Science Skills

The Reactome Analysis Service provides endpoints for biological pathway analysis, identifier mapping, and species comparison. The base URL for all requests is https://reactome.org/AnalysisService.

API Categories

• Database: Retrieve database name and version.
• Identifier: Analyze single identifiers across species or with projection to Homo sapiens.
• Identifiers - Batch Analysis: Perform overrepresentation or expression analysis on multiple identifiers via POST (body), file upload, or URL.
• Token - Result Retrieval: Use a token to retrieve, filter, or summarize analysis results.
• Download: Export results in JSON, gzipped JSON, or CSV formats.
• Mapping: Map identifiers across different species.
• Import: Re-import previously exported JSON data.
• Report: Download PDF reports.
• Species Comparison: Compare Homo sapiens to other species.

The foldseek-structural-search skill allows you to perform 3D structural searches of proteins against various databases (such as PDB, AlphaFold, CATH, and MGnify) using the Foldseek web server API.

Critical Requirement: This tool only works with physical 3D coordinate files. You must provide a path to a .cif, .mmcif, or .pdb file. It cannot search using protein sequences, gene names, or UniProt IDs. If you only have a sequence, you must first download the structure (e.g., using an AlphaFold fetch tool).

Supported Databases: When specifying databases, you must choose from the following allowed list:

• afdb50
• afdb-swissprot
• pdb100
• BFVD
• mgnify_esm30
• cath50
• gmgcl_id
• bfmd
• afdb-proteome

When searching for specific product names, drug names, or categorical terms, always append .exact to the field name. Without this, the API tokenizes multi-word values, leading to noisy partial matches.

Example: Precise Brand Name Search

uv run scripts/openfda_query.py search --category drug --endpoint label \
  --search 'openfda.brand_name.exact:"ADVIL"' \
  --limit 5 --output /tmp/advil_label.json

Note on Variants: Many brand names include suffixes (e.g., "TYLENOL Extra Strength"). If an .exact search returns 0 results, try searching without .exact to identify the available variants, then re-query using the full exact name found.

Aggregation: The .exact suffix is also required when using --count_field to ensure you are aggregating whole phrases rather than individual words.

When using the ENCODE SCREEN GraphQL API (https://factorbook.api.wenglab.org/graphql), keep the following in mind:

• Pagination/Limits: If a query is too large, it may return an error. To avoid this, split your requests by dividing coordinates or accessions into smaller chunks.
• Composability: While multiple queries can be batched in a single request, it is often more efficient to use the provided Python script abstractions unless a highly specific custom GraphQL query is required.

When attempting to find a protein sequence, follow this priority order for maximum efficiency:

1. Direct accession: fetch-protein (GenPept/RefSeq)
2. CDS translation: cds-translate (nucleotide/CDS accession)
3. PubMed-linked: pubmed-proteins (PMID + gene name)
4. Locus lookup: locus-protein (locus tag + organism)
5. Gene + organism: gene-protein (gene name + organism)
6. Patent search: patent-search (patent number or keywords)
7. Organism + length: organism-length (last resort)

Multisequence Conservation

• Conservation Index: (Fully Conserved Columns) / (Total Alignment Columns). Use this to quantify the percentage of residues that are 100% identical across the entire alignment set. This identifies the core evolutionary signature of the protein family.

All list endpoints support pagination using --limit and --offset.

# First page: 2 results starting at offset 0
uv run scripts/chembl_api.py molecule --limit 2 --offset 0 --output /tmp/page1.json

# Second page: next 2 results starting at offset 2
uv run scripts/chembl_api.py molecule --limit 2 --offset 2 --output /tmp/page2.json

The response includes a page_meta object containing total_count, limit, offset, next, and previous links. Use successive --offset values to page through large result sets.

This skill is designed to retrieve and analyze AlphaFold predicted structures (mmCIF) and Predicted Aligned Error (PAE) matrices for a specific UniProt Accession ID.

Use this skill when:

• You have a specific UniProt Accession ID.
• You need structural confidence metrics (pLDDT).
• You need domain boundary analysis or disorder assessment.

Do NOT use this skill when:

• The user only provides a protein name, gene name, or amino acid sequence (ask for a UniProt ID first).
• The user wants to search for structural homologs (use Foldseek instead).
• The user wants to run AlphaFold predictions on a custom sequence.
• The user needs experimental PDB structures (use RCSB PDB instead).

Use the ClinVar skill when you need to access the 'clinical ground truth' for human genomic variations, specifically for:

• Finding current clinical significance (e.g., Pathogenic, Benign, VUS) and star ratings (review status).
• Fetching clinician notes, assertion criteria, or rationales.
• Retrieving preferred condition names and associated HPO terms.
• Finding variant controls (e.g., searching for all Pathogenic variants in a specific gene).
• Identifying conflicting interpretations and the organizations submitting them.

Do NOT use ClinVar for:

• Global population allele frequencies (use gnomAD).
• Protein biological roles or inheritance patterns (use OMIM).
• Predicting mechanistic effects of novel mutations (use AlphaGenome).
• Patient surveillance schedules (use GeneReviews).
• 3D structural models (use PDB / AlphaFold).