Simpati: patient classifier identifies signature pathways based on similarity networks for the disease prediction

BACKGROUNDPathway-based patient classification is a supervised learning task which implies a model learning pathways as features to predict the classes of patients. The counterpart of enrichment tools for the pathway analysis are fundamental methods for clinicians and biomedical scientists. They allow to find signature cellular functions which help to define and annotate a disease phenotype. They provide results which lead human experts to manually classify patients. It is a paradox that pathwaybased classifiers which natively resolve this objective are not strongly developed. They could simulate the human way of thinking, decipher hidden multivariate relationships between the deregulated pathways and the disease phenotype, and provide more information than a probability value. Instead, there are currently only two classifiers of such kind, they require a nontrivial hyperparameter tuning, are difficult to interpret and lack in providing new insights. There is the need of new classifiers which can provide novel perspectives about pathways, be easy to apply with different biological omics and produce new data enabling a further analysis of the patients. RESULTSWe propose Simpati, an innovative and interpretable patient classifier based on pathway-specific patient similarity networks. The first classifier to adopt ad-hoc novel algorithms for such graph type. It standardizes the biological high-throughput dataset of patients profiles with a propagation algorithm that considers the interconnected nature of the cells molecules for inferring a new activity score. This allows Simpati to classify with dense, sparse, and non-homogenous omic data. Simpati organizes patients molecules in pathways represented by patient similarity networks for being interpretable, handling missing data and preserving the patient privacy. A network represents patients as nodes and a novel similarity measure determines how much every pair act co-ordinately in a pathway. Simpati detects signature biological processes based on how much the topological properties of the related networks separate the patient classes. In this step, it includes a new cohesive subgroup detection algorithm to handle patients not showing the same pathway activity as the other class members. An unknown patient is then classified by a unique recommender system which considers how much is similar to known patients and distant from being an outlier. Simpati outperforms previously published classifiers on five cancer datasets described with two biological omics, classifies well with sparse data, identifies more relevant pathways associated to the patients disease than the competitors and has the lowest computational requirements. CONCLUSIONSimpati can serve as generic-purpose pathway-based classifier of patient classes. It provides signature pathways to unveil the altered biological mechanisms of a disease phenotype and to classify patients according to the learnt pathway-specific similarities. The signature condition and patient prediction can be deciphered considering the patient similarity networks which must reveal the members of a patient class more cohesive and similar than the non-members. Simpati divides the pathways in up and downinvolved. Upinvolved when the signaling cascades generated by the altered molecules of the disease patients impact stronger the pathway than the ones of the control class. We provide an R implementation, a graphical user interface and a visualization function for the patient similarity networks. The software is available at: https://github.com/LucaGiudice/Simpati