Annotate a single bacterial genome assembly

Hand Claude Code one freshly assembled bacterial (or archaeal) genome; get back a standardized, NCBI-compatible annotation — coding sequences, rRNA/tRNA/ncRNA, CRISPR arrays, replicon features — plus a feature-count summary and a circular plot, all from a single skill.

   
Problem class Data analysis
Subject areas Immunology and Microbiology, Molecular and Cellular Biology
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Workstation with GPU

Problem

You have de novo assembled one isolate — a clinical strain, an environmental isolate, a lab construct — and the FASTA of contigs tells you nothing biological until it is annotated. Genome characterization papers almost always open with a single-isolate Bakta (or Prokka) run before any comparative or resistance analysis (Santhosh et al., BMC Genomics 2025). The task is to call genes consistently against a curated reference and emit standard files (GFF3, GenBank, FASTA of proteins) that every downstream tool consumes. “Solved” looks like: drop one assembly in, get back a version-stamped, database-pinned annotation with a feature table you can read at a glance and a CDS count you can sanity-check against the expected genome size.

The footguns are small but real: running an annotator with the wrong genus/translation-table assumption, forgetting to set --complete for closed genomes versus draft contigs, and not recording the database version (so the annotation can’t be reproduced later). This recipe surfaces them and stops at the annotated genome — comparative pan-genome work across many isolates is a separate, heavier recipe.

  1. Install the Bakta skill. It ships in the SciAgent-Skills collection — clone once and load as a plugin:

    git clone https://github.com/jaechang-hits/SciAgent-Skills

    Then inside Claude Code run /plugin install sciagent-skills and confirm Bakta appears under /plugin → Installed. Bakta needs its UniRef-derived database downloaded once (full DB ~70 GB; a light DB is available if disk is tight). Use the Prokka skill instead for legacy/Prokka-compatible output or a non-bacterial kingdom.

  2. Stage the assembly. One genome per file — assembly.fasta (or .fna). Note up front whether it is a closed/complete genome (single circular replicon[s]) or a draft (many contigs); it changes one flag.

  3. Run Bakta with the right replicon assumption. A minimal prompt:

    Use the Bakta skill on assembly.fasta with the full database.
This is a DRAFT genome (many contigs) — do not pass --complete.
Settings:
  - genus/species hint: <Escherichia coli>   (improves gene naming)
  - default translation table 11 for bacteria
  - write GFF3, GenBank, and the protein FASTA to annot/
Then report a feature summary: CDS, rRNA, tRNA, tmRNA, ncRNA,
CRISPR arrays, and the Bakta DB version used.

    For a closed genome, say so and have Claude add --complete. The database version in the report is what makes the run reproducible — keep it.

  4. Sanity-check the feature counts. Open the summary: a typical 5 Mb bacterial genome carries roughly 4,500–5,500 CDS and a full complement of tRNAs plus several rRNA operons. A very low CDS count, or zero rRNA, usually means a fragmented/contaminated assembly or a wrong kingdom — fix the input before trusting the annotation.

  5. (Optional) Layer on AMR / virulence calls. Bakta annotates structure (CDS, ncRNA, CRISPR, oriC) but is not a dedicated resistance caller. For an AMR/virulence profile, run the protein FASTA from annot/ against a curated database (CARD, VFDB) as a separate step — published single-isolate characterizations pair Bakta with exactly this (Santhosh et al., 2025). No such database is catalogued as a Claude skill today, so run it from the CLI.

  6. Hand off downstream. The GFF3 is the input to the bacterial pan-genome recipe once you have more isolates; the protein FASTA feeds homology and functional-enrichment work.

Why this assembly

Rung 2 of the simplicity ladder — one cataloged skill solves the whole problem. Annotation is a single, well-bounded step with a fragile parameter surface (replicon completeness, translation table, database version) that the Bakta skill encapsulates. Rung 1 (plain Claude Code) can drive the Bakta binary, but Bakta is a heavyweight tool with a 70 GB database and easy-to-miss flags; the skill pins the invocation and surfaces the database version for reproducibility. Rung 3+ is unnecessary: this is one genome, one tool — the multi-genome comparative case is the separate, rung-3 pan-genome recipe.

Availability

Fully open. Bakta is GPL-3.0; the SciAgent-Skills wrapper is CC BY 4.0. All computation runs locally on your assembly — no account, no API key, no external upload. The Bakta database is a one-time free download. FASTA, GFF3, and GenBank are open formats.

Compute requirements

Workstation-class, driven by the database footprint rather than CPU. Bakta annotates a typical 5 Mb bacterial genome in a few minutes on a multi-core CPU; its alignment-free identification benefits from many cores and ~16 GB RAM, and the full database needs ~70 GB disk (use the light DB if disk-constrained, at some loss of annotation depth). No GPU is required — the “GPU workstation” tier reflects the RAM/disk and core count that make annotation comfortable, not a CUDA dependency.

Evidence

Reported. The Prokka/Bakta single-isolate annotation step is the field-standard opening move in bacterial genome characterization. A 2025 study assembled an environmental Burkholderia thailandensis isolate de novo and functionally annotated it with Bakta (then layered KEGG/GO/CARD/VFDB for AMR and virulence) — the exact single-genome assembly this recipe wraps (Santhosh et al., BMC Genomics 2025). Bakta itself outperformed contemporary annotators on functional annotation and database cross-references at comparable runtime (Schwengers et al., Microb. Genom. 2021).

No head-to-head benchmark of the agent-driven single-genome annotation versus a hand-typed Bakta command is published — the skill buys a pinned invocation, a recorded database version, and a feature-count sanity check, not a new method.

Alternatives considered

  • Plain Claude Code, no skill (rung 1). Fine if Bakta and its database are already installed and on PATH; the skill is worth it for the pinned flags and the reproducibility-friendly database-version report.
  • Prokka instead of Bakta. Reach for the Prokka skill when a downstream tool expects Prokka-style GFF3, or for a non-bacterial kingdom. Bakta is the better default for modern bacterial/archaeal annotation.
  • NCBI PGAP. The right choice when you intend to submit the genome to GenBank, which requires PGAP-conformant annotation. PGAP is not catalogued as a Claude skill and is heavier; use it at submission time, not for interactive characterization.
  • The pan-genome recipe (rung 3). Reach for it the moment you have several isolates and want core/accessory partitions — it runs this annotation step identically across every genome, then clusters with Roary.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.