Feature selection and dimensionality reduction in multi-omics

High-throughput multi-omics technologies generate high-dimensional, heterogeneous and noisy datasets, making robust biomarker discovery and disease modelling challenging. This review critically examines dimensionality reduction and representation strategies for multi-omics analysis while distinguishing three related but different goals: feature selection, feature extraction/representation learning, and multi-omics integration. We use sparse canonical correlation analysis (sCCA), variational autoencoders (VAEs), and graph convolutional networks (GCNs) as representative case studies spanning sparse linear, probabilistic deep-learning, and graph-based approaches. Rather than treating these methods as interchangeable, we compare their data requirements, assumptions, interpretability, ability to accommodate non-linearity and missing modalities, computational demands, and validation needs. Applications across cancer, cardiovascular and neurodegenerative disease, and across bulk, single-cell, spatial and imaging multi-omics are summarized in comparative tables. We conclude with practical guidance for matching method families to biological questions and emphasize that robust biomarker claims require leakage-free model development, stability assessment and independent validation.