Patients’ networks for clinical phenotype/outcome prediction

Methods for phenotype and outcome prediction in bioinformatics usually rely on supervised models that employ set of biomarkers to discriminate between patients but do not directly consider the functional or genetic relationships between them. Instead, the field of “Network Medicine” adopts the whole set of connections between biomolecular components to e.g. ranking genes with respect to a given phenotype. Since our aim is to consider at the same time the relationships between individuals and their a priori available phenotypic information, we developed a network-based method called Sample-Net (S-Net) to rank patients with respect to an outcome of interest. S-Net builds a network where nodes are samples/patients and edges similarities between them based on their clinical or genetic profiles (e.g. expression profiles). S-Net exploits the topological features of the net to assign a score to each individual and the application of a graph-kernel (e.g. random walk kernel) enriches the net with new informative edges. Then a semi-supervised method, a local learning strategy based on the guilt-by-association principle, assigns a score to each sample representing its odd to show the considered outcome and allowing its ranking. We compared S-Net with supervised and semi-supervised methods on real datasets including Pancreatic, Melanoma, Ovarian, Breast, Colorectal and Colon cancer samples. S-Net not only is competitive with state-of-the-art methods but it can improve the predictions for patients difficult to classify with other methods. The graph of samples can be easily visualized to gain clues about the relationships between samples, considering the phenotype associated and predicted for each patient. S-Net is able to effectively predict the phenotype, while the graphical representation of the patients’ net provides significant insights into their biomolecular stratification. A fast implementation of S-Net is available online: https://github.com/GliozzoJ/S-Net.