Unlocking precision diagnostics: A multimodal framework integrating metabolomics with advanced machine learning techniques.

Abstract

Integrating multiple omics modalities is a crucial strategy in cancer research, particularly in metabolomics, enabling early detection and detailed exploration of cancer biomarker signatures. This study evaluates five strategies for integrating metabolomics data from liquid chromatography-mass spectrometry, gas chromatography-mass spectrometry, and nuclear magnetic resonance. Deep Transfer Learning and Multiple Kernel Learning demonstrated superior performance, significantly improving classification accuracy, sensitivity, and robustness compared to single-modality analyses. Deep Transfer Learning employed a custom autoencoder for feature extraction followed by artificial neural network classification, while Multiple Kernel Learning optimized kernel matrices across different modalities. Feature extraction in the Deep Transfer Learning approach, combined with the selection of important features and subsequent analysis, revealed elevated levels of monounsaturated phospholipids such as phosphatidylcholine 30:1, phosphatidylethanolamine 32:1, and sphingomyelin 32:1 in HER2-positive cases. Additionally, β-alanine, gluconic acid, and N-acetylaspartic acid were increased, whereas 5'-deoxy-5'-methylthioadenosine and nicotinamide were decreased. These methods advance cancer detection, biomarker discovery, and the development of precise diagnostic and therapeutic tools while offering robust and adaptable strategies for multi-omics data integration across diverse biological datasets.