Akademik Veri Yönetim Sistemi
Intelligent Systems for Molecular Biology (ISMB) & European Conference on Computational Biology , Liverpool, İngiltere, 20 - 24 Temmuz 2025, ss.648-649, (Özet Bildiri)
We propose TransFusionKG, a novel model that integrates TransE-based translational embeddings with a graph denoising transformer to enhance the discovery of novel drug-disease associations. Built on a structured biomedical knowledge graph synthesized from SemMedDB, repoDB, and UMLS, our model aims to address the limitations of static embedding models in capturing high-order semantics and noise in biomedical graphs. TransE initializes entity and relation embeddings by preserving translational distance among triples, while a transformer-based encoder with attention over node neighborhoods is used to refine the embeddings by accounting for global graph structure. To further improve generalizability, we incorporate a score-based graph denoising module that learns to denoise the corrupted embeddings through iterative refinement, improving prediction stability. The model is trained using known associations and relations, and evaluated on unapproved drug-disease pairs from repoDB, enabling realistic testing for drug repositioning. Candidate pairs are ranked using a distance-based scoring function, where smaller vector distances indicate stronger predicted associations. TransFusionKG highlighted 'Temozolomide TREATS Rhabdomyosarcoma' among its top-ranked predictions, a relationship supported in biomedical literature where temozolomide is used-particularly in combination therapies-as a treatment option for rhabdomyosarcoma. Conversely, low-scoring predictions like 'Racepinephrine TREATS Postoperative Pain' lacked supporting evidence in biomedical literature, demonstrating TransFusionKG's effectiveness in down-ranking unsupported or speculative associations. Our results demonstrate that TransFusionKG produces refined biomedical embeddings capable of accurately ranking candidate drug-disease associations, offering a robust and interpretable computational method to support drug repositioning and accelerate hypothesis generation in biomedical research.