tooluniverse-protein-structure-retrieval
Fetch and evaluate protein structures from major databases using experimental resolution or pLDDT quality metrics.
Install
mkdir -p .claude/skills/tooluniverse-protein-structure-retrieval && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/6365" && unzip -o skill.zip -d .claude/skills/tooluniverse-protein-structure-retrieval && rm skill.zipInstalls to .claude/skills/tooluniverse-protein-structure-retrieval
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Protein structure retrieval from RCSB PDB, PDBe, and AlphaFold with disambiguation, quality assessment (resolution, R-factor, pLDDT), and metadata. Distinguishes high-quality experimental (X-ray under 2 Angstrom) vs predicted vs medium-quality structures. Use for fetching protein structures, structure-quality comparison, and selecting structures for drug design or modeling.Key capabilities
- →Disambiguate proteins by organism
- →Fetch structure metadata from RCSB/PDBe
- →Assess structural quality via resolution or pLDDT
- →Retrieve ligand information
- →Compare experimental and predicted models
How it works
It queries standardized biological databases (RCSB, PDBe, AlphaFold) and filters results based on defined quality metrics (resolution/pLDDT).
Inputs & outputs
When to use tooluniverse-protein-structure-retrieval
- →Select high-resolution structures for drug design
- →Compare experimental vs. predicted protein models
- →Retrieve ligand information for specific protein IDs
- →Validate structural quality using pLDDT or resolution data
About this skill
Protein Structure Data Retrieval
Retrieve protein structures with disambiguation, quality assessment, and comprehensive metadata.
IMPORTANT: Always use English terms in tool calls. Respond in the user's language.
LOOK UP DON'T GUESS: Never assume PDB IDs, resolution, or availability. Always query RCSB/PDBe and AlphaFold to confirm.
Domain Reasoning
Not all structures are equal. X-ray <2 A is high-quality for drug design. Cryo-EM 3-4 A is good for fold but not side chains. AlphaFold is excellent for well-folded domains but unreliable for disordered regions. Always check pLDDT (AlphaFold) or resolution (experimental) before drawing conclusions.
Workflow
Phase 0: Clarify (if needed) → Phase 1: Disambiguate Protein → Phase 2: Retrieve Structures → Phase 3: Report
Phase 0: Clarification (When Needed)
Ask ONLY if: protein name ambiguous (e.g., "kinase"), organism not specified, unclear if experimental vs AlphaFold needed. Skip for: specific PDB IDs, UniProt accessions, unambiguous protein+organism.
Phase 1: Protein Disambiguation
# By PDB ID: direct retrieval
# By UniProt: get AlphaFold + search experimental structures
af_structure = tu.tools.alphafold_get_prediction(uniprot_id=uniprot_id)
# By protein name: search
result = tu.tools.PDBeSearch_search_structures(query=protein_name)
Identity Checklist
- Protein name/gene identified, organism confirmed
- UniProt accession (if available), isoform/variant specified (if relevant)
Phase 2: Data Retrieval (Internal)
Retrieve silently. Do NOT narrate the process.
pdb_id = "4INS"
# Search, metadata, quality, ligands, similar structures
result = tu.tools.PDBeSearch_search_structures(query=name)
metadata = tu.tools.get_protein_metadata_by_pdb_id(pdb_id=pdb_id)
exp = tu.tools.RCSBData_get_entry(pdb_id=pdb_id)
quality = tu.tools.PDBeValidation_get_quality_scores(pdb_id=pdb_id)
ligands = tu.tools.PDBe_get_structure_ligands(pdb_id=pdb_id) # data.ligands[]
similar = tu.tools.PDB_search_similar_structures(query=pdb_id, search_type="structure") # data.results[] with rank, score
# By UniProt accession: all PDB entries, and ligand-binding residues across entries
entries = tu.tools.PDBeSIFTS_get_all_structures(uniprot_accession=uniprot_id) # data.pdb_entries[]
sites = tu.tools.PDBe_KB_get_ligand_sites(uniprot_accession=uniprot_id) # data.ligands[]
# PDBe additional data
summary = tu.tools.pdbe_get_entry_summary(pdb_id=pdb_id)
molecules = tu.tools.pdbe_get_entry_molecules(pdb_id=pdb_id)
# AlphaFold (when no experimental structure, or for comparison)
af = tu.tools.alphafold_get_prediction(uniprot_id=uniprot_id)
Fallback Chains
| Primary | Fallback |
|---|---|
| RCSB search | PDBe search |
| get_protein_metadata | pdbe_get_entry_summary |
| Experimental structure | AlphaFold prediction |
| get_protein_ligands | PDBe_get_structure_ligands (per PDB ID) or PDBe_KB_get_ligand_sites (per UniProt accession) |
| No usable AlphaFold model (multi-domain protein, low pLDDT for a specific range) | SwissModel_get_models filtered to that residue range — the repository holds per-domain homology models AlphaFold's single full-length model may not resolve well |
SWISS-MODEL Repository (Homology Models + Re-Indexed Experimental Structures)
Use SwissModel_get_summary(uniprot_id=...) for a one-call "does a homology model exist and how good is it" check — it returns the single best-coverage entry across everything SWISS-MODEL has for that accession. Important: the "best" entry it returns can be an experimental PDB structure re-indexed by UniProt residue range (verified live: for EGFR/P00533, get_summary's best_model was an X-RAY DIFFRACTION entry, not a homology model) — check best_model.method, don't assume "best model" means "computed homology model."
