A second paper redo: *E. coli* and *Salmonella* Kentucky from chickens in Bangladesh
On 2026-11-14, the same kind of agent as in the first paper redo got another
genome paper, its raw reads and irit-mcp, with no shell, no web access and no other analysis
tools. This time there were two species and six isolates, and irit-mcp had the tools the first
redo asked for: QUAST-style assembly statistics (assembly_stats) and depth from read mapping
(coverm_genome). Everything ran on the project laptop.
Result:
- Matched: all six sequence types (every allele exact), genome size within 0.13 %, GC within 0.01 points, and every acquired resistance gene in Table 1 apart from one gene and some allele names that come from using a different tool.
- Not checkable: clonal complex, plasmid replicons and serovar; irit has no tool for them.
- Problems the agent found in the paper:
- Twelve blaTEM alleles listed for one isolate where the assembly has one truncated blaTEM.
- Read counts and coverage that don’t follow from the deposited reads.
- Two isolates reported resistant to ceftriaxone and ertapenem with no gene that explains it.
- irit problems: the agent reported twelve. Two were real irit bugs and are fixed; the rest were setup on our side, missing tools, or upstream.
The paper
Section titled “The paper”Amena B, Islam M, Nizami TA, Barua H, Biswas PK. Draft genome sequences of multidrug-resistant Escherichia coli and Salmonella enterica serovar Kentucky isolates from chicken samples in Bangladesh. Microbiology Resource Announcements 15(3):e01429-25, published 9 February 2026. doi:10.1128/mra.01429-25, CC BY 4.0. Reads: SRA SRR30733610–SRR30733615 (BioProject PRJNA1162444).
Why this paper:
- Different from the first redo: two species instead of one, and a lab setting close to Indonesia’s (poultry farms and live bird markets).
- Data: Illumina NextSeq 2000 reads, 1.6 GB gzipped for all six isolates.
- Methods: FastQC, fastp, Shovill (SPAdes), QUAST, Kraken2, ResFinder, PlasmidFinder and PubMLST, run on Galaxy. irit-mcp wraps some of these and has alternatives for others.
- Table 1 has concrete values per isolate to check: ST, clonal complex, genome size, GC, contigs, N50, reads, coverage, resistance genes and plasmid replicons.
| Agent | Claude Code, headless (claude -p), default model (Claude Opus 5.5), built-in tools limited to Read/Write/Edit/Glob/Grep. The only MCP server was irit-mcp. |
| irit | Built from main at 861d753 (after 0.4.0: assembly_stats, coverm_genome, kleborate) |
| Tools | fastp 1.4.0, Shovill 1.4.2 (SPAdes 3.15.5), mlst 2.35.0, AMRFinderPlus 4.2.7, CheckM2 1.1.0, CoverM 0.8.0, via Bioconda environments |
| Databases | dblock: amrfinderplus-db-4.2-2026-10-29.1, checkm2-db-v3, gtdbtk-r232-ani-screen, gtdbtk-r232-placement-lite, all reused from the shared store |
| Limits | irit-mcp and every tool it started ran in an 8 GB, no-swap cgroup. The machine has 12 cores and was shared with nanopore runs and builds. |
| Brief | The first redo’s brief, adapted: check every Table 1 value, report species and quality, and say what looks wrong in the data or the paper. |
The session took 106 minutes and 52 turns, and cost US$5.43 at API prices. Compute steps, from
the run record (started/ended of each call):
| Step | Time |
|---|---|
| leanqc, all 12 read files | 28 s |
| fastp, per isolate | 11–24 s |
| Shovill (SPAdes), per isolate | 7.7–20.8 min |
| mlst, per isolate | 7–9 s |
AMRFinderPlus --plus, per isolate |
23–46 s |
| CheckM2, all six | 10 min |
| CoverM genome, per isolate | 20–83 s |
Two setup mistakes on our side limited the run:
- sheetgen: its binary wasn’t copied into irit-mcp’s bin folder, so the agent wrote the sample sheet by hand.
