tooluniverse-structural-proteomics — скилл для Claude Code
Скилл tooluniverse-structural-proteomics для Claude Code собирает структурные и протеомные данные для проверки лекарственной мишени. Помогает оценить участки связывания, лиганды и уверенность структурных моделей.
Apache-2.0 240 установок ★ 1 798 обновлён 2026-10-24
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a claude-codeЧто делает tooluniverse-structural-proteomics
Скилл ищет данные о белке и его структурах через инструменты ToolUniverse: от экспериментальных структур PDB и предсказаний AlphaFold до лигандов BindingDB, антительных структур SAbDab и NMR-записей BMRB. Он также помогает изучать домены, мембранную топологию, каталитические центры и внутренне неупорядоченные участки.
Для анализа с расчётами скилл предписывает запускать Python через Bash и сообщать полученные результаты. К примеру, разрешение структуры влияет на допустимые выводы: при значении выше 3 Å надёжной считается только основа белка, а участок связывания — ненадёжным. MDDB возвращает метаданные опубликованных симуляций, но не файлы траекторий; MobiDB сейчас недоступен.
Какие задачи решает
- найти структуры белка и сравнить их уверенность
- оценить участок связывания и доступные данные о лигандах
- проверить антительные интерфейсы и мутации
- изучить топологию мембранного белка или каталитический центр
Примеры запросов агенту
- Найди структуры белка по UniProt и сравни их разрешение с прогнозом AlphaFold.
- Оцени возможные участки связывания и найди для этой мишени лиганды с данными BindingDB.
- Поищи опубликованные MD-симуляции для белка и перечисли параметры траекторий.
Как установить tooluniverse-structural-proteomics
1Через Skills CLI (Vercel)
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a claude-code2Вручную: положить файлы
~/.claude/skills/tooluniverse-structural-proteomics/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/tooluniverse-structural-proteomics/
1Через Skills CLI (Vercel)
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a codex2Вручную: положить файлы
.agents/skills/tooluniverse-structural-proteomics/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a cursor2Вручную: положить файлы
.cursor/skills/tooluniverse-structural-proteomics/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install mims-harvard/ToolUniverse skills/tooluniverse-structural-proteomics2Вручную: положить файлы
.github/skills/tooluniverse-structural-proteomics/SKILL.md1Через Skills CLI (Vercel)
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a gemini-cli2Командой платформы
gemini skills install https://github.com/mims-harvard/ToolUniverse --path skills/tooluniverse-structural-proteomicsВручную: положить файлы
.devin/skills/tooluniverse-structural-proteomics/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о tooluniverse-structural-proteomics
Как установить tooluniverse-structural-proteomics в Claude Code?
Выполни npx skills add mims-harvard/ToolUniverse --skill tooluniverse-structural-proteomics -a claude-code. Файл скилла находится по пути ~/.claude/skills/tooluniverse-structural-proteomics/SKILL.md.
Что делает tooluniverse-structural-proteomics?
Скилл объединяет поиск структурных и протеомных данных о мишени: структур PDB и AlphaFold, лигандов, доменов, участков связывания и других характеристик белка.
Получает ли скилл файлы MD-симуляций?
Нет. MDDB даёт метаданные опубликованных траекторий, например температуру и длительность симуляции, но не сами файлы траекторий.
Исходный текст
SKILL.md · Apache-2.0
---
name: tooluniverse-structural-proteomics
description: Structural biology plus proteomics integration for drug target validation. Combines PDB experimental structures, AlphaFold predictions, GPCRdb, SAbDab antibody structures, ProteinsPlus binding-site prediction, BindingDB ligand-affinity data, BMRB NMR data, CATH/InterPro fold classification, intrinsic-disorder prediction (MobiDB/DisProt/IUPred3), membrane-protein topology (OPM/TopDB/PDBTM/ChannelsDB), enzyme catalytic-site data (M-CSA), PDBe ligand/compound chemistry, and MD simulation trajectory metadata (MDDB). Use for druggability assessment, binding-site characterization, ligand-pocket analysis, structural-confidence scoring (resolution, pLDDT), antibody-target interface analysis, disordered-region assessment, membrane-protein topology, catalytic-mechanism lookup, and finding published MD simulations for a target.
disable-model-invocation: true
---
# Structural Proteomics for Drug Target Validation
Comprehensive structural data integration using ToolUniverse tools across PDB, AlphaFold, GPCRdb, SAbDab, and proteomics databases for drug target validation.
