Biologically Supervised Graph Embedding for Disease Comorbidity Prediction: a Versatile Framework for Various Embedding Techniques and Classifiers

Comorbidity is essential for understanding and managing diseases and is believed to arise at the genetic level mutations connected through the protein-protein interactions (PPI) within the human interactome. However, the complexity and incompleteness of the human interactome pose challenges in extracting useful features for comorbidity prediction. In this study, we introduce a new framework called Biologically Supervised Graph Embedding (BSE) to select the most relevant features from the graph embedding capturing disease subgraphs relation representation, which can lead to improving the accuracy of disease comorbidity prediction. Our investigation into BSE's impact on both centered and uncentered embedding methods showcases its consistent superiority over the state-of-the-art techniques and its adeptness in selecting features enriched with vital biological insights, thereby improving prediction performance significantly, up to 50% when measured by ROC AUC score. Detailed analysis indicates that BSE consistently and substantially extract features with relatively higher ratio of disease associations to gene connectivity, affirming its potential in uncovering latent biological factors affecting comorbidity. We further demonstrate statistically significant enhancements across various embedding methods using different metrics and classifiers, highlighting BSE's versatility and broad applicability. This study highlights BSE's ability to introduce novel avenues for precise disease comorbidity predictions and other potential applications.