Roary: rapid large-scale prokaryote pan genome analysis (Page et al. 2015)

1. Citation

Page AJ, Cummins CA, Hunt M, Wong VK, Reuter S, Holden MTG, Fookes M, Falush D, Keane JA, Parkhill J. 2015. “Roary: rapid large-scale prokaryote pan genome analysis.” Bioinformatics 31(22):3691–3693. DOI 10.1093/bioinformatics/btv421. Open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC4817141/ (also https://academic.oup.com/bioinformatics/article/31/22/3691/240757).

2. Access note

Read via PMC full text (PMC4817141). This is a short “Applications Note” (~3 pages). Full body text, abstract, and Tables 1–2 read. Supplementary material NOT accessed — some implementation details (exact CD-HIT iterative step values, full command-line flag list) are not in the main text and may live only in the supplement or the software docs; flagged below where relevant.

3. Thesis (1 sentence)

Roary rapidly builds large-scale prokaryote pan genomes — identifying core and accessory genes across thousands of same-species samples — on a standard desktop without compromising accuracy.

4. Problem & context

Prokaryote population sequencing studies now routinely contain hundreds to thousands of isolates. Existing pan-genome tools do not scale to such datasets (paper shows several competitors failing to complete within 5 days on 1000 samples; see §6). Roary targets that computational-infeasibility gap while preserving accuracy.

5. Method / approach

Pipeline (as described in the main text):

Coding regions extracted from input and converted to protein sequences; filtered to remove partial sequences.
Protein sequences iteratively pre-clustered with CD-HIT (the specific iterative percent-identity step values are NOT given in the main text — likely supplement/software; [summarizer-inferred] gap).
All-against-all comparison with BLASTP on the reduced sequences at a user-defined percentage sequence identity, default 95%.
Sequences clustered with MCL.
CD-HIT pre-clustering results merged with MCL results.
Using conserved gene neighborhood information, homologous groups containing paralogs are split into groups of true orthologs.
A graph is constructed of cluster relationships based on order of occurrence in the input.
Isolates are clustered based on gene presence in the accessory genome.

Input: one annotated assembly per sample in GFF3 format, such as produced by Prokka, where all samples are from the same species.

Default parameter of record: BLASTP percentage sequence identity default = 95% (user-configurable). NOTE: the main-text passage read does not print the command-line flag name (guidance expected -i); flag not stated in the article body — do not import from software docs (see §12 Q1).

Output/tooling: “A suite of command line tools is provided to interrogate the dataset providing union, intersection and complement.” (Specific output filenames not enumerated in main text.)

6. Key claims / findings

Scaling: “Using a single CPU Roary can produce a pan genome consisting of 1000 isolates in 4.5 hours using 13 GB of RAM, with further speedups possible using multiple processors.” (Abstract)
Runtime vs competitors (real S. typhi data, Table 2):
- 8 samples: Roary 44 s; PanOCT 1,457 s; LS-BSR 2,585 s; PGAP 41,397 s.
- 1000 samples: Roary 15,465 s (≈4.5 h); PGAP and PanOCT failed to complete within 5 days.
Accuracy on simulated 12-genome data (Table 1): Roary 994 core / 1017 total, 0 incorrect splits, 0 incorrect merges; PGAP 991/1012, 0 splits, 4 merges; PanOCT 993/1015, 1 split, 1 merge; LS-BSR 974/994, 0 splits, 23 merges.
Core definition: “Core is defined as a gene being in at least 99% of samples.” (This is the ONLY presence threshold/band the paper states — see §9 and §12 Q2.)

7. Load-bearing statements (VERBATIM — license permits; CC BY, see §License)

“An all-against-all comparison is performed with BLASTP on the reduced sequences with a user defined percentage sequence identity (default 95%).” — Methods.
“The input to Roary is one annotated assembly per sample in GFF3 format … such as that produced by Prokka … where all samples are from the same species.” — Methods.
“Core is defined as a gene being in at least 99% of samples.” — Methods/results.
“Using a single CPU Roary can produce a pan genome consisting of 1000 isolates in 4.5 hours using 13 GB of RAM, with further speedups possible using multiple processors.” — Abstract.
“Roary is implemented in Perl and is freely available under an open source GPLv3 license.” — availability.

8. Stated scope, assumptions, limitations

Same-species assumption: input samples are expected to be “from the same species.”
Input must be pre-annotated (GFF3, e.g. from Prokka) — Roary consumes annotations rather than producing them.
Accuracy claims benchmarked on 12 simulated genomes (Table 1) and S. typhi real data (Table 2); no broader taxonomic generality claimed in-text.

9. Failure modes / invalidity patterns

The paper does NOT state any explicit handling of annotation errors or annotation inconsistency; it assumes annotated GFF3 input of the same species (see §12 Q3). It makes no in-text claim to detect, correct, or be robust to mis-annotation — a confident silence, relevant to downstream inflation critiques but not addressed by this source itself.
Paralog handling is the named accuracy mechanism: “homologous groups containing paralogs are split into groups of true orthologs” via conserved gene neighborhood. Table 1 frames accuracy as incorrect splits / incorrect merges — those two counts are the paper’s own accuracy diagnostics.

10. What the source does NOT address

No soft-core / shell / cloud percent bands. The terms “soft core,” “shell,” “cloud” do NOT appear; no bands such as 95–99%, 15–95%, or <15% are stated. Only “core = ≥99% of samples” is given.
No explicit command-line flag names in the main text (no -i/-cd printed in body read).
Exact CD-HIT iterative identity step values not given in main text.
No annotation-error handling / robustness claims.
Output filenames not enumerated.

11. Open questions / ambiguities

Whether the 95% default is applied only at the BLASTP stage or also gates CD-HIT pre-clustering (text ties the 95% default specifically to BLASTP).
The provenance of the widely-cited soft-core/shell/cloud bands: NOT in this paper; likely software docs or later convention (not resolved here — do not attribute to this source).

12. Guidance answers

Q1 — default BLASTP(P) identity threshold + flag? Default = 95%, user-defined, applied at the BLASTP all-against-all step (verbatim §7 #1). The flag name is NOT printed in the article body (guidance anticipated -i); not stated here — do not import from software docs.

Q2 — core/soft-core/shell/cloud presence bands? Only “Core = gene in ≥99% of samples” is in the paper. Soft-core, shell, and cloud bands are NOT stated — the terms and their percent bands (e.g. soft-core 95–99%, shell 15–95%, cloud <15%) do not appear in this paper. This paper is NOT a citable origin for the full four-band convention; only the ≥99% core cutoff is sourced here.

Q3 — annotation-error handling vs clean-annotation assumption? The paper states no error handling; it assumes annotated GFF3 input “from the same species” (verbatim §7 #2). No claim of robustness to mis-annotation. (Faithful capture of what THIS paper says about its own assumptions.)

Q4 — expected input? “one annotated assembly per sample in GFF3 format … such as that produced by Prokka … where all samples are from the same species” (verbatim §7 #2).

Q5 — scaling numbers? 1000 isolates in 4.5 h using 13 GB RAM on a single CPU (verbatim §7 #4). Table 2: 8 samples in 44 s; 1000 samples in 15,465 s; competitors PGAP/PanOCT failed within 5 days.

License

CC BY (Creative Commons Attribution License). Article page states: ”© The Author 2015. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution License.” Software itself is GPLv3. Mode used: verbatim quoting permitted — §7 contains short verbatim load-bearing quotes.