## LOOK UP DON'T GUESS
- PDB structures/resolutions: `PDBeSIFTS_get_best_structures` and `RCSBGraphQL_get_structure_summary`
- AlphaFold confidence: `alphafold_get_summary`
- Ligands/affinities: `PDBe_get_structure_ligands` and `BindingDB_get_ligands_by_uniprot`
- Protein-protein binding mutations (ddG): `SKEMPI_search_by_structure`/`SKEMPI_search_by_protein`/`SKEMPI_get_mutation` -- BindingDB's protein-protein equivalent, for interface-mutation affinity effects rather than small-molecule ligands
- Druggability: `ProteinsPlus_predict_binding_sites`
## COMPUTE, DON'T DESCRIBE
When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.
## Domain Reasoning
Resolution determines valid conclusions: <2A = atom positions visible; 2-3A = side chains reliable, drug design supported; >3A = backbone only, binding site unreliable. Do not over-interpret low-resolution structures.
---
## Tool Inventory
### PDB (RCSB)
`RCSBAdvSearch_search_structures` (query_type, query_value, rows), `RCSBData_get_entry` (entry_id), `RCSBGraphQL_get_structure_summary` (pdb_id), `RCSBGraphQL_get_ligand_info` (pdb_id), `RCSB_get_chemical_component` (comp_id)
### PDB (PDBe)
`pdbe_get_entry_summary` (pdb_id), `PDBe_get_structure_ligands` (pdb_id), `PDBe_get_bound_molecules` (pdb_id), `PDBeSearch_search_structures` (query, rows), `PDBeSIFTS_get_best_structures` (uniprot_id), `PDBeSIFTS_get_all_structures` (uniprot_id), `PDBe_KB_get_ligand_sites` (pdb_id), `PDBe_KB_get_interface_residues` (pdb_id), `PDBeValidation_get_quality_scores` (pdb_id)
### PDBe PISA
`PDBePISA_get_interfaces` (pdb_id), `PDBePISA_get_assemblies` (pdb_id)
### AlphaFold
`alphafold_get_prediction` (qualifier=UniProt), `alphafold_get_summary` (qualifier), `alphafold_get_annotations` (qualifier)
### Binding Sites
`ProteinsPlus_predict_binding_sites` (pdb_id, chain), `BindingDB_get_ligands_by_uniprot` (uniprot_id), `BindingDB_get_ligands_by_pdb` (pdb_id), `BindingDB_get_targets_by_compound` (smiles)
### Ligand/Compound Chemistry (PDBe Compounds)
`PDBe_get_compound_summary` (comp_id — 3-letter PDB chemical-component code) — formula, weight, InChI/InChIKey, SMILES, systematic names, and cross-references to PubChem/DrugBank/EPA CompTox for one ligand. `PDBe_get_compound_structures` (comp_id) — **see gotcha below, does not return an exhaustive structure list**.
### Foldseek
`Foldseek_search_structure` (sequence, mode="tmalign"), `Foldseek_get_result` (ticket)
### GPCRdb
`GPCRdb_get_protein` (protein), `GPCRdb_get_structures` (protein), `GPCRdb_get_ligands` (protein), `GPCRdb_get_mutations` (protein). Accepts entry names, gene symbols (auto-converted to `{symbol.lower()}_human`), or UniProt accessions.
