods; Figure legends

  3. ShownIn a diverse global K. pneumoniae collection, Panaroo estimated the largest core genome and produced the fewest clusters with conflicting gene annotations, consistent with less erroneous merging.

    328 genomes with nine QC outliers removed; core genes at the 99% threshold: Panaroo 3372 (default) and 3376 (sensitive mode), PIRATE 3318, Roary 1800; Panaroo had the fewest conflicting annotations within clusters, PanX second and COGsoft the most (hypothetical and fragment annotations excluded). · Greater Internal Consistency in a Diverse…

  4. SuggestedGene and gene-triplet structural-variant association tests on Panaroo output for N. gonorrhoeae recovered resistance candidates and split a common plasmid into variants, the tetM-bearing ones tracking tetracycline resistance.

    Euro-GASP, 1054 isolates from 20 countries, phenotypes for seven antibiotics, pyseer: 67 genes and 138 triplets significant (adjusted p ≤ 0.05); insertion of transmembrane gene group_1138 linked to ceftriaxone resistance; a plasmid present 430 times resolved into variants, one with four triplets negatively associated with tetracycline resistance, others carrying tetM (group_1999). · Pyseer Association Analysis with Panaroo Finds…; Figure legends

  5. SuggestedApplying the Finitely Many Genes model to Panaroo output across pneumococcal lineages linked larger effective pangenomes to higher recombination and serotype diversity, and lower gene gain rates to invasiveness.

    51 Global Pneumococcal Sequence Clusters with dated phylogenies; effective pangenome size vs recombination rate (r/m): Spearman 0.53, p<0.001; vs serotype count: Spearman 0.51, p=0.001; lower gene gain rate associated with significant invasive-disease odds ratio (p=0.04); gene loss rate same direction (p=0.08). · Improved Methods for Analysing Pangenome…; Figure legends

Methods, data and reproducibility

Panaroo clusters genes from Prokka GFF3 files with CD-HIT at 98% identity, links clusters adjacent on any contig into a graph, then cleans it: pruning low-support degree-one nodes (contig ends, contamination), merging mistranslated genes, collapsing neighbouring families at 70% identity, and refinding missing genes with edlib.

  • Highly clonal isoniazid-resistant M. tuberculosis outbreak assemblies from London, annotated with Prokka and used as a negative control · Mycobacterium tuberculosis · n = 413 genome assemblies
  • Simulated pangenomes from E. coli reference ASM584v2 (Infinitely Many Genes model), reads simulated with Mason or ART and assembled with SPAdes · Escherichia coli (simulated) · n = three replicates of 100 genomes per parameter set · NC000913 (ASM584v2)
  • Reference genome used as the source of 10kb contaminating segments (Poisson mean 1) added in simulations · Staphylococcus epidermidis · ASM764v1
  • Globally sourced K. pneumoniae genomes from human and animal hosts · Klebsiella pneumoniae · n = 328 genomes (nine removed as QC outliers)
  • Euro-GASP isolates from European countries collected September to November 2013, with antimicrobial susceptibility for seven antibiotics · Neisseria gonorrhoeae · n = 1054 isolates (subsets of 10, 100, 1000 for benchmarking)
  • Global Pneumococcal Sequencing project clusters with reliable dated phylogenies, plus recombination, invasiveness and serotype data from Gladstone et al. · pneumococcus · n = 51 GPSCs
  • Archived GFF3, FASTA and intermediate post-processing files for replication · 10.5281/zenodo.3599800
Methods & tools
Panaroo (v1.0.0; Python 3.6+), CD-HIT (v4.8.1), edlib (v1.3.4), Prokka (v1.13.3) with Prodigal, Mason (v2.0.9) and ART (v2.5.8) read simulators, SPAdes (v3.13.0), dendropy (v4.4.0), Kingman coalescent simulation, pyseer (v1.3.0), Roary (v1.007002) and PIRATE (v1.0), PanX (v1.5.1) and COGsoft (v201204), Finitely Many Genes and Infinitely Many Genes models, Mash / Mash screen, Spydrpick, Cytoscape
Software & environment
Python 3.6+; Panaroo v1.0.0 with CD-HIT v4.8.1 and edlib v1.3.4; Prokka v1.13.3, SPAdes v3.13.0, Mason v2.0.9, ART v2.5.8, dendropy v4.4.0, pyseer v1.3.0; MIT licence
Assay conditions & compute
Benchmarked on 10, 100 and 1000 N. gonorrhoeae isolates with 5 CPUs per tool: Panaroo, Roary and PIRATE used similar CPU time and memory; PanX and COGsoft used the most, and COGsoft did not finish the largest set within a week.
Reproducibility gaps
Original sequencing accessions for the M. tuberculosis, K. pneumoniae, Euro-GASP and GPS collections are not given in the text; per-tool runtime and memory values are shown only graphically.
Limitations (6)
  • The default contamination filter, which recursively prunes low-support nodes of degree one or less, can discard rare plasmids; a more sensitive setting is offered to keep such rare elements. · stated by the authors
  • Real-data comparisons lack ground truth; for K. pneumoniae, within-cluster annotation conflicts serve as a proxy for incorrect merging, and conflicts involving hypothetical genes or fragments are ignored. · stated by the authors
  • Gene-triplet structural variant calls can capture large rearrangements, but the authors note such associations are often harder to interpret and need manual inspection of the pangenome graph in Cytoscape. · stated by the authors
  • Benchmarks reached 1000 isolates; scaling to hundreds of thousands or millions of genomes is stated as future performance work. · stated by the authors
  • Results link the group_1138 insertion to ceftriaxone resistance, but the Figure 5c legend reports triplets around group_1138 as tetracycline-associated; these associations also lack experimental validation. · noted in this summary
  • Simulations derive all genomes from a single E. coli reference and give Prokka a BLAST database of the true simulated proteins, which may understate annotation difficulty; per-scenario error counts appear only in figures. · noted in this summary

Read this if you build bacterial pangenomes from draft assemblies and need to know how annotation errors inflate accessory genomes, or want graph-based structural-variant inputs for pan-GWAS with pyseer.

Citations
1,349
FWCI
29.10
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHorsfield '23Le '24Page '15Arnoux '26Dabbaghie '23Manzano-Mor '23
59% semantic matchBioinformatics
Samuel Horsfield, Gerry Q. Tonkin‐Hill, Nicholas J. Croucher+1 more·Genome Research·Sep 1, 2023· 34 citations