Factors Affecting Network-Based Gene Prediction Across Diverse Diseases

There are many current efforts to integrate biological interaction data with disease information in order to predict new genes associated with complex diseases. Network-based learning methods such as logistic regression can utilize this information to identify disease genes, and are typically applied to protein-protein interaction networks. However, little is reported about what factors influence the performance of these network-based methods. Here, we explore features that affect networkbased disease gene prediction performance. We devise two cross-validation schemes to evaluate the impact of various parameters, settings and disease qualities across a wide range of diseases. We demonstrate that including gene regulatory interactions and including low-confidence disease genes improves disease gene prediction performance. Further, network connectivity among high-confidence disease genes is a strong indicator of prediction performance. We demonstrate that network and input features can have a dramatic effect on prediction performance, and these should be carefully considered when designing network-based algorithms to find new disease genes.