README.md

CGBench Dataset

This repository contains the CGBench dataset - a comprehensive benchmarking framework for evaluating scientific reasoning in language models using ClinGen (clinicalgenome.org) data. The dataset enables evaluation of language models' ability to extract, interpret, and explain fine-grained results from scientific publications in clinical genetics.

Dataset Overview

CGBench formulates three separate tasks:

1. GCI Evidence Extraction - Gene-disease association evidence extraction
2. VCI Evidence Scoring - Variant pathogenicity evidence scoring
3. VCI Evidence Sufficiency - Evidence sufficiency determination

Directory Structure

GCI/ - Gene Curation Interface Data

Contains data related to gene-disease association evidence extraction tasks.

Core Files

Clingen-Gene-Disease-Summary-2025-03-31.csv

• Summary of all ClinGen gene-disease validity classifications
• Columns:
  • GENE SYMBOL: HGNC gene symbol
  • GENE ID (HGNC): HGNC gene identifier
  • DISEASE LABEL: Human-readable disease name
  • DISEASE ID (MONDO): MONDO ontology disease identifier
  • MOI: Mode of inheritance (AD=Autosomal Dominant, AR=Autosomal Recessive, XL=X-linked)
  • SOP: Standard Operating Procedure version used for curation
  • CLASSIFICATION: Gene-disease association strength (Definitive, Strong, Moderate, Limited)
  • ONLINE REPORT: URL to full ClinGen report
  • CLASSIFICATION DATE: Date of classification
  • GCEP: Gene Curation Expert Panel responsible for classification

SOP/ - Standard Operating Procedures

experimental_evidence/ Contains JSON files defining evidence categories and scoring criteria:

• SOP5.json through SOP11.json: Different versions of standard operating procedures
• Each file contains structured evidence categories with titles and descriptions for:
  • Biochemical Function (A & B)
  • Protein Interaction
  • Expression (A & B)
  • Functional Alteration
  • Model Systems
  • Rescue experiments

evidence_tables/

experimental_evidence/evidence_cleaned_fulltext.csv

• Complete dataset of experimental evidence with full-text paper content
• Columns:
  • Label: Evidence description/title
  • Experimental Category: Type of evidence (e.g., "Model Systems Non-human model organism")
  • Reference: Citation with PMID
  • Explanation: Detailed explanation of the evidence
  • Score Status: Whether evidence was scored
  • Points (default points): Numerical score assigned
  • Reason for Changed Score: Explanation if score was modified
  • url: Link to ClinGen assertion
  • primary_index: Internal identifier
  • pmid: PubMed identifier

Train/Test Splits:

• train.csv, test.csv: Standard splits for model training/evaluation
• train_datesplit.csv, test_datesplit.csv: Date-based splits to avoid data leakage

pubmed/

experimental_evidence.csv

• PubMed abstracts associated with evidence entries
• Links evidence to published literature

VCI/ - Variant Curation Interface Data

Contains data for variant pathogenicity assessment tasks.

Core Files

clingen_vci_pubmed_fulltext.csv

• Complete VCI dataset with full-text papers
• Columns:
  • entry_index: Unique identifier
  • variant: HGVS variant notation
  • hgnc_gene: Gene symbol
  • disease: Associated disease
  • mondo_id: MONDO disease identifier
  • assertion: Pathogenicity classification (Pathogenic, Likely Pathogenic, Uncertain Significance, etc.)
  • mode_inheritance: Inheritance pattern
  • expert_panel: Responsible VCEP (Variant Curation Expert Panel)
  • pub_date: Publication date
  • evidence_code: ACMG/AMP evidence code applied
  • met_status: Whether evidence criteria was met
  • pmid: Associated PubMed ID
  • comments: Curator comments
  • summary: Evidence summary
  • summary_comments: Additional summary information

Data Variants:

• clingen_vci_pubmed_fulltext_dedup_pmid.csv: Deduplicated by PMID
• clingen_vci_pubmed_fulltext_vceps.csv: VCEP-specific subset
• clingen_vci_pubmed_var_na_filtered.csv: Filtered for missing variants

erepo.tabbed_2025-02-25.txt

• Raw ClinGen Evidence Repository export
• Tab-delimited format with comprehensive variant classification data
• Key Columns:
  • #Variation: Variant identifier
  • ClinVar Variation Id: ClinVar identifier
  • HGVS Expressions: All HGVS representations
  • Applied Evidence Codes (Met/Not Met): ACMG/AMP criteria
  • Summary of interpretation: Detailed reasoning
  • Expert Panel: Responsible VCEP

pubmed_id_to_text.csv

• Maps PubMed IDs to full-text content
• Columns:
  • pmid: PubMed identifier
  • abstract: Paper abstract
  • full_text: Complete paper text when available

Task-Specific Splits

split_evidence_score/ Evidence scoring task data:

• train.csv, test.csv: Basic train/test splits
• train_merged.csv, test_merged.csv: Merged with additional metadata
• Additional path column indicates source VCEP specification file

split_evidence_sufficiency/ Evidence sufficiency determination task:

• train.csv, test.csv: Basic splits
• train_dedup.csv, test_dedup.csv: Deduplicated versions
• test_dedup_missing.csv: Test set with missing data scenarios

parsing_csr_criteria/

Contains tools and data for parsing Criteria Specification Registry (CSR) documents:

• cspec_version_guide.csv: Maps VCEP specifications to versions
• cspec_version_guide_processed.csv: Processed version mappings
• version_csv_individual/: Individual CSV files for each VCEP specification version
  • Format: {VCEPName}_version={X.Y.Z}.csv
  • Contains parsed criteria specifications for each expert panel

Python Scripts:

• get_versions.py: Extract version information
• parse_on_date.py: Parse specifications by date
• scrape_criteria.py: Scrape criteria from registry
• scrape_criteria_fn.py, scrape_criteria_versions.py: Helper functions

Data Usage Notes

Evidence Codes

The dataset uses ACMG/AMP variant interpretation guidelines with evidence codes like:

• PVS1: Very strong pathogenic evidence
• PS1-4: Strong pathogenic evidence
• PM1-6: Moderate pathogenic evidence
• PP1-5: Supporting pathogenic evidence
• BA1: Stand-alone benign
• BS1-4: Strong benign evidence
• BP1-7: Supporting benign evidence

Expert Panels (VCEPs)

Various disease-specific Variant Curation Expert Panels contribute data:

• Phenylketonuria VCEP (PAH gene)
• Cardiomyopathy VCEP
• BRCA1/BRCA2 VCEP (ENIGMA)
• And many others

Data Splits

• Standard splits: Random train/test division
• Date splits: Temporal division to prevent data leakage
• Deduplicated versions: Remove duplicate entries by PMID or other criteria

File Formats

• CSV: Comma-separated values for tabular data
• JSON: Structured data for evidence category definitions
• TXT: Tab-delimited raw exports from ClinGen systems

Citation

If you use this dataset, please cite the CGBench paper (publication details forthcoming).

License

MIT License - see main repository for details.