open-biosciences/biosciences-competency-questions-sample

Open Biosciences Competency Questions (Sample)

Dataset Description

A curated collection of 15 competency questions (CQs) for evaluating and guiding knowledge graph construction in biosciences research. Each question includes structured entities with standardized CURIEs, gold-standard knowledge graph paths using BioLink predicates, executable multi-API workflow steps, and source provenance.

What Are Competency Questions?

Competency questions are natural language questions that a knowledge graph or ontology must be able to answer. Coined by Gruninger & Fox (1995), they serve as requirements specifications for knowledge representation systems -- defining scope, driving schema design, and providing testable acceptance criteria.

In biosciences, CQs bridge the gap between what a researcher wants to know ("By what mechanism does Palovarotene treat FOP?") and the structured graph traversals needed to answer it (Drug --[agonist_of]--> Gene --[regulates]--> Gene --[associated_with]--> Disease).

Why This Dataset?

Existing biomedical QA benchmarks (PubMedQA, BioASQ, MedQA, PrimeKGQA) provide large-scale evaluation but lack:

Feature	This Dataset	PrimeKGQA	BioASQ	MedReason
Standardized CURIEs (HGNC, CHEMBL, MONDO)	Yes	No	Partial	No
BioLink-typed entities & predicates	Yes	No	No	No
Gold standard graph paths	Yes	SPARQL	Triples	Text
Executable API workflow steps	Yes	No	No	No
Multi-database federation	Yes	Single KG	PubMed	Single KG
Fuzzy-to-Fact discovery protocol	Yes	No	No	No

This dataset is small by design -- it prioritizes depth of annotation per question over breadth, making each CQ a complete, reproducible research workflow specification.

Supported Tasks

• Knowledge Graph Evaluation: Test whether a KG can answer domain research questions
• Multi-Hop Biomedical Reasoning: Benchmark for graph traversal across genes, drugs, diseases, pathways
• API Workflow Validation: Executable steps for testing biosciences MCP server integrations
• Entity Resolution Benchmarking: CURIE-grounded evaluation of fuzzy-to-fact discovery
• Ontology Engineering: Template for authoring new competency questions

Dataset Structure

Schema

Field	Type	Description
cq_id	string	Unique identifier (e.g., "cq1")
question	string	Natural language research question
category	string	Domain category (e.g., "Drug Mechanism of Action", "Synthetic Lethality")
disease_area	string	Therapeutic area (e.g., "Rare Disease (FOP)", "Oncology")
reasoning_type	string	Graph reasoning pattern required
complexity	string	intermediate or advanced
num_hops	int	Number of hops in the gold standard path
key_entities	list[dict]	Entities with name, curie, type (BioLink), role
gold_standard_path	string	Expected node-edge-node traversal chain
workflow_steps	list[string]	Ordered API calls to answer the question
source	string	Origin (DrugMechDB, DALK, STRING 2025, BioThings, Feng et al., Original)
source_reference	string?	Academic citation if applicable
group_id	string	Graphiti knowledge graph persistence target
apis_used	list[string]	APIs required (HGNC, ChEMBL, STRING, etc.)
biolink_edges	list[dict]	Typed edges with source, target, predicate

Reasoning Types

Type	Description	Example CQs
multi_hop_traversal	Follow a chain of typed edges	cq1, cq2, cq4, cq8, cq11, cq14
network_expansion	Expand outward from anchor node(s)	cq3
directed_traversal	Follow regulatory direction (activation/inhibition)	cq5
bidirectional_expansion	Expand both upstream and downstream	cq6
federated_multi_hop	Multi-hop across federated APIs	cq7
comparative_analysis	Compare entities side-by-side	cq9
set_difference	Find elements in set A not in set B	cq10
aggregation	Count, group, rank across entities	cq12
ranking	Order by computed metric	cq13
temporal_analysis	Track progression over time	cq15

Categories

Category	Count	CQs
Drug Mechanism / Repurposing	3	cq1, cq2, cq7
Gene-Protein Networks	2	cq3, cq6
Therapeutic Target Discovery	2	cq4, cq10
Synthetic Lethality	2	cq8, cq14
Signaling / Pathway Validation	2	cq5, cq11
Drug Safety	1	cq9
Clinical Trial Landscape	2	cq12, cq13
Regulatory Analysis	1	cq15

Example Instance (cq1)

{
  "cq_id": "cq1",
  "question": "By what mechanism does Palovarotene treat Fibrodysplasia Ossificans Progressiva (FOP)?",
  "category": "Drug Mechanism of Action",
  "disease_area": "Rare Disease (FOP)",
  "reasoning_type": "multi_hop_traversal",
  "complexity": "intermediate",
  "num_hops": 3,
  "key_entities": [
    {"name": "Palovarotene", "curie": "CHEMBL:2105648", "type": "biolink:SmallMolecule", "role": "RAR-gamma agonist"},
    {"name": "RARG", "curie": "HGNC:9866", "type": "biolink:Gene", "role": "Retinoic acid receptor gamma"},
    {"name": "ACVR1", "curie": "HGNC:171", "type": "biolink:Gene", "role": "BMP receptor (causal gene)"},
    {"name": "FOP", "curie": "MONDO:0007606", "type": "biolink:Disease", "role": "Target disease"}
  ],
  "gold_standard_path": "Drug(Palovarotene) --[agonist]--> Protein(RARG) --[regulates]--> Protein(ACVR1) --[causes]--> Disease(FOP)",
  "workflow_steps": [
    "Anchor: chembl_search_compounds('palovarotene') -> CHEMBL:2105648",
    "Enrich: chembl_get_compound('CHEMBL:2105648') -> Max Phase 4, indications",
    "Mechanism: ChEMBL /mechanism -> RARG agonist",
    "Target Gene: hgnc_get_gene('HGNC:171') -> ACVR1 details",
    "Disease Link: opentargets_get_associations('ENSG00000115170') -> FOP association",
    "Persist: graphiti add_memory(group_id='cq1-fop-mechanism')"
  ],
  "source": "DrugMechDB",
  "apis_used": ["ChEMBL", "HGNC", "Open Targets"],
  "biolink_edges": [
    {"source": "CHEMBL:2105648", "target": "HGNC:9866", "predicate": "biolink:agonist_of"},
    {"source": "HGNC:9866", "target": "HGNC:171", "predicate": "biolink:regulates"},
    {"source": "HGNC:171", "target": "MONDO:0007606", "predicate": "biolink:gene_associated_with_condition"}
  ]
}

