Development of an interpretable ensemble model to predict immune checkpoint inhibitor therapy response in bladder cancer patients

Abstract

Identifying robust biomarkers for bladder cancer and immunotherapy continues to be a challenge in the field of computational biology, as pertinent datasets often suffer from the “curse of dimensionality.” In this study, we explore the effectiveness of ensemble feature selection methods to discover potential and validate existing biomarkers that are predictive of sensitivity to immune checkpoint inhibitor (ICI) therapy, administered using Pembrolizumab. The thesis utilizes the GSE111636 dataset with 11 samples and over 70,000 features to simulate the high-dimensionality problem and uses a tailored ensemble modeling approach to tackle it. To be more specific, the research applies four popular feature selection algorithms - Support Vector Machine Recursive Feature Elimination (SVM-RFE), Random Forest Recursive Feature Elimination (RF-RFE), Mutual Information (MI), and Logistic Regression Recursive Feature Elimination (LR-RFE) - both individually and in ensembles. To aggregate the results of the ensemble models, a frequency-based majority voting system is adopted. Furthermore, keeping in mind the limitations of the dataset, Leave-One-Out Cross-Validation (LOOCV) is also used to improve on the generalizability and robustness and reduce bias of the feature selection methods. Two deep learning models - one based on neural networks with L1 Regularization and another Concrete Autoencoder - were also used to carry out feature selection. A final comparison between all the models demonstrated that ensemble models outperformed both single models and deep learning models by selecting a greater number of biologically relevant gene features, most of which have been previously implicated in bladder cancer progression and immune response, while also showcasing more stable feature selection metrics. Thus, this study highlights the utility of ensemble modeling in low sample count, high-dimensional gene expression datasets, underscoring its potential for advancing biomarker research in oncology.