To get genuine computed homology models specifically, call SwissModel_get_models(uniprot_id=..., provider="swissmodel") — provider="pdb" isolates the re-indexed experimental entries instead, and omitting provider returns both. Other filters: range (residue window, useful for large multi-domain proteins where one region has poor AlphaFold confidence) and template (a specific PDB template ID).
SwissModel_download_pdb(uniprot_id=..., provider=...) fetches the actual ATOM/HETATM coordinate text (not just a URL) for downstream docking/visualization — same filters as get_models. SwissModel_get_models_batch(uniprot_ids=[...]) resolves up to 250 accessions in one call, useful when checking model availability across a gene list before deciding which proteins need AlphaFold/ESMFold instead.
Quality-metric caveat (verified live): qmean_global/qmean_z_score were null on every real homology-model entry tested (EGFR's 2 SWISS-MODEL models) — QMEAN is frequently absent, not a metric you can always rely on. Use coverage (fraction of the UniProt sequence the model spans), template (which PDB structure it was built from), and method as the more consistently populated quality signals; treat QMEAN as a bonus when present, not a required check.
When to reach for SWISS-MODEL vs. AlphaFold: AlphaFold (above) is the default single-model reference for any UniProt-reviewed protein. Reach for SWISS-MODEL specifically when (a) you need multiple alternative models built from different templates to compare, (b) a large protein's AlphaFold confidence is poor in one region and a domain-specific homology model with better template coverage might do better there, or (c) you're checking many accessions at once and want batch lookup rather than N separate AlphaFold calls.
3D-Beacons (Meta-Aggregator Across All Structure Providers)
ThreeDBeacons_get_structure_summary(accession=...) is the best FIRST call when you don't yet know which structure source has the most/best coverage for a UniProt accession — it queries PDBe, SWISS-MODEL, AlphaFold DB, AlphaFill, and ModelArchive simultaneously and returns a by_provider count plus a by_category breakdown (EXPERIMENTALLY DETERMINED / TEMPLATE-BASED / AB-INITIO). Real example (verified live): EGFR/P00533 → total_structures: 426, by_provider: {"PDBe": 412, "SWISS-MODEL": 2, "AlphaFold DB": 6, "AlphaFill": 1, "ModelArchive": 5} — confirming this protein is overwhelmingly covered by experimental structures already, and that the 2 SWISS-MODEL entries (Phase above) are a small fraction of what's actually available; don't stop at SWISS-MODEL's own listing if 3D-Beacons shows a much richer PDBe count.
ThreeDBeacons_get_structures(accession=..., category=..., provider=..., max_results=...) returns the individual structure/model records (not just counts) — filter by category or provider to narrow down before fetching detail elsewhere. ThreeDBeacons_get_annotations(accession=..., type=..., provider=...) maps residue-level annotations (e.g. type="DOMAIN", type="BINDING") onto the protein's 3D models — real example: BRCA1/P38398 with type="DOMAIN" returns real domain-to-residue mappings, while a type="BINDING" query on P04637 can legitimately return zero annotations (annotation_count: 0) — an empty result here means no such annotation exists in the aggregated sources, not a broken call.
Workflow: use ThreeDBeacons_get_structure_summary as the first orientation step for any UniProt accession, before deciding whether to drill into PDBe/RCSB (Phase 2), AlphaFold, or SWISS-MODEL specifically.
PDB-REDO (Re-Refined Experimental Structures)
PDB-REDO automatically re-refines every X-ray PDB entry with current software/parameters, often improving on the original deposition's refinement quality. PDB_REDO_get_structure_quality(pdb_id=...) returns detailed refinement metrics (unit cell axes, B-factors, real-space/working correlation coefficients CCFFIN/CCWFIN, resolution DATARESH, completeness) for the re-refined structure — real example (verified live): PDB 4hjo (an EGFR structure) → DATARESH: 2.75, CCWFIN: 0.93, COMPLETED: 96.9. PDB_REDO_get_version_info(pdb_id=...) returns re-refinement provenance/versioning metadata for the same entry.
