Heterojen Biyomedikal Verinin Bilgi Çizgeleri ve Derin Öğrenme Tabanlı Analizi ile Protein Fonksiyonlarının Otomatik Tahmini

Abstract

Proteins are vital for cellular processes, and accurately determining their functions is crucial for understanding complex biological mechanisms. Computational approaches have emerged as alternatives to expensive and time-consuming experimental methods, leveraging publicly available data in biomedical databases to predict protein functions. However, existing methods often rely on a single data type, limiting their ability to capture the multifaceted functional complexity of proteins. Geometric deep learning offer new algorithms that can be utilized to address these issues by integrating diverse biological entities and relationships sourced from multiple databases using heterogeneous graphs. In this thesis study, we propose a heterogeneous graph learning approach and its implementation as a computational method for Gene Ontology (GO) based large-scale protein function prediction. For this, we first constructed a comprehensive biological knowledge graph by obtaining and integrating data from 14 different biomedical databases. Using this dataset, we trained function prediction models using graph neural networks, i.e., the heterogeneous graph transformer architecture. Performance evaluation on benchmark datasets indicated superior performance compared to baseline methods across all GO categories, while achieving comparable results to top predictors. Our model demonstrated excellent performance in predicting high-information-content molecular function terms, ranking among the top three models. To assess the biological relevance of predicted functional relationships, we conducted a use-case study for selected proteins, showcasing our approach's ability to identify unknown functions with limited available information. This study contributes to the existing literature by investigating protein function prediction using geometric deep learning on highly heterogeneous biomedical data.