Abstract
Reliable links between genes and diseases are central to biomedical research; however, many computational methods overlook the semantic and hierarchical layers of ontologies, missing indirect relationships and producing shallow association scores. We propose an ontology-driven framework for gene-disease association mining that integrates hierarchical knowledge from the Gene Ontology and Disease Ontology. Our text-mining pipeline processes PubMed text by cleaning, annotating, and extracting sentence-level co-occurrences of biomarker-related terms. We evaluated and compared well-known association rule mining algorithms, namely Apriori, FP-Growth, and Eclat, and applied a tie-aware rank-based transformation to correct for non-normal distributions of association scores. The resulting Athar Semantic Enriched Association (ASEA) score combines entity-specific associations with Hierarchical Ontology Associations, with an enhanced Apriori variant showing superior performance in capturing direct and indirect associations. Benchmarking against the Comparative Toxicogenomics Database, ASEA detected 17 high-grade associations (30.4% more than Apriori and Eclat, 88.9% more than FP-Growth). In total, ASEA produced 185 associations, compared with 217 for Apriori, 166 for Eclat, and 71 for FP-Growth. Among these, 21 belong to high-confidence databases (Case 1), 28 are supported by substantial literature, but not yet high-confidence (Case 2), 39 have low/intermediate database support with no strong literature (Case 3), and 22 are purely speculative (Case 4), including 12 particularly novel associations absent from the curated resources. Overall, this framework provides a transparent and extensible pipeline for biomedical knowledge discovery, combining statistical co-occurrence with ontology-driven enrichment to retrieve established knowledge and generate reliable predictions for precision medicine and hypothesis-generation.