Predicting temozolomide response in low-grade glioma patients with large-scale machine learning

Abstract Background Temozolomide is the primary chemotherapeutic agent and first-line treatment for low-grade glioma. Although low-grade gliomas are generally less aggressive than high-grade gliomas, they can eventually progress into high-grade gliomas, making it crucial to maximise the efficacy of initial treatment. Methods We analysed data from 109 patients with low-grade gliomas in The Cancer Genome Atlas to evaluate the predictive performance of 12 machine learning classification algorithms for temozolomide response, using six types of omics data. Cross-validation and bootstrapping bias correction were applied to compare these models with a conventional biomarker-based model using promoter methylation status of O6-methylguanine-DNA methyltransferase. The Matthews Correlation Coefficient (MCC) was used as the primary evaluation metric. Results The microRNA-based model using the Extreme Gradient Boosting algorithm achieved the best performance (MCC = 0.447), outperforming both the automated machine learning method JADBio (MCC = 0.250) and the biomarker-based model (MCC = 0.331). Incorporating clinical variables, such as patient age and Karnofsky score, further improved predictive power, with the logistic regression model with optimal model complexity achieving the highest MCC (0.483). Feature importance analysis on the best model revealed six predictive microRNAs, including three tumour-related factors (miR-335, let-7f, and miR-7-2) and three potential biomarkers (miR-204, miR-6513, and miR-376). Discussion This study systematically demonstrates the potential of large-scale analyses combining machine learning and omics data to predict temozolomide response, offering superior predictive accuracy compared with standard biomarkers. However, validation in independent clinical datasets remains necessary before clinical translation.