### SAbDab
`SAbDab_search_structures` (query/antigen), `SAbDab_get_structure` (pdb_id), `TheraSAbDab_search_therapeutics` (query), `TheraSAbDab_search_by_target` (target)
### Domains
`InterPro_get_protein_domains` (uniprot_id), `Pfam_get_protein_annotations` (uniprot_id), `UniProt_get_entry_by_accession` (accession)
### Fold/Domain Classification (CATH + InterPro member databases)
`CATH_get_superfamily` (superfamily_id), `CATH_get_domain_summary` (domain_id), `CATH_list_funfams` (superfamily_id), `CATH_get_funfam` (superfamily_id, funfam_number), `InterPro_list_member_databases` (no params), `InterPro_list_member_signatures` (member_database, page_size), `InterPro_get_member_signature` (member_database, accession), `InterPro_get_structures_for_entry` (interpro_id, page_size)
### Domain Architecture and Clans (InterPro, Pfam-accession-keyed)
`InterPro_get_protein_domain_architecture` (accession), `InterPro_get_structures_for_domain` (pfam_accession, max_results), `InterPro_get_clan_members` (clan_accession, max_results) — **InterPro's API is intermittently flaky, see Workflow 5b below before assuming a "not found" result means the entry doesn't exist**
### Proteomics
`ProteomeXchange_search_datasets` (query), `ProteomeXchange_get_dataset` (dataset_id)
### Molecular Dynamics Simulation Metadata (MDDB)
`MDDB_search_projects` (query, limit), `MDDB_get_project` (accession) — simulation parameters (force field, temperature, ensemble, simulated time, frames, trajectory size) for published MD trajectories; metadata only, not the trajectory files
### BMRB (NMR data)
`BMRB_search_by_keyword` (term, database="macromolecules"|"metabolomics"), `BMRB_search_by_sequence` (sequence), `BMRB_get_entries_by_pdb_id` (pdb_id), `BMRB_get_entries_by_uniprot` (uniprot_id), `BMRB_get_entry` (entry_id), `BMRB_get_entry_citation` (entry_id), `BMRB_get_validation` (entry_id), `BMRB_search_chemical_shifts` (entry_id or search filters)
### Intrinsic Disorder (MobiDB, DisProt, IUPred3)
`MobiDB_get_protein` (accession), `MobiDB_get_consensus` (accession) — **currently unreachable, see Limitations**. `DisProt_search` (query, page_size), `DisProt_get_entry` (accession — DisProt ID or UniProt accession). `IUPred3_predict_disorder` (accession, iupred_type="long"|"short"|"anchor")
### Membrane Protein Structure (OPM, TopDB, PDBTM, ChannelsDB)
`OPM_search_structures` (query, limit), `TopDB_get_topology` (identifier), `PDBTM_get_topology` (pdb_id), `ChannelsDB_get_channels_pdb` (pdb_id), `ChannelsDB_get_channels_cofactor` (pdb_id)
### Catalytic Site Atlas (M-CSA)
`MCSA_get_entry` (mcsa_id), `MCSA_search_enzymes` (enzyme_name, ec_number, uniprot_id, max_pages, limit)
---
## Workflow 1: Find All Structures for a Drug Target
```
Phase 0: Resolve protein → UniProt ID, gene symbol, organism
Phase 1: PDBeSIFTS_get_best_structures → RCSBGraphQL_get_structure_summary → PDBeValidation
Phase 2: alphafold_get_prediction/summary → compare pLDDT with experimental coverage
Phase 3: IF GPCR → GPCRdb; IF antibody target → SAbDab/TheraSAbDab
Phase 4: InterPro/Pfam domain mapping → identify unresolved regions
Phase 5: Summary table (PDB ID, method, resolution, ligands, coverage, quality)
```
**Decisions**: Resolution <2.5A for drug design. X-ray > Cryo-EM > NMR > AlphaFold for binding sites. Holo > apo structures. When a target's best experimental structure IS an NMR structure (common for small/disordered proteins where crystallography fails), or when dynamics/flexibility data matters more than a single static pocket geometry, pull the underlying chemical-shift/assignment data via BMRB (Workflow 4) rather than treating the PDB coordinate file as the only NMR evidence available.
## Workflow 4: NMR Structural/Dynamics Data (BMRB)
```
Phase 1: Resolve target -> BMRB_search_by_keyword (protein name) or BMRB_search_by_sequence (exact sequence match)
or, if you already have an NMR PDB entry -> BMRB_get_entries_by_pdb_id (cross-reference to the matching BMRB entry/entries)
or BMRB_get_entries_by_uniprot (UniProt accession -> linked BMRB entries)
Phase 2: BMRB_get_entry(entry_id) -> full deposition: saveframes for entry_information, assigned chemical shifts,
sample conditions, method. BMRB_get_entry_citation(entry_id) -> BibTeX citation for the deposition.