Dataset Creation

Curation Rationale

These 15 CQs were designed to cover the core reasoning patterns needed for AI-powered biosciences research:

• Mechanistic questions (cq1, cq5, cq11): How does drug X work? What regulates gene Y?
• Discovery questions (cq2, cq7, cq10): What drugs could be repurposed? What targets are uncovered?
• Safety questions (cq9): What are the off-target risks?
• Validation questions (cq8, cq14): Can we confirm synthetic lethal relationships?
• Landscape questions (cq12, cq13, cq15): What does the trial/regulatory landscape look like?

The CQs span rare diseases (FOP, NGLY1 deficiency, Huntington's), common diseases (Alzheimer's, cancer), and cross-cutting concerns (drug safety, clinical trial prioritization).

Source Data

Source	CQs	Description
DrugMechDB	cq1, cq2	Drug mechanism database patterns
Li et al. (2024) DALK	cq3, cq4	Alzheimer's KG from 9,764 PubMed papers
Szklarczyk et al. (2025) STRING	cq5, cq6	Directed regulatory networks
Callaghan et al. (2023) BioThings	cq7	Federated KG from 34 biomedical APIs
Feng et al. (2022)	cq14	209 synthetic lethal gene pairs
Original research	cq8-13, cq15	Novel questions for this project

Annotations

Each CQ was:

1. Authored by domain researchers following the Fuzzy-to-Fact protocol
2. Validated by executing workflow steps against live biosciences MCP API servers (12 FastMCP servers covering HGNC, UniProt, ChEMBL, Open Targets, STRING, BioGRID, Ensembl, Entrez, PubChem, IUPHAR, WikiPathways, ClinicalTrials.gov)
3. Persisted to a Graphiti knowledge graph with dedicated group_id namespaces

CQ Type Classification (Keet & Khan 2024)

Following the ROCQS taxonomy of competency question types:

CQ Type	Abbreviation	Present In
Scoping	SCQ	cq3 ("What genes are implicated..."), cq12, cq15
Validating	VCQ	cq11 ("How do we validate..."), cq14
Relationship	RCQ	cq1, cq5, cq6, cq8, cq9
Foundational	FCQ	Implicit in BioLink typing
Metaproperty	MpCQ	Not directly represented

Considerations

Known Limitations

• Small size: 15 CQs -- designed as a sample/template, not a large-scale benchmark
• CURIE staleness: Database identifiers may change as HGNC, ChEMBL, etc. update
• API dependency: Workflow steps require access to external biosciences APIs
• Domain skew: Oncology and rare disease overrepresented; infectious disease underrepresented
• English only: All questions and annotations in English

Relation to Other Datasets

• Complements PrimeKGQA (84K pairs, single-KG SPARQL) with multi-API federation
• Extends BioASQ concept by adding executable graph paths
• Parallels MedReason (32K pairs) but with CURIE-grounded rather than text-only reasoning
• Builds on CQ-to-SPARQL-OWL (234 CQs) methodology for biosciences domain

Future Extensions

• Paul Zamora's 12 oncology CQs (Doxorubicin toxicity, tumor microenvironment, NSCLC synthetic lethality)
• Additional CQs from community contributions
• Validation result datasets with API response snapshots
• HuggingFace Datasets versioning for reproducible evaluation

References

1. Gruninger, M. & Fox, M.S. (1995). "The Role of Competency Questions in Enterprise Engineering." IFIP AICT.
2. Ren, Y. et al. (2014). "Towards Competency Question-Driven Ontology Authoring." ESWC 2014.
3. Wisniewski, D. et al. (2019). "Analysis of Ontology Competency Questions and their Formalizations in SPARQL-OWL." J. Web Semantics.
4. Bezerra, C. et al. (2023). "Use of Competency Questions in Ontology Engineering: A Survey." ER 2023.
5. Keet, C.M. & Khan, Z.C. (2024). "Discerning and Characterising Types of Competency Questions for Ontologies." EKAW 2024.
6. Li, D. et al. (2024). "DALK: Dynamic Co-Augmentation of LLMs and KG." EMNLP 2024 Findings.
7. Szklarczyk, D. et al. (2025). "STRING 2025: protein networks with directionality." Nucleic Acids Res.

Citation

@dataset{open_biosciences_cq_2026,
  title={Open Biosciences Competency Questions (Sample)},
  author={Open Biosciences Project},
  year={2026},
  url={https://huggingface.co/datasets/open-biosciences/biosciences-competency-questions-sample},
  license={MIT}
}

Project Context

Part of the Open Biosciences platform -- a multi-repo workspace for AI-powered biosciences research. Maintained by the Research Workflows Engineer (Agent #6).