PREDICTING DISEASE-GENE ASSOCIATIONS VIA MACHINE LEARNING

Abstract

In the quest to elucidate disease etiology and develop advanced diagnostic and treatment tools, knowing disease-gene relationships is of great importance. Traditional approaches based on manual curation fall short due to limited scalability and precision. On the other hand, graph neural networks (GNN) enable the analysis of complex relational data within biological networks. Although the GNN-based methods developed to date have produced positive results in predicting unknown biological relationships, there is a current need to develop new models with high prediction performance and generalisation capabilities for usability in biology and medicine. In this thesis study, we propose GLADIGATOR (Graph Learning bAsed DIsease Gene AssociaTiOn pRediction), a deep learning model designed with the encoder-decoder architecture to predict disease-gene associations. GLADIGATOR creates a heterogeneous graph that primarily integrates two types of biological components, genes and diseases, and the connections between them. The model was trained using gene-gene, disease-disease and gene-disease relationships existing in source biological databases, as well as protein sequence representations generated by the Prot-T5[1] protein language model and disease representations generated by the BioBert[2] language model, as node feature vectors. As the outcome of the analyses conducted, it was observed that i GLADIGATOR had superior prediction accuracy. Additionally, the model was positioned as the highest performer among 14 different disease-gene association prediction methods. Literature-driven studies on selected predictions have confirmed the biological relevance of predicted novel associations and highlighted the effectiveness of the GNN-based approach in identifying potential candidate genes for specific diseases. These results may provide valuable information for discovering new drugs as a result of future experimental validation analyses. GLADIGATOR has not only enriched computational approaches developed for disease-gene association prediction but also emphasised the transformative abilities of GNNs in biomedical research by potentially accelerating the discovery of new biological relationships.