Gene-Disease Association Extraction is the biomedical NLP task of automatically identifying relationships between genes, genetic variants, and human diseases from scientific literature — populating the knowledge bases that drive Mendelian disease gene discovery, polygenic risk score construction, cancer driver identification, and precision medicine by extracting the genetic-disease links documented across millions of biomedical publications.

What Is Gene-Disease Association Extraction?

• Task Definition: Relation extraction identifying (Gene/Variant, Disease, Association Type) triples from biomedical text.
• Association Types: Causal (gene mutation causes disease), risk (variant increases susceptibility), therapeutic target (gene modulation treats disease), biomarker (gene expression indicates disease state), complication (disease causes gene dysregulation).
• Key Databases Populated: DisGeNET (1.1M gene-disease associations), OMIM (Mendelian genetics), ClinVar (variant-disease clinical significance), COSMIC (cancer somatic mutations), PharmGKB (pharmacogenomics).
• Key Benchmarks: BC4CHEMD (chemical-gene), BioRED (multi-entity relation), NCBI Disease Corpus, CRAFT Corpus.

The Association Extraction Challenge

Gene-disease associations in literature come in many forms:

Direct Causal Statement: "Mutations in CFTR cause cystic fibrosis." → (CFTR gene, Cystic Fibrosis, Causal).

Statistical Association: "The rs12913832 SNP in OCA2 is associated with blue eye color (p < 10−300)." → (rs12913832 variant, eye color phenotype, GWAS association).

Mechanistic Description: "Overexpression of HER2 drives proliferation in breast cancer by activating the PI3K/AKT pathway." → (ERBB2/HER2, Breast Cancer, Driver).

Negative Association: "No significant association between APOE ε4 and Parkinson's disease was found in this cohort." → Negative/null finding — critical to prevent false positive database entries.

Speculative/Hedged: "These data suggest LRRK2 may be involved in sporadic Parkinson's disease." → Uncertain evidence — must be distinguished from confirmed associations.

Entity Recognition Challenges

• Gene Name Ambiguity: "CAT" is the gene catalase but also an English word. "MET" is the hepatocyte growth factor receptor but also a preposition.
• Synonym Explosion: TP53 = p53 = tumor protein 53 = TRP53 = FLJ92943 — gene entities have dozens of aliases.
• Variant Notation: "p.Glu342Lys," "rs28931570," "c.1024G>A" — three notations for the same SERPINA1 variant causing alpha-1 antitrypsin deficiency.
• Disease Ambiguity: "Cancer," "tumor," "malignancy," "neoplasm," "carcinoma" — hierarchical disease terms requiring OMIM/DOID normalization.

Performance Results

Clinical Applications

Rare Disease Diagnosis: When a patient's whole-exome sequencing reveals a variant of uncertain significance (VUS) in a poorly characterized gene, automated gene-disease extraction can find publications describing similar variants in similar phenotypes.

Cancer Driver Analysis: Mining literature for somatic mutation-cancer associations populates COSMIC and OncoKB — databases used by oncologists to interpret tumor sequencing reports.

Drug Target Validation: Gene-disease association strength (number of independent studies, effect sizes) is a key predictor of the probability that targeting the gene will treat the disease.

Pharmacogenomics: CYP2D6, CYP2C9, and other pharmacogene-drug interaction associations extracted from literature directly inform FDA drug labeling with genotype-guided dosing recommendations.

Gene-Disease Association Extraction is the genetic medicine knowledge engine — systematically mining millions of publications to build the gene-disease knowledge base that connects genomic variants to clinical phenotypes, enabling precision medicine applications from rare disease diagnosis to oncology treatment selection.