Structural Alignment (bioSkills)

A Claude Code skill for the case where sequence alignment has run out of signal: pick the right structural aligner, run it, and read TM-score, DALI Z-score, RMSD and lDDT against the cutoffs that actually mean something.


Type	Claude Skill
Supplier	GPTomics bioSkills (community OSS, MIT)
Availability	GA — part of the bioSkills collection
Pricing	Free / OSS (MIT). The aligners are installed separately and are individually free: Foldseek and FoldMason (GPL-3.0, Steinegger Lab), TM-align / US-align (Zhang Lab, free for academic use), open-source PyMOL. DALI is used as a web server or via DaliLite from the Holm lab
Capabilities	Read/Write — Claude runs the skill’s workflow locally (Bash/Python), not as an MCP tool
Verified	works · 2026-08-17
Security	caution · 2026-08-17 — GPTomics/bioSkills is now archived upstream; MIT and skill dir confirmed unchanged

How to install

bioSkills is not an npm package — skills are plain markdown/code read directly by the agent. Clone the repo, then either run the installer for the whole category or copy the single skill directory.

Claude Code — clone and install via the bundled script:
```
git clone https://github.com/GPTomics/bioSkills
cd bioSkills
./install-claude.sh --categories "alignment"
```
The installer copies matching skills into ~/.claude/skills/ (default target). Use ./install-claude.sh --list to preview the skills first.
Claude Code / other agents — copy just this one skill:
```
cp -r bioSkills/alignment/structural-alignment ~/.claude/skills/
```
(run from inside your clone — the previous step left you in bioSkills/; otherwise replace bioSkills/ with the absolute path of your clone, e.g. /Users/you/repos/bioSkills).
Prerequisites — the aligners are CLI tools the skill drives, none are bundled. All four are on bioconda (versions checked 2026-08-15, and each satisfies the skill’s stated minimum):
```
conda install -c conda-forge -c bioconda foldseek tmalign usalign foldmason
```
foldseek 8.ef4e960 (skill wants 8+), tmalign 20240303 (20220412+), usalign 20241201 (20231222+), foldmason 2.7bd21ed (1+). Confirm with foldseek version, TMalign (prints usage), USalign -h, foldmason version.
Python side — BioPython 1.83+ for Bio.PDB.Superimposer, plus open-source PyMOL 3.0+ if you want to render the superposition:
```
pip install "biopython>=1.83"
conda install -c conda-forge pymol-open-source
```
DALI — not a conda package. It is reached through the Dali server (free, submit-and-wait) or the standalone DaliLite distribution from the Holm lab. Treat it as an external service: uploading an unpublished structure sends it off your machine.

What it does

Routes a structure-comparison question to the tool that answers it, then interprets the score:

Database-scale search — Foldseek’s 3Di alphabet indexes backbone geometry as sequence, giving structural search at ~10³–10⁶ structures/second; Foldseek-Multimer extends this to complexes.
Pairwise superposition — TM-align for single chains where the residue correspondence is unknown; US-align for multi-chain complexes, RNA and DNA (-mm 1 -ter 0); Bio.PDB.Superimposer (Kabsch/SVD) when the atom correspondence is already known and you just need the transform.
Flexible and distance-matrix methods — FATCAT for alignments needing twists and chain breaks, CE for combinatorial extension, DALI for distance-matrix alignment and its Z-score.
Structural MSA — FoldMason for multiple structure alignment at scale; T-Coffee Expresso / 3D-Coffee for hybrid sequence–structure alignment when only some members have structures.

Score cutoffs the skill carries: TM-score > 0.5 = same fold, > 0.8 = equivalent topology, < 0.2 = random similarity — and for chains under 60 residues a length-aware Gumbel p-value instead of the raw TM-score. DALI Z-score > 20 definitely homologous, 8–19 probable, 2–8 candidate (confirm with TM-score), < 2 not significant. RMSD < 2 Å over > 100 residues is a strong superposition, < 1.5 Å excellent. lDDT > 0.6 for a correctly modelled residue.

Primary use cases: remote-homology detection below the sequence twilight zone, deciding whether two folds are the same, superposing predicted against experimental models, building a structural MSA.

Notes

The skill’s organising idea is a sequence-identity ladder that tells you when to stop aligning sequences: at ≥ 40% identity ordinary sequence dynamic programming is sufficient; 25–40% calls for sensitive sequence methods (MMseqs2, jackhmmer, HHsearch); 15–25% for profile–profile methods, or Foldseek if structures exist; below 15% sequence alignment is noise and structural or protein-language-model aligners are required. The stated exception is families under strong structural or functional constraint, which can stay reliable down to roughly 12–15% identity when confirmed by a profile–profile search.

The most consequential practical warning concerns predicted models: mask residues with pLDDT < 70 before Foldseek indexing — their backbone coordinates encode as effectively random 3Di letters and contaminate hits below about TM-score 0.4. Feeding a full-length AlphaFold model with long disordered tails into a structural search is the standard way to generate confident-looking nonsense.

Two boundaries worth keeping straight. RMSD is not a property of a structure pair — it depends on the superposition and the atom selection — so TM-score, which is length-normalised, is the comparable number; and Bio.PDB.Superimposer does not solve the correspondence problem, so it is the wrong tool the moment you do not already know which atom pairs with which.

Upstream skill front-matter name is bio-alignment-structural; upstream directory alignment/structural-alignment; primary_tool is Foldseek. Complements Foldseek Structural Search (the standalone Foldseek surface), Protein MCP Server (TM-align/jFATCAT and Foldseek reachable as MCP tools rather than local CLIs), Geometric Analysis (RMSD, SASA and dihedrals on a single structure), Structure Validation (pLDDT and PAE interpretation before you index a model), and Multiple Sequence Alignment for the sequence-side step this skill takes over from.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.