Phase 3: BMRB_get_validation(entry_id) -> AVS (assigned-value-set) analysis: per-residue/per-atom chemical-shift
typing and assignment-quality scores -- use this before trusting a shift value, the same way
PDBeValidation quality scores gate trust in an X-ray/cryo-EM structure.
Phase 4 (optional, slow): BMRB_search_chemical_shifts -> raw per-atom chemical-shift table for an entry
(large response, 15-20s+ for popular entries -- only pull this when the actual shift values are
needed, e.g. for a downstream chemical-shift-based dynamics or secondary-structure calculation).
```
**Real example** (verified live): PDB `1D3Z` (an NMR ubiquitin structure) -> `BMRB_get_entries_by_pdb_id` returns BMRB entries `11505`, `11547`, `15047`, ... (BLAST-matched by sequence, not a strict 1:1 PDB<->BMRB mapping -- expect several candidate entries and pick by title/method) -> `BMRB_get_entry("11505")` confirms title "Alternative structure of Ubiquitin" -> `BMRB_get_validation("11505")` returns the AVS chemical-shift-typing analysis for that deposition.
**Tool note**: `BMRB_search_by_keyword`'s `term` parameter accepts a plain protein/molecule name (e.g. `"ubiquitin"`, `"calmodulin"`) -- not `keyword`, despite the tool name.
## Workflow 5: Fold and Domain-Signature Classification (CATH + InterPro Member Databases)
CATH classifies a protein *domain*'s 3D fold (Class/Architecture/Topology/Homologous-superfamily) and further groups domains into functionally-coherent FunFams within a superfamily. This is complementary to, not a replacement for, `InterPro_get_protein_domains`/`Pfam_get_protein_annotations` above: those two answer "what domains does *this protein* have," while CATH and the InterPro member-database tools here let you go the other direction — starting from a fold/superfamily/signature and finding everything that belongs to it (structures, domains, or the raw per-database signature catalog behind InterPro's integrated entries).
```
Phase 1: CATH_get_domain_summary(domain_id) -> C.A.T.H classification + residue count for one PDB-chain-domain
(domain_id format: <pdb_id><chain><domain_number>, e.g. "1cukA01")
Phase 2: CATH_get_superfamily(superfamily_id) -> homologous-superfamily name + domain/family counts across all of CATH
Phase 3: CATH_list_funfams(superfamily_id) -> functional sub-families within that superfamily, ranked by member count
Phase 4: CATH_get_funfam(superfamily_id, funfam_number) -> detail for one specific FunFam
```
**Real example** (verified live): `CATH_get_domain_summary(domain_id="1cukA01")` -> `cath_id: "2.40.50.140.116.1.1.1.1"`, superfamily `2.40.50.140` ("Nucleic acid-binding proteins", Mainly Beta) -> `CATH_get_superfamily("2.40.50.140")` -> 2,879 total domains in this superfamily -> `CATH_list_funfams("2.40.50.140")` -> 1,285 FunFams, top one "30S ribosomal protein S12" (6,057 members).
**Tool note**: `CATH_get_superfamily` takes `superfamily_id`, not `cath_id` — passing `cath_id` fails parameter validation even though the tool's own response echoes a `cath_id` field.
```
Phase 1: InterPro_list_member_databases() -> catalog of all 14 member databases (pfam, cathgene3d, ssf, panther,
cdd, profile, smart, ncbifam, prosite, prints, hamap, pirsf, sfld, antifam) with entry counts
Phase 2: InterPro_list_member_signatures(member_database, page_size) -> browse that database's raw signature catalog,
each entry showing its `integrated_interpro` cross-reference if InterPro has folded it into an integrated entry
Phase 3: InterPro_get_member_signature(member_database, accession) -> full detail for one signature (name, GO terms,
protein/structure/taxa counts)
Phase 4: InterPro_get_structures_for_entry(interpro_id, page_size) -> real PDB structures (with experiment type and
resolution) whose chains carry a given *integrated* InterPro entry
```
**Real example** (verified live): `InterPro_get_member_signature(member_database="pfam", accession="PF00069")` -> "Protein kinase domain", `integrated_interpro: "IPR000719"`, 1,364,822 proteins / 5,477 structures / 36,542 taxa carry this signature -> `InterPro_get_structures_for_entry(interpro_id="IPR000719")` -> real PDB hits including kinase-inhibitor co-crystals (e.g. `10dj`, Fyn kinase + saracatinib, 2.22Å) — a direct bridge from "this protein has a kinase domain" to "here are solved structures of that domain, several already in complex with a drug."
