Intrinsic-dimension analysis for guiding dimensionality reduction and data fusion in multi-omics data processing

Jessica Gliozzo, Mauricio Soto Gomez, Valentina Guarino, Arturo Bonometti, Alberto Cabri, Emanuele Cavalleri, Justin T. Reese, Peter N. Robinson, Marco Mesiti, Giorgio Valentini, Elena Casiraghi

Abstract: Highlights•Unbiased id estimation: We address the overlooked problem of intrinsic dimensionality (id) in multi-omics data, proposing a principled approach for unbiased id estimation.•Principled dimensionality reduction: A tailored DR approach ensures robust multi-modal data characterization and effectively mitigates information loss.•View-specific DR pipeline: Our block-analysis framework customizes DR to each view with a novel two-step strategy for improved dimensionality reduction.•Comprehensive evaluation: Testing nine TCGA cancer datasets reveals the impact of proper DR on multi-omics integration and predictive performance.•Improving data fusion and predictions: Our DR pipeline enhances fusion algorithms and achieves better survival prediction accuracy using interpretable classifiers.