An Efficient Machine Learning Method to Solve Imbalanced Data in Metabolic Disease Prediction

The increase of obesity, its related diseases and the high incidence of metabolic diseases as a whole, constitute a major public health problem on a global scale. New strategies that allow for the discovery of novel metabolic disease-related genes are necessary to develop new treatments. In this paper, we proposed an efficient method to predict metabolic disease genes, solving the problem of imbalanced data. The method combined protein-protein interactions and miRNA-target interactions to construct integrated networks, whose topological properties can be used as features to train machine learning classifiers. We applied different strategies to optimize imbalanced class. The best model of gradient boosting achieved a significant F1-score of 0.82. When testing the model with non-disease genes, we predicted 549 candidates, out of which 123 were validated indirectly from literature to be related to metabolic diseases. The remaining genes' functions were investigated by gene enrichment analysis, revealing their association with diseases known to co-occur with metabolic diseases, such as cancer and cardiovascular conditions. These results indicated that this method contributed to the identification of novel metabolic disease-related genes.