A Comparative Analysis of Machine Learning Models on Graph Embeddings for Disease-Protein Association

Understanding protein-disease associations (PDAs) is crucial for uncovering disease mechanisms and identifying novel therapeutic targets. Graph-based representations of biological networks like PPI (protein-protein interaction) and PDA (protein disease associations) are used for disease prediction. To enable efficient use of machine learning algorithms for classification, these graphs must be converted to meaningful node embeddings that could capture topological and relational information from graphs. The purpose of this study is to assess how well the various embedding techniques work in representing disease-gene association networks and to evaluate their performance in downstream classification tasks. We applied three embedding techniques: Node2Vec, Fast Random Projection (FastRP), and GraphSAGE to convert the network into lowdimensional vector representations. These embeddings were then used as input features for machine learning classifiers: random forest, adaboost, Catboost and SVM. The models were trained to distinguish Alzheimer's disease-related genes from non-associated ones. We conducted a comparative analysis of different node embeddings and classifiers using various performance metrics. An evaluation of graph embeddings was conducted in the context of Alzheimer's disease classification, revealing their potential and effectiveness in capturing meaningful patterns for disease-gene association analysis.