Classification of Diabetic Patients using a Network Representation of Their Metabolism

Studies on Type 2 Diabetes Mellitus (T2DM) rely on specific metabolic networks to represent the intricate relationships between metabolites. Accurate classification requires analyzing network characteristics, such as distance graphs and topological similarities, and identifying features that effectively capture these aspects. This study focuses on deriving metabolic networks and applying graph embeddings to achieve optimal feature representation and classification performance. We extract metabolic networks from large patient cohorts and targeted tissues, comprising metabolism and gene expression data. We label patients into three groups: T2DM, non-T2DM, and Healthy based on the occurrence of T2DM enzymes in the referenced dataset. We build classification models using traditional machine learning techniques and Graph Neural Networks (GNNs) approaches based on extracted features. The models are evaluated on several statistical tests, identifying the best classification model for new patient data. The impact of interference factors in normalized feature data and perturbation on classification performance is also analyzed.