Build a phylogenetic tree from a set of sequences

Hand Claude Code a FASTA of homologous sequences (viral genomes, a microbial marker gene, a protein family); get back a trimmed multiple alignment, a model-selected maximum-likelihood tree with bootstrap support, and a publication-ready annotated tree figure — without hand-wiring the MAFFT → IQ-TREE 2 → ETE3 command chain.

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

Problem

A microbiologist or viral phylodynamicist who has just pulled a set of sequences — outbreak isolate genomes from GISAID/GenBank, 16S/marker amplicons from a culture collection, or a protein family from UniProt — needs the same canonical pipeline every time: align the sequences (MAFFT), optionally trim poorly aligned columns, pick a substitution model, infer a maximum-likelihood tree with branch support (IQ-TREE 2’s ModelFinder + ultrafast bootstrap, or FastTree when speed matters), root it, and render an annotated figure. The individual tools are standard, but stitching them together — getting the MAFFT flags right, parsing ModelFinder output, passing the right alignment to IQ-TREE, mapping tip labels to metadata in the figure — is repetitive boilerplate that is easy to get subtly wrong (a mis-set bootstrap count, an unrooted tree presented as rooted). “Solved” looks like: drop a FASTA in the project, name the metadata column you want to colour tips by, and get back aligned.fasta, tree.treefile with support values, and a labelled tree.png.

  1. Install the Phylogenetics (Claude Skill). From the K-Dense scientific-agent-skills collection:

    npx skills add K-Dense-AI/scientific-agent-skills

    Enable the phylogenetics skill when prompted. It wraps MAFFT (alignment), IQ-TREE 2 (maximum-likelihood with ModelFinder + ultrafast bootstrap), FastTree (fast NJ/ML), and ETE3/FigTree (visualization). The skill declares its own Python/binary dependencies in its SKILL.md; install them on first use.

  2. Place the inputs in the project. You need:

    • sequences.fasta — your homologous nucleotide or amino-acid sequences, one record per taxon, with stable IDs in the headers.
    • metadata.tsv (optional) — one row per sequence ID, with columns you want to annotate tips by (collection date, host, lineage, country).

  3. Invoke the skill with the file paths and the analysis you want. A minimal prompt:

    Use the phylogenetics skill on data/sequences.fasta.

Steps:
  1. Align with MAFFT (use --auto; report the strategy it selected).
  2. Trim alignment columns with >50% gaps; report columns kept vs
     dropped and write out/aligned.fasta.
  3. Infer a maximum-likelihood tree with IQ-TREE 2: let ModelFinder
     pick the substitution model, run 1000 ultrafast bootstrap
     replicates and 1000 SH-aLRT replicates. Write out/tree.treefile.
  4. Midpoint-root the tree.
  5. Render the tree with ETE3 to figures/tree.png: show bootstrap
     support on internal nodes, colour tips by `host` from
     data/metadata.tsv, and collapse nodes with <50% support.

  4. Read the model and support critically. Note which substitution model ModelFinder selected (GTR+G, HKY, LG+G, etc.) and report it — reviewers will ask. Treat internal nodes with ultrafast-bootstrap < 95% or SH-aLRT < 80% as unresolved; do not over-interpret a poorly supported clade. Confirm the rooting choice is defensible: midpoint rooting is a convenience, not a biological claim — if you have a known outgroup, re-run with -o <outgroup_id> instead.

  5. Hand off downstream. The Newick tree.treefile drops into ETE Toolkit for orthology/paralogy event detection or NCBI-taxonomy annotation, or into BEAST/Nextstrain for time-scaled phylodynamics (outside this skill). For microbial-community context, the same FASTA can feed the 16S diversity recipe — that recipe consumes the tree this one produces as its UniFrac input.

Why this assembly

Rung 2 of the simplicity ladder. Plain Claude Code can write the MAFFT/IQ-TREE invocations from scratch, but the parameter surface is wide (alignment strategy, trimming threshold, model selection, bootstrap type and count, rooting) and small slips silently change the tree — an unrooted tree drawn as rooted, or a too-low bootstrap count that overstates support. The skill encodes the canonical align → model-select → ML-infer → annotate workflow as one discoverable action with sensible defaults, which is the right grain for a single-stage analytical task. No need to escalate to a multi-tool harness or an autonomous system: tree building is a well-defined, self-contained problem.

Availability

Fully open. MAFFT (BSD), IQ-TREE 2 (GPL-2.0), FastTree (GPL-2.0), and ETE3 (GPL-3.0) are all open-source; the K-Dense skill wrapper ships in the same OSS collection (license not stated upstream). The skill makes no external API calls — all computation runs locally on your sequences. FASTA and Newick are open formats. No subscription, institutional account, or API key required.

Compute requirements

Laptop-sufficient for typical inputs. A few hundred sequences of viral-genome length (~30 kb) or a 16S marker (~1.5 kb) align in seconds-to-minutes with MAFFT --auto and infer in a few minutes with IQ-TREE 2 + 1000 ultrafast-bootstrap replicates on a modern multi-core laptop with 16 GB RAM; IQ-TREE parallelizes across cores (-T AUTO). The ML tree search is the heaviest step. Datasets in the thousands of full genomes push you toward FastTree (the skill exposes it) for a fast approximate tree, or toward an HPC IQ-TREE run. No GPU is used.

Evidence

Proposed. No documented end-to-end attempt of “Claude Code + the phylogenetics skill on a real sequence set” with quantitative pass/fail is known to the curator. The closest evidence is component-level and class-level:

Alternatives considered

  • Plain Claude Code, no skill (rung 1). Works — Claude can write the MAFFT and IQ-TREE 2 commands directly if the binaries are on PATH. Reach for this when you need a one-off custom step the skill does not expose (a non-standard partition model, a constraint tree). Reach for the skill when you want a repeatable, documented prompt template across studies.
  • Nextflow nf-core/phylogeny-style pipeline (no agent). The right tool when you run the same tree build at scale, on a cluster, hundreds of times — but that is workflow-automation overkill for an interactive one-off. See the Nextflow catalog page if your use case is batch.
  • Biomni (rung 4). The Biomni agent exposes alignment/tree primitives inside a 150-tool environment. Reach for it when the tree is one node of a larger multi-stage analysis (e.g., assemble genomes → call variants → build tree → date the MRCA); reach for the focused skill when the tree is the whole job.

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.