Abstract
Advancements in genome-wide profiling techniques have revolutionized the generation of large-scale genomic data, enabling the successful identification and classification of cancer tissue samples based on their gene expression patterns. Despite these strides, the widespread adoption of such methodologies, particularly in constructing single sample predictors (SSPs) from gene expression profiles, has long faced challenges due to the lack of calibration across different gene expression measurement technologies. Nonetheless, research has highlighted the feasibility of phenotype classification through the expression order of multiple genes. Traditional TSP-based methods, while interpretable based on "relative expression reversals", struggle with complex pattern misclassification when more than two gene expression comparisons are involved in the decision boundary. To overcome this limitation, we propose a method that extends TSP rules by constructing rank-based trees, enabling the accommodation of gene-gene comparisons. To address overfitting, we incorporate two ensemble strategies, boosting, and random forest, to enhance the robustness of the rank-based trees. Our implementation of ensembled rank-based trees in boosting with LogitBoost cost and random forests on 12 binary and multi-class cancer gene expression datasets demonstrates superior classification accuracy compared to K-Top Scoring Pairs (k-TSP) and Nearest Template Prediction (NTP) methods. Furthermore, we elucidate the process of variable selection and the generation of concise yet precise decision rules from rank-based trees for interpretive purposes. Our research findings consistently underscore the substantial promise of ensemble rank-based trees for enhancing disease classification through gene expression data. Our software is available at https:// CRAN.R-project.org/package=ranktreeEnsemble.