Identify a bacterial isolate from its 16S rRNA sequence

Hand Claude Code a 16S rRNA gene sequence from a bacterial isolate (a colony-PCR Sanger read or an assembled 16S contig) and get back a ranked table of the closest named reference organisms, with percent identity and coverage interpreted against the accepted species/genus thresholds — anchored to the exact reference-database release you searched.

   
Problem class Data analysis
Subject areas Immunology and Microbiology
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

You picked a colony, ran 16S colony PCR with the universal primers (27F/1492R), and got back a Sanger read (or you pulled the 16S gene out of a draft assembly). The everyday microbiology question is “what is this?” The canonical answer is a similarity search of the 16S sequence against a curated database of type-strain 16S sequences, then reading the top hit against the accepted thresholds: ≥98.7% identity to a type strain supports a species-level match, and ~94.5–95% supports genus level; below that you may be looking at a novel taxon (Kim et al., Int. J. Syst. Evol. Microbiol. 2014; Yarza et al., Nat. Rev. Microbiol. 2014).

The mechanics are a blastn search, but the footguns are real: trusting a top hit that only covers half the read (a chimeric or truncated Sanger trace); reporting “species X” off a 96% hit that only supports genus; searching the general nt database (full of misannotated environmental clones) instead of a curated type-strain 16S set; and — the reproducibility killer — not recording which database release the call came from, so it can’t be re-checked when the database moves. “Solved” looks like: point the agent at your 16S FASTA and one pinned 16S reference database, get back a hits.csv ranked by identity with coverage and a threshold-based assignment column, plus a provenance record naming the exact database release.

  1. Install the BLAST (Bio-MCP) server. It wraps NCBI BLAST+ and exposes makeblastdb, blastn, and blastp as MCP tools. Follow the catalog page’s clone-and-pip install -e . steps and register it with claude mcp add. The BLAST+ binaries must be on PATH first.

  2. Download and pin a curated 16S reference database. Use NCBI’s targeted 16S_ribosomal_RNA BLAST database (type-strain and reference 16S sequences), fetched with the BLAST+ helper:

    update_blastdb.pl --decompress 16S_ribosomal_RNA

    Record the release date printed by update_blastdb.pl (the set is versioned by NCBI build date) and compute a sha256 of the downloaded .tar.gz. These are your pinned inputs. (SILVA or the GTDB 16S export are equally valid pinned alternatives — pick one and record its release.)

  3. Confirm the read is clean before searching. A raw Sanger trace has low-quality ends and may need reverse-complementing; a good 16S query is ~1,300–1,500 bp of unambiguous sequence. Trim obvious low-quality tails and drop reads with excessive Ns — a short or chimeric query produces confident-looking but wrong hits.

  4. Have the agent run the search into a committed script. A prompt:

    Using the bio-blast MCP server, write me a script identify_16s.py that:
  1. Runs blastn of my query (query_16s.fasta) against the local
     16S_ribosomal_RNA database, with -perc_identity 90,
     -max_target_seqs 10, and outfmt 6 (qseqid sseqid pident length
     qlen slen mismatch gapopen evalue bitscore stitle).
  2. Computes query coverage = length / qlen for each hit and keeps
     hits with coverage >= 0.90.
  3. Sorts hits by pident (then bitscore) and writes hits.csv with
     the reference accession, organism name (from stitle), pident,
     coverage, alignment length, and evalue.
  4. Adds an `assignment` column per the standard thresholds:
     pident >= 98.7 -> "species-level"; 94.5 <= pident < 98.7 ->
     "genus-level"; pident < 94.5 -> "below genus / possible novel".
  5. Writes provenance.json: BLAST+ version, the 16S database
     release date and .tar.gz sha256, the query_16s.fasta sha256,
     the identity/coverage cutoffs, run date, and model id.
Commit identify_16s.py; do not paste results back as prose I
can't audit.

    Pin the environment with a requirements.txt (BLAST+ via conda, plus pandas). Keep identify_16s.py, the pinned env, and provenance.json under version control alongside the query FASTA and the database release string + hash.