**Tool note**: `InterPro_get_member_signature` requires BOTH `member_database` and `accession` — the accession alone (e.g. `PF00069`) is not sufficient because the same-shaped accession could theoretically collide across member databases.
## Workflow 5b: Whole-Protein Domain Architecture and Pfam Clans
Three more InterPro tools, keyed by Pfam accession/UniProt accession rather than the integrated InterPro ID used above — a different entry point for the same underlying database:
```
InterPro_get_protein_domain_architecture(accession) -> ALL Pfam domains on one protein at once, with
exact residue start/end and the integrated InterPro ID for each -- the "what domains does this
whole protein have, with coordinates" complement to InterPro_get_member_signature's "tell me about
one specific signature" above
InterPro_get_structures_for_domain(pfam_accession, max_results) -> PDB structures containing a given
Pfam domain -- the Pfam-accession-keyed sibling of InterPro_get_structures_for_entry(interpro_id)
above; use whichever accession type (Pfam PFxxxxx vs integrated IPRxxxxxx) you already have,
they answer the same question from the two different InterPro identifier namespaces
InterPro_get_clan_members(clan_accession, max_results) -> member Pfam families in a clan (Pfam's own
superfamily grouping) -- distinct from CATH's superfamily/FunFam grouping in Workflow 5 above:
CATH groups by solved 3D fold, Pfam clans group by sequence-profile relatedness; the two systems
can (and often do) disagree on where the boundary of a "related family" sits
```
**Real example** (verified live, same protein as Workflow 1): `InterPro_get_protein_domain_architecture(accession="P04637")` (TP53) -> 4 real Pfam domains with exact coordinates: PF08563 "P53 transactivation motif" (6-30), PF18521 "Transactivation domain 2" (35-59), PF00870 "P53 DNA-binding domain" (100-288), PF07710 "P53 tetramerisation motif" (319-357) -- matches the tetramerisation/DNA-binding boundaries already used in this skill's other workflows. `InterPro_get_clan_members(clan_accession="CL0016")` -> real clan "PKinase" with 30+ member Pfam families.
**Reliability warning (verified live, repeatedly): this InterPro endpoint family is intermittently flaky.** The exact same call (`InterPro_get_structures_for_domain(pfam_accession="PF00069")`) returned "Not found in InterPro: PF00069" on two consecutive attempts, then succeeded on a third with no change to the request. `InterPro_get_protein_domain_architecture` showed the same pattern in reverse (failed via `tu test`'s canned examples, succeeded via a fresh `tu run` moments later, on the identical accession). **Treat a "Not found in InterPro" error from any tool in this cluster as inconclusive, not as proof the accession doesn't exist — retry 1-2 times before reporting an entry is missing.**
## Workflow 6: Intrinsic Disorder Assessment (MobiDB + DisProt + IUPred3)
These three tools answer the same underlying question — "is this region of the protein structured or disordered?" — with different evidence quality, from highest to lowest confidence:
```
Tier 1 (curated experimental evidence): DisProt_search(query) -> DisProt_get_entry(accession)
-> real, literature-curated disordered-region boundaries with the experimental method that
established them (NMR, circular dichroism, limited proteolysis, etc.)
Tier 2 (aggregated predictions + curation): MobiDB_get_protein(accession) / MobiDB_get_consensus(accession)
-> merges DisProt curation with multiple computational predictors into one consensus view
**currently unreachable, see Limitations below**
Tier 3 (single fast predictor, sequence-only): IUPred3_predict_disorder(accession, iupred_type=...)
