Structural Bioinfo Skills
skblnw/structural-bioinfo-skills
Automação de tarefas estruturais em biologia computacional
Ver no GitHubConjunto de skills Python para biologia estrutural que automatizam tarefas comuns: busca de estruturas no PDB, extração de monômeros e ligantes, mapeamento de epítopos para proteínas parentais, e busca de complexos pMHC/TCR. Usa apenas stdlib Python, sem dependências externas ou chaves de API.
Instalação
npx convez-ai skblnw/structural-bioinfo-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Busca e extração de estruturas proteicas do RCSB PDB
- Mapeamento de epítopos imunológicos para proteínas parentais
- Análise de complexos pMHC/TCR para pesquisa em imunologia
- Predição de estrutura secundária em modelos AlphaFold
Veredito

Ferramenta especializada e bem-construída para pesquisadores em biologia estrutural e imunologia. Elimina repetição de tarefas rotineiras com Python puro, mas requer conhecimento técnico em bioinformática.
README
A tradução para o português deste README ainda está na fila.
Structural Bioinfo Skills
Four Claude Code skills for everyday structural-biology grunt work: fishing the right PDB entries out of RCSB, splitting them into clean monomer+ligand files, mapping immune epitopes back to their parent proteins, and collecting pMHC / TCR-pMHC complexes. Stdlib Python only — no pip install, no API keys.
Quickstart
In Claude Code:
/plugin marketplace add skblnw/structural-bioinfo-skills
/plugin install structural-bioinfo
Then just ask Claude in natural language — the right skill triggers on intent. No CLI to memorise.
Why these exist
Every structural-bio project starts the same way: paste the same RCSB GraphQL query, the same UniProt chain-resolution lookup, the same IEDB peptide-search hack, into yet another notebook. These skills are the versions actually used in my pipeline, cleaned up and given triggers Claude can recognise. Hand it a peptide, a UniProt accession, or a PDB ID — get back a CSV, a clean PDB, or a Markdown+HTML report.
The skills compose: holostructure search → extractor for drug-target work; epitope-host → pMHC/TCR search for immunology work. Each is also fine on its own.
Skills
- pdb-holostructure-search — Search the RCSB Protein Data Bank for structures of a target protein (or its homologs) bound to small-molecule ligands. Outputs a CSV table ready for batch extraction.
- pdb-extractor — Download PDB/CIF structures from RCSB, resolve protein chains via UniProt (GraphQL), and extract clean PDB files containing only the target monomer and its bound ligands. Handles mmCIF-to-PDB conversion automatically. Supports single-PDB and CSV batch modes.
- search-epitope-host — Map an epitope peptide (or a list of them) to its parent host protein via the IEDB IQ-API, verify the position on the canonical UniProt sequence, and report the best 3D structure: experimental PDB entries first, AlphaFold prediction as fallback. Outputs Markdown + HTML reports.
- epitope-secondary-structure — Consume a
search-epitope-hostoutput bundle and compute, for each epitope, the secondary structure it adopts in its parent — a per-position probability over the 8 DSSP states (H,G,I,E,B,T,S,C) aggregated across the parent's experimentally solved structures, plus a simplified 3-state (H/E/C) consensus. Prioritizes experimental PDBs and considers all of them; AlphaFold-only parents are skipped and flagged. 8-state DSSP is computed locally withmkdssp(the RCSB Data API serves only coarse 3-state); RCSB SIFTS is used for the UniProt→structure residue mapping. - af-secondary-structure — Assign per-residue secondary structure to an AlphaFold model (or any PDB/mmCIF) — the 8-state DSSP code (H,G,I,E,B,T,S,C) and the reduced 3-state (H/E/C) for every sequence position — with built-in pLDDT confidence filtering so unreliable disordered regions can be flagged or masked. Reads pLDDT from the B-factor column, classifies the four AlphaFold confidence bands (very-low/low/confident/very-high), masks low-pLDDT positions (
X, cutoff configurable, default 70), and reports the contiguous disordered segments. DSSP is computed with mdtraj (nomkdsspbinary; reads AlphaFold mmCIF directly, which mkdssp 3.x cannot). The AF-model counterpart toepitope-secondary-structure(which is epitope-centric and uses experimental PDBs). Handles one file or a whole batch/directory. - pdb-pmhc-tcr-search — Search the RCSB Protein Data Bank for pMHC and TCR-pMHC complex structures by epitope peptide sequence. Auto-classifies MHC class I vs II; splits TCR-pMHC complexes into derivative
<pdb>_pmhc.pdband<pdb>_tcr.pdbfiles for downstream docking / MD / pharmacophore work. Redundant copies are kept and grouped; the lowest-resolution member is flaggedrepresentative. - esm-featurize — Turn one or more protein sequences into ESM Cambrian (ESM C 300M) embeddings, written to disk as compressed NumPy archives. One sequence in → one
.npzout. Representation-only — no generation, masking, structure prediction, or Forge API. Requires a localesmcconda env with theesmpackage. - pharmacophore-analyzer — Compute a 3D pharmacophore from a protein-ligand co-crystal PDB using RDKit. Identifies the bound ligand, restores correct bond orders and hydrogens via a SMILES template (auto-fetched from the RCSB Chemical Component Dictionary), runs the
BaseFeatures.fdeffeature factory on the ligand, and keeps only features that engage the protein (structure-based / interaction-filtered, LigandScout-style). - pharmacophore-report-generator — Consolidate per-PDB pharmacophore JSONs from N≥3 co-crystals of the same receptor into a single screening-ready markdown report. Surfaces conserved receptor residues that anchor each feature, computes retention statistics, and — when no consensus is supplied — derives one via Kabsch alignment of the orthosteric set followed by greedy single-linkage clustering.
- pharmacophore-dashboard — Turn the same per-PDB pharmacophore JSONs into a single self-contained, interactive HTML report — the visual counterpart to
pharmacophore-report-generator's markdown. Embeds all data and renders a 3D pharmacophore explorer (3Dmol.js: receptor pocket + ligand + colored feature spheres + H-bond/salt-bridge lines), feature-retention charts, sortable per-structure and residue-engagement tables, and the consensus virtual-screening query. Auto-discovers structures, locates the cleaned PDBs, computes statistics, and overlays a consensus query when one is present.
Typical pipeline
# 1. Find every holostructure of ATR
python plugins/structural-bioinfo/skills/pdb-holostructure-search/scripts/search_holostructures.py Q13535 -o atr_holo.csv
# 2. Download and extract clean monomer+ligand PDBs
python plugins/structural-bioinfo/skills/pdb-extractor/scripts/download_and_extract.py --csv atr_holo.csv --uniprot Q13535 --output-dir raw/
For an epitope-side workflow, search-epitope-host and pdb-pmhc-tcr-search are the entry points.
Requirements
- Python 3.6+ (stdlib only — no pip install)
- Internet access for RCSB PDB, UniProt, IEDB, and the AlphaFold DB
Updating
This repo is a downstream mirror of my local plugin at ~/.claude/plugins/structural-bioinfo/. Sync is automated via the sync-structural-bioinfo skill living at ~/.claude/skills/sync-structural-bioinfo/ — ask Claude to "sync structural-bioinfo" and it mirrors, commits, and pushes in one shot.
License
MIT — see LICENSE.