Uncovering subtype-specific metabolic signatures in breast cancer through multimodal integration, attention-based deep learning, and self-organizing maps.

This study integrates multimodal metabolomic data from three platforms-LC-MS, GC-MS, and NMR-to systematically identify biomarkers distinguishing breast cancer subtypes. A feedforward attention-based deep learning model effectively selected 99 significant metabolites, outperforming traditional static methods in classification performance and biomarker consistency. By combining data from diverse platforms, the approach captured a comprehensive metabolic profile while maintaining biological relevance. Self-organizing map analysis revealed distinct metabolic signatures for each subtype, highlighting critical pathways. Group 1 (ER/PR-positive, HER2-negative) exhibited elevated serine, tyrosine, and 2-aminoadipic acid levels, indicating enhanced amino acid metabolism supporting nucleotide synthesis and redox balance. Group 3 (triple-negative breast cancer) displayed increased TCA cycle intermediates, such as α-ketoglutarate and malate, reflecting a metabolic shift toward energy production and biosynthesis to sustain aggressive proliferation. In Group 4 (HER2-enriched), elevated phosphatidylcholines and phosphatidylethanolamines suggested upregulated mono-unsaturated phospholipid biosynthesis. The study provides a framework for leveraging multimodal data integration, attention-based feature selection, and self-organizing map analysis to identify biologically meaningful biomarkers.