  5. Read the assignment critically. A single 16S gene resolves genus reliably but often cannot separate closely related species (e.g., within Bacillus or the Enterobacteriaceae), where multiple references tie at >99%. Treat a species-level call from 16S alone as provisional: if the top few hits are different species at near-identical identity, report the genus and flag the ambiguity. Definitive species delineation needs whole-genome average nucleotide identity (ANI) or multi-locus typing, not 16S.

  6. Re-run and hand off. Because the database release and cutoffs are pinned in provenance.json, re-running identify_16s.py reproduces the assignment until you deliberately bump the database (which the provenance record makes visible). The identified organism name feeds the genus/species hint of the bacterial-genome-annotation recipe; the 16S sequence itself can join a set for the phylogenetic-tree recipe to place the isolate among relatives.

Why this assembly

Rung 2 of the simplicity ladder — one cataloged MCP server does the whole job. Plain Claude Code (rung 1) can shell out to blastn, but the BLAST MCP gives the agent first-class makeblastdb/blastn tools and keeps the invocation and the pinned database release auditable, which is exactly what makes an identification reproducible. Rung 3+ is unnecessary: a similarity search of one gene against one curated database, read against published thresholds, is a single well-bounded step. This recipe is the single-isolate identification counterpart to the community-level 16S diversity recipe and the multi-isolate resistome screen — different question, same BLAST-MCP tool.

Availability

Fully open. The BLAST MCP server is MIT-licensed and BLAST+ is public-domain NCBI software. The 16S_ribosomal_RNA database is freely downloadable from NCBI; SILVA and GTDB alternatives are free for academic use (check SILVA’s terms for commercial use). All computation runs locally on your sequence — no account, no API key, no upload.

Compute requirements

Laptop-sufficient. The 16S_ribosomal_RNA database is a few hundred MB; a single blastn of one ~1.5 kb query against it returns in seconds on a modern laptop with 8 GB RAM. No GPU. Batch-identifying a plate of hundreds of colony reads is embarrassingly parallel — loop the same script per query, or pass a multi-FASTA and group results by qseqid.

Evidence

Reported. 16S rRNA similarity search for isolate identification is one of the most-used procedures in microbiology, and the identity thresholds this recipe applies are the community-standard reference points: 98.7% for species and ~94.5% for genus (Kim et al., Int. J. Syst. Evol. Microbiol. 64:346 (2014); corroborated by the SILVA-scale analysis of Yarza et al., Nat. Rev. Microbiol. 12:635 (2014)). Countless species-description and isolate-characterization papers run exactly this blastn-against-a-curated-16S-database step (e.g., the type-strain characterizations returned for any new sp. nov., which report the top 16S identity before confirming with ANI/dDDH).

No head-to-head benchmark of the agent-driven BLAST-MCP assembly versus a hand-typed blastn command is published — the MCP loop buys first-class BLAST tools, a pinned database release, and a recorded provenance file, not a new method. The underlying BLAST+ homology search and the identity thresholds are the validated components; the recipe’s contribution is making the call reproducible.

Alternatives considered

  • Plain Claude Code, no MCP (rung 1). Fine if BLAST+ is already installed; the MCP gives the agent structured makeblastdb/blastn tools and a cleaner audit trail. Reach for rung 1 for a single one-off lookup you will not repeat.
  • NCBI web BLAST against nt (no agent). Convenient but two traps: the general nt database is polluted with misannotated environmental clones, so a curated 16S set is safer, and a web run leaves no pinned-release provenance. Use the web tool for a quick eyeball; use this recipe when the identification goes in a record.
  • Whole-genome ANI / MLST for definitive species calls. When 16S ties several species (common), reach for average nucleotide identity against reference genomes or multi-locus sequence typing. Neither an ANI tool (fastANI/pyani) nor an MLST caller is catalogued as a Claude tool today — surfaced as a missing component to the catalog curator. 16S identification is the fast first pass; ANI is the confirmation.

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.