-> a real-time computed disorder score per residue; "long" = long disordered regions,
"short" = short disordered segments, "anchor" = disorder that becomes ordered upon binding
a partner (protein-binding-induced folding)
```
**Real example** (verified live): p53/P04637 — `DisProt_get_entry("P04637")` confirms curated disordered regions (DP00086, "Cellular tumor antigen p53"); `IUPred3_predict_disorder(accession="P04637", iupred_type="long")` returns a real per-residue disorder-score profile independently, useful for a quick check even without DisProt curation existing for a given protein.
**When to use which**: check DisProt first if you need citable, experimentally-validated boundaries. Use IUPred3 for any protein (curated or not) when you just need a fast disorder profile. Reach for MobiDB only once it is confirmed reachable again (see Limitations) — do not report a "no disorder data" conclusion based on a MobiDB timeout; that is a connectivity failure, not evidence of an ordered protein.
## Workflow 7: Membrane Protein Structure (OPM + TopDB + PDBTM + ChannelsDB)
```
Phase 1: OPM_search_structures(query, limit) -> solved membrane-protein structures positioned in a lipid
bilayer, with geometric/energetic properties: hydrophobic thickness, tilt angle, and
transfer_energy_kcal_per_mol (more negative = more favorable membrane insertion)
Phase 2: TopDB_get_topology(identifier) -> curated per-segment topology (Inside/Membrane/Outside regions)
for a named protein, cross-species; identifier is typically a TopDB/UniProt-style ID
(e.g. "OPSD_HUMAN"), not a bare gene symbol
Phase 3: PDBTM_get_topology(pdb_id) -> structure-derived per-chain topology for one specific PDB entry:
tm_type ("alpha"/"beta"/"non_tm"), num_tm_segments, is_membrane_embedded
Phase 4: ChannelsDB_get_channels_pdb(pdb_id) / ChannelsDB_get_channels_cofactor(pdb_id) -> detected
channels/tunnels/pores through the structure (general access tunnels vs. cofactor-specific
access tunnels) for a given PDB entry
```
**Real example** (verified live, cross-validated across all four tools on rhodopsin): `OPM_search_structures(query="rhodopsin")` finds real bacteriorhodopsin/archaerhodopsin family members with thickness/tilt data; `TopDB_get_topology("OPSD_HUMAN")` -> 7 transmembrane segments, reliability score 92.81; `PDBTM_get_topology("1f88")` -> chain A, `tm_type: "alpha"`, `num_tm_segments: 7` — the same helix count TopDB reported independently, a useful cross-check between curated (TopDB) and structure-derived (PDBTM) topology. `ChannelsDB_get_channels_pdb` on a real ion channel (`1bl8`, KcsA potassium channel) returns real annotated pore/tunnel data; **not every PDB entry has ChannelsDB coverage** — `1f88` (rhodopsin) returned "Protein with ID '1f88' not found in ChannelsDB" even though OPM/TopDB/PDBTM all have data for it, so a ChannelsDB miss does not mean the protein lacks a real pore.
## Workflow 8: Enzyme Catalytic Mechanism (M-CSA)
M-CSA (Mechanism and Catalytic Site Atlas) curates the actual catalytic residues and reaction mechanism for enzymes with solved structures — distinct from BRENDA/kinetics data (rates/constants) and from generic active-site predictions (ProteinsPlus above, which is pocket-geometry-based, not mechanism-curated).
```
Phase 1: MCSA_search_enzymes(enzyme_name=... | ec_number=... | uniprot_id=...) -> candidate M-CSA entries
Phase 2: MCSA_get_entry(mcsa_id) -> full catalytic machinery: catalytic residues, mechanism steps,
EC number(s), reference structure
```
**Real example** (verified live): `MCSA_search_enzymes(enzyme_name="lysozyme")` -> `mcsa_id: 203`, "lysozyme (glycosyl hydrolase 22 family)", EC 3.2.1.17, reference UniProt P00698 -> `MCSA_get_entry(203)` for the full mechanism detail.