When to use: prefer PDB-REDO's re-refined metrics over the original PDB deposition's stated resolution/R-factors when precision matters for downstream drug-design decisions — re-refinement can meaningfully change R-free and even correct minor model errors from the original deposition.
Phase 3: Report Structure Profile
Present as a Structure Profile Report. Hide search process. Include:
- Search Summary: query, organism, experimental + AlphaFold structure counts
- Best Structure: PDB ID, UniProt, organism, method, resolution, date, quality assessment
- Experimental Details: method, resolution, R-factor, R-free, space group
- Composition: chains, residues (coverage%), ligands, waters, metals
- Bound Ligands: ligand ID, name, type, binding site
- Binding Site Details (for drug discovery): location, key residues, druggability
- Alternative Structures: ranked by quality with resolution, method, ligands
- AlphaFold Prediction: UniProt, model version, pLDDT confidence distribution, use cases
- Structure Comparison: resolution, completeness, ligands across structures
- Download Links: PDB/mmCIF/AlphaFold formats, database URLs
Quality Assessment
Experimental Structures
| Tier | Criteria |
|---|---|
| Excellent | X-ray <1.5A, complete, R-free <0.22 |
| High | X-ray <2.0A OR Cryo-EM <3.0A |
| Good | X-ray 2.0-3.0A OR Cryo-EM 3.0-4.0A |
| Moderate | X-ray >3.0A OR NMR ensemble |
| Low | >4.0A, incomplete, or problematic |
Resolution Use Cases
<1.5A: atomic detail, H-bond analysis. 1.5-2.0A: drug design. 2.0-2.5A: structure-based design. 2.5-3.5A: overall architecture. >3.5A: domain arrangement only.
AlphaFold Confidence (pLDDT)
90: very high, experimental-like. 70-90: good backbone. 50-70: uncer
Content truncated.
When not to use it
- →Analyzing non-protein biological data
- →When structure files are too large for memory
Limitations
- →Does not verify experimental method validity
- →Depends on availability in public databases
How it compares
It enforces strict biological quality checks before data retrieval, preventing the use of low-resolution or inappropriate structures.
Compared to similar skills
tooluniverse-protein-structure-retrieval side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| tooluniverse-protein-structure-retrieval (this skill) | 1 | 5mo | No flags | Advanced |
| torchdrug | 3 | 11mo | Review | Advanced |
| string-database | 2 | 11mo | Review | Intermediate |
| transformer-lens-interpretability | 2 | 10mo | Review | Advanced |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by mims-harvard
View all by mims-harvard →You might also like
torchdrug
davila7
Graph-based drug discovery toolkit. Molecular property prediction (ADMET), protein modeling, knowledge graph reasoning, molecular generation, retrosynthesis, GNNs (GIN, GAT, SchNet), 40+ datasets, for PyTorch-based ML on molecules, proteins, and biomedical graphs.
string-database
davila7
Query STRING API for protein-protein interactions (59M proteins, 20B interactions). Network analysis, GO/KEGG enrichment, interaction discovery, 5000+ species, for systems biology.
transformer-lens-interpretability
davila7
Provides guidance for mechanistic interpretability research using TransformerLens to inspect and manipulate transformer internals via HookPoints and activation caching. Use when reverse-engineering model algorithms, studying attention patterns, or performing activation patching experiments.
denario
davila7
Multiagent AI system for scientific research assistance that automates research workflows from data analysis to publication. This skill should be used when generating research ideas from datasets, developing research methodologies, executing computational experiments, performing literature searches, or generating publication-ready papers in LaTeX format. Supports end-to-end research pipelines with customizable agent orchestration.
tooluniverse
jimmc414
Use this skill when working with scientific research tools and workflows across bioinformatics, cheminformatics, genomics, structural biology, proteomics, and drug discovery. This skill provides access to 600+ scientific tools including machine learning models, datasets, APIs, and analysis packages. Use when searching for scientific tools, executing computational biology workflows, composing multi-step research pipelines, accessing databases like OpenTargets/PubChem/UniProt/PDB/ChEMBL, performing tool discovery for research tasks, or integrating scientific computational resources into LLM workflows.
tooluniverse-binder-discovery
mims-harvard
Discover novel small molecule binders for protein targets using structure-based and ligand-based approaches. Creates actionable reports with candidate compounds, ADMET profiles, and synthesis feasibility. Use when users ask to find small molecules for a target, identify novel binders, perform virtual screening, or need hit-to-lead compound identification.