From bulk to single-cell and spatial data: An AI framework to characterise breast cancer metabolic dysregulations across modalities

Due to tumour heterogeneity, cellular metabolic activities and tumour proliferation rates can vary across patients, limiting the predictive and prognostic power of population-based metabolic biomarkers. Genome-scale metabolic models have been developed to address this challenge. These models can simulate patient-specific metabolic reaction rates and mechanistically elucidate cellular metabolic alterations. Simultaneously, phenotypes and other omics data can be integrated to provide valuable insights into specific tumour behaviour and molecular mechanisms underlying cancer biology. However, integrative approaches that combine multiple data modalities with metabolic modelling remain largely undeveloped. We propose a unified framework based on interpretable multi-modal machine learning, integrating different data modalities, including transcriptomics, clinical data, and genome-scale metabolic modelling, to stratify breast cancer patients. By integrating validation analysis across data scales, from bulk to single-cell and spatial data, our framework combines mechanistic knowledge from metabolic modelling with machine learning to characterise molecular and metabolic dysregulations across breast cancer risk groups. This unique mechanistic interpretation approach, combining data-driven and biologically-driven knowledge, provides a comprehensive understanding of breast cancer biology to improve personalised cancer therapy and reveal key biomarkers across multiple scales, including patient-specific, cell-specific, and spot-specific prognostic signatures.