---
## Workflow 2: Identify Binding Pocket Ligands
```
Phase 1: PDBe_get_structure_ligands + RCSBGraphQL_get_ligand_info + PDBe_KB_get_ligand_sites
Phase 2: ProteinsPlus_predict_binding_sites → druggability score, pocket residues
Phase 3: BindingDB_get_ligands_by_pdb/uniprot → Ki, Kd, IC50
Phase 4: RCSB_get_chemical_component for key ligands, or PDBe_get_compound_summary for
the same 3-letter comp_id if you need InChI/InChIKey/SMILES/systematic names and
cross-references (PubChem/DrugBank/EPA CompTox) rather than RCSB's structural fields
```
**Filter artifacts**: GOL, EDO, SO4, PEG, ACT, CL, NA. Keep cofactors (ATP, NAD, HEM) and catalytic metals (ZN, MG) if relevant.
**Real example** (verified live): `PDBe_get_compound_summary({"comp_id":"ATP"})` -> formula C10H16N5O13P3, weight 507.181, InChIKey `ZKHQWZAMYRWXGA-KQYNXXCUSA-N`, cross-references including BRENDA and Probes And Drugs. `PDBe_get_compound_summary({"comp_id":"HEM"})` -> `formal_charge: null` (a real schema mismatch, verified live: the tool's `return_schema` declares `formal_charge` as a non-nullable integer, but HEM's actual charge is reported as null — the call still succeeds and the rest of the data is correct, this is a documentation/schema gap, not a broken tool).
**Gotcha (verified live, do not present as exhaustive): `PDBe_get_compound_structures` does not reliably return "all structures containing this compound" despite its description.** Tested on ATP (which appears in thousands of real PDB structures — kinases, ATPases) and HEM (hemoglobins, cytochromes, peroxidases): both calls returned exactly ONE PDB entry each (`1b0u` for ATP, `3ia3` for HEM). Tracing the tool's implementation confirms why: when its primary per-compound structure list is empty, it silently falls back to the compound's `first_observed_in` field (which is inherently a single structure) and still labels that result `pdb_entries`. Treat a `pdb_entries` result from this tool as "one example structure containing this compound," not as a complete list — for an exhaustive structure search by ligand, use `RCSBGraphQL_get_ligand_info` or `RCSBAdvSearch_search_structures` instead.
## Workflow 9: MD Simulation Trajectory Metadata (MDDB)
MDDB indexes 15,000+ molecular dynamics simulation trajectories (MoDEL, BioExcel, DESRES-ANTON, and community depositions) — metadata only (force field, temperature, ensemble, simulated time, frame count, trajectory size), not the trajectory files themselves. Use when a target already has (or you want to check whether it has) a published MD simulation, as a complement to the static structures the rest of this skill documents.
```
Phase 1: MDDB_search_projects(query, limit) -> candidate project accessions, by protein name or PDB id
Phase 2: MDDB_get_project(accession) -> full simulation parameters for one project
```
**Real example** (verified live): `MDDB_search_projects({"query":"6M71","limit":5})` -> project `MD-A001UA` -> `MDDB_get_project({"accession":"MD-A001UA"})` -> a real 10 microsecond DESRES-ANTON simulation of the SARS-CoV-2 nsp7-nsp8-nsp12 RNA polymerase complex (Amber ff99SB-ILDN, NPT ensemble, 310 K, 8,335 frames, ~2.5 GB) with a text description of a specific structural finding from the simulation (a disordered N-terminal region folding into a kinase-like fold during the run). Note `total_simulated_time_ps` is summed across all replicas listed under `md_replica_count`, not per-replica.
## Workflow 3: Cross-Validate Drug Binding
```
Phase 1: Find co-crystal structures → filter for drug/analogs
Phase 2: BindingDB affinity data (Ki, Kd, IC50)
Phase 3: ProteinsPlus + PDBe-KB binding site characterization
Phase 4: PDBeValidation quality → binding site well-resolved?