- Species check: the brief said “no web access” and didn’t say irit’s own tools may read from
GTDB’s server. So the agent didn’t let
gtdbprep_anifetch the 23 reference sketches it needed (75 MB). The first redo hit the same wording; the brief now has to say it.
The species check was run afterwards, outside the agent (below). Kleborate wasn’t installed; these aren’t Klebsiella, so it wouldn’t have helped.
Results against Table 1
Section titled “Results against Table 1”The agent’s full table is in its report; this is the summary.
| Row | Verdict | Detail |
|---|---|---|
| ST | 6 of 6 match | E. coli ST1589, ST2705, ST1196 (Achtman scheme) and Salmonella ST198 ×3, every allele exact |
| Clonal complex | not checked | mlst reports no clonal complex |
| Genome size | 5 match, 1 close | Within 0.13 %, but only when counting contigs of 200 bp or more (below) |
| GC | 6 of 6 match | Within 0.01 points |
| Contigs | 5 close, 1 differs | 6–13 % fewer for five isolates at ≥ 200 bp; BAS-3 has 25 % more |
| N50 | 2 match, 1 close, 3 differ | BAE-3 and BAS-1 within 4 bp; others −26 % to +48 % |
| Reads | 6 differ | The paper’s counts are 92.9–94.2 % of the read pairs in SRA |
| Coverage | 6 differ | The paper’s values are 0.82–0.90 times our mapped depth |
| Resistance genes | 1 match, 5 close | Below |
| Plasmid replicons | not checked | irit has no replicon typing (PlasmidFinder) |
- Contig cut-off: the paper doesn’t say which minimum contig length Table 1 uses. The agent
ran
assembly_statsat both 500 bp (QUAST’s default) and 200 bp (what GenBank accepts).- At 200 bp, the genome sizes agree to 43–6,711 bp; at 500 bp the gaps are about ten times larger.
- So Table 1’s size and contig count most likely describe the deposited 200 bp assemblies. Its GC fits the 500 bp numbers slightly better.
- Reads and coverage:
- Reads: the paper’s counts are a steady 93–94 % of the pairs in SRA, below even our post-fastp counts (99.3–99.6 %).
- Coverage: our mapped depth (CoverM) and Shovill’s own estimate agree within 3 %. The paper’s coverage is 10–18 % lower, and its lower read counts explain only about 6 % of that.
- The paper doesn’t say how it counted either.
- Resistance genes were found with a different tool (AMRFinderPlus instead of ResFinder), so
some differences are naming:
- Missing: aac(6’)-Iaa, a chromosomal Salmonella gene, in all three Salmonella isolates. ResFinder reports it; this AMRFinderPlus database doesn’t, with or without the organism option.
- Names: TEM-1A/1B differ only by silent changes, so AMRFinderPlus, which works on proteins, reports blaTEM-1. For BAE-2, AMRFinderPlus calls aac(3)-IIe where the paper has aac(3)-IIa, and aadA2 where it has aadA17.
- Partial hits: several genes sit at contig ends, for example tet(A) in BAE-3, found as a 59 % fragment. A cut-off on gene coverage would explain why ResFinder didn’t list it.
Species, afterwards
Section titled “Species, afterwards”The species check the agent couldn’t run was run afterwards with gtdbprep ani --species-only
(GTDB R232) on the agent’s six assemblies. It fetched 23 sketches (75 MB) in 30 s.
| Isolate | Closest GTDB reference (ANI, AF) | GTDB R232 species |
|---|---|---|
| BAE-1 | GCA_047199055.1 (96.91, 0.85) | ECMA0423 sp047199055 |
| BAE-2 | GCF_003697165.2 (96.88, 0.78) | Escherichia coli |
| BAE-3 | GCA_047199055.1 (96.83, 0.87) | ECMA0423 sp047199055 |
| BAS-1, BAS-2, BAS-3 | GCF_000006945.2 (98.42–98.61) | Salmonella enterica |
Two of the three “E. coli” get a placeholder genus name in GTDB R232.
- Where they land: BAE-1 and BAE-3 are closest to a reference GTDB R232 files under the placeholder genus ECMA0423. The E. coli species representative is the next hit, at 96.6.
