A Machine Learning Framework for Tumour Classification Using Transcriptomic and Multi-Omics Datasets

Abstract

Cancer is a biologically heterogeneous disease characterized by molecular alterations across multiple regulatory
layers, necessitating robust computational modelling for accurate diagnosis and biomarker discovery. The increasing availability of
high-dimensional genomic and multi-omics datasets from largescale initiatives such as The Cancer Genome Atlas (TCGA)
has enabled the development of machine learning approaches for cancer classification. However, challenges including extreme
dimensionality, feature redundancy, and class imbalance continue affect model stability and generalization performance.
In this study, we propose a reproducible integrative machine learning framework for tumor versus normal classification and
biomarker identification using gene expression and multi-omics TCGA data. The methodology employs Extreme Gradient Boosting
(XGBoost) for embedded feature selection to identify then most informative molecular variables from tens of thousandsof features. The selected features are subsequently used to train ensemble classifiers including Logistic Regression, Random
Forest, and Support Vector Machine models. To ensure unbiased performance estimation and prevent data leakage, a stratified five-fold cross-validation strategy is adopted. Experimental evaluation on breast and lung cancer datasets demonstrates strong discriminative performance, with the XGBoost–Random Forest model achieving mean classification accuracies exceeding 99%, along with high ROC-AUC and Cohen’s Kappa values. Furthermore, multi-omics integration improves classification robustness by capturing complementary molecular signals across biological layers. The results indicate that XGBoost-driven feature selection combined with ensemble learning provides a scalable, interpretable, and effective framework for high-dimensional cancer classification and biomarker discovery.