Phase 5: AlphaFold + Foldseek structural comparison
Phase 6: GPCR-specific (if applicable) → active/inactive states, pharmacology, resistance mutations
Phase 7: Antibody-specific (if applicable) → epitope mapping
Phase 8: Evidence integration
```
---
## Tool Parameter Gotchas
| Tool | Mistake | Correct |
|------|---------|---------|
| `alphafold_get_prediction/summary` | `uniprot_id` | `qualifier` |
| `GPCRdb_get_protein` | `gene_name` | `protein` |
| `PDBeSIFTS_get_best_structures` | gene symbol | `uniprot_id` (e.g., "P04637") |
| `Foldseek_search_structure` | `mode="3diaa"` | `mode="tmalign"` |
| `SAbDab_search_structures` | `name` | `query` or `antigen` |
| `RCSB_get_chemical_component` | `ligand_id` | `comp_id` |
| `BMRB_search_by_keyword` | `keyword` | `term` |
| `CATH_get_superfamily` | `cath_id` | `superfamily_id` |
| `InterPro_get_member_signature` | `accession` alone | `member_database` + `accession` (both required) |
| `TopDB_get_topology` | bare gene symbol | a TopDB/UniProt-style identifier (e.g. `"OPSD_HUMAN"`) |
| `PDBe_get_compound_structures` | trusting `pdb_entries` as exhaustive | it falls back to a single `first_observed_in` example when the primary list is empty (verified live) — use `RCSBGraphQL_get_ligand_info` for a real exhaustive search |
---
## Evidence Grading
| Tier | Confidence |
|------|------------|
| T1 | Co-crystal (<2.5A) + binding affinity data |
| T2 | Experimental structure + computational prediction |
| T3 | AlphaFold + pocket analysis + known ligand analogs |
| T4 | Homology model or low-resolution only |
## Interpretation
| Metric | High | Acceptable | Caution |
|--------|------|-----------|---------|
| Resolution | <2.0A (X-ray) / <3.0A (cryo-EM) | 2.0-2.5A / 3.0-4.0A | >3.0A / >4.5A |
| R-free | <0.25 | 0.25-0.30 | >0.30 |
| AlphaFold pLDDT | >90 | 70-90 | <70 (disordered) |
DoGSiteScorer >0.6 = druggable; <0.4 = unlikely druggable. PISA assemblies should be cross-validated with SEC-MALS/native MS.
## Limitations
- BindingDB: 60s+ for popular targets
- AlphaFold: lacks ligand context
- GPCRdb: Class A-F GPCRs only
- PDBePISA: `operation` is internal, not a public parameter
- BMRB: `BMRB_get_entries_by_pdb_id` is a BLAST-based sequence match, not a curated 1:1 PDB<->BMRB cross-reference -- a query can return several candidate entries (or none) for a real NMR PDB ID; `BMRB_search_chemical_shifts` is slow (15-20s+) for well-studied entries, only call it when raw shift values are actually needed
- MobiDB: `MobiDB_get_protein` and `MobiDB_get_consensus` were verified live (repeatedly, with a 90s timeout) to hang indefinitely against their configured endpoint (`https://mobidb.org/api/download`) -- confirmed independently via direct `curl` to the same host (20s with no response). This is an upstream connectivity issue, not a query-parameter mistake. Use DisProt (curated) or IUPred3 (fast predictor) instead until this is confirmed working again -- do not retry MobiDB with a longer timeout expecting it to resolve.
- ChannelsDB: coverage is per-PDB-entry, not universal -- a real membrane protein with confirmed topology elsewhere (OPM/TopDB/PDBTM) can still return "not found in ChannelsDB" for that specific PDB ID (verified live: `1f88` rhodopsin)
- InterPro domain-architecture cluster (`InterPro_get_protein_domain_architecture`/`_get_structures_for_domain`/`_get_clan_members`): the underlying InterPro API is intermittently flaky -- the identical query can 404 twice then succeed on a third attempt with no request change (verified live, repeatedly). A "Not found in InterPro" result is inconclusive on the first attempt; retry before concluding the accession/domain genuinely isn't in InterPro.
- `PDBe_get_compound_summary`: `formal_charge` is `null` for at least some neutral compounds (verified live: HEM) despite the tool's `return_schema` declaring it a non-nullable integer -- the call still succeeds with correct data otherwise, treat a null charge as "not reported as charged," not as a failure.
- `PDBe_get_compound_structures`: see the Workflow 2 gotcha above -- verified live to return only one example PDB entry even for ubiquitous ligands (ATP, HEM), not an exhaustive list.
© mims-harvard · Apache-2.0 · mims-harvard/ToolUniverse