- Not a new genus: GTDB R232 splits the E. coli/Shigella complex into several species with placeholder genus names. The E. coli K-12 test genome in the ANI write-up lands on another of them (GCA_047199095.1, G047199095).
- What to report: by the usual clinical definition all three are E. coli, as the paper says. A GTDB name in a report on E. coli needs that explanation next to it.
What the agent found in the paper
Section titled “What the agent found in the paper”- BAS-3’s twelve blaTEM alleles.
- What Table 1 says: TEM-1B, -30, -70, -104, -148, -176, -198, -207, -217, -220, -230 and -234.
- What the assembly has: one blaTEM, cut off at a contig end (62.6 % coverage, 100 % identity).
- Why: a fragment matches many TEM alleles equally well, and ResFinder lists every tie. The table should say one partial blaTEM, not twelve β-lactamases.
- Read counts and coverage don’t follow from the deposited reads, and the paper doesn’t say how it counted them (above).
- Phenotype without a genotype:
- BAS-1 and BAS-2 are reported resistant to ceftriaxone and ertapenem. Neither the paper’s list nor AMRFinderPlus finds an ESBL, AmpC or carbapenemase gene in them, only blaTEM-1.
- BAE-2 (ertapenem resistant, with only blaDHA-1) is similar.
- The cause could be porin or efflux changes that neither tool reports, or the susceptibility test itself.
- Allele calls at contig ends:
- BAE-2 has several genes as partial copies at contig ends: aac(3)-IId 53.5 %, aadA2 65.4 % and tet(A) 74.4 %.
- In short-read drafts, gene presence is reliable but allele-level names for such genes are not.
The agent also judged that BAE-3’s larger assembly (5.77 Mb) is accessory DNA rather than a second organism. CheckM2 contamination is 0.61 %, and the paper lists the most plasmid replicons for this isolate.
What it says about irit
Section titled “What it says about irit”Fixed after the run:
- isotable ignored the sheet’s read files. The sheet listed
reads/SRR30733615_1.fastq.gzfor BAE-1, yet each leanqc report got a row of its own (12 spurious “incomplete” rows). With--sheet, a report named after a read file in the sheet’s FASTQ columns now joins that sample. On the agent’s outputs the table now has exactly the six samples. - Contig cut-off: isotable always counted contigs of 500 bp or more, so 200 bp statistics
couldn’t be compared in the table. It now has
--min-contig(irit-mcpmin_contig). - Output names: irit-mcp refused AMRFinderPlus, mlst and CoverM outputs not named
<sample>.<tool>.tsv. That blocked a side run that shouldn’t join the table. Other names are now allowed, with a warning that isotable won’t read them. dblock_envin a new folder: it errors until a project exists. Its description now says to calldblock_addfirst.
Setup and environment, not irit code:
- sheetgen missing and the species-check wording: above.
- CheckM2: the Bioconda environment has scikit-learn 1.9.1 against models saved with 1.6.1, and irit-mcp passes on scikit-learn’s warning. The results look plausible (100 % complete for all six), but the environment should pin scikit-learn.
- AMRFinderPlus warns on every call that
$CONDA_PREFIXisn’t set; harmless when the database is given. - Shovill 1.4.2: it needs at least 8 GB for
--ram, the same as the cap. It worked here.
Gaps, now on the roadmap:
- Plasmid replicon typing (PlasmidFinder or similar) and ResFinder. A whole Table 1 row couldn’t be checked, and the resistance comparison had to cross tools.
- Salmonella serotyping (SISTR or SeqSero2), and a Kraken2 tool. dblock can already fetch Kraken2 databases, but irit-mcp can’t run Kraken2.
- Clonal complex: mlst doesn’t report it. A lookup table from PubMLST would cover it.
- Read checksums: the agent couldn’t check the reads’ MD5s. During this work, four ONT downloads (two files, twice each) arrived corrupt with no download error; Flye only noticed partway through (nanopore write-up). irit has no way to check read files against ENA/SRA checksums.
The brief, the agent’s transcript, its report, the run record and replay.sh are kept in the
project’s lab folder, not in the repository. The paper is CC BY 4.0; its Table 1 is summarised here
and not copied.