TG-ME-Met: An Integrative Transformer-Graph Variational Autoencoder Framework for Multi-Modal Characterization of Tumor Microenvironments through Spatial Transcriptomics and Metabolomics

The tumor microenvironment (TME) is a complex ecosystem of malignant, stromal, and immune cells whose spatial organization influences tumor progression, therapeutic response, and clinical outcomes. Although spatial transcriptomics has enabled detailed characterization of cellular heterogeneity, integrating additional molecular layers is essential for a more comprehensive understanding of tumor biology. Here, we present TG-ME-Met, an extension of the Transformer-Graph Microenvironment Explorer (TG-ME), a multi-modal computational framework that combines transformer architectures, Graph Variational Autoencoders (GraphVAE), and metabolomics integration to characterize tumor niches. TG-ME-Met integrates spatial transcriptomics, morphological imaging, metabolomics, multi-omics datasets, molecular interaction networks, and other biological data sources.

The framework employs a multi-stage pipeline including data normalization, spatial integration, morphological feature extraction, gene expression quantification, metabolite profiling, single-cell characterization, and tumor niche identification. The transformer module captures complex molecular interactions, while the GraphVAE incorporates spatial neighborhood information to identify biologically meaningful microenvironmental structures. Application of TG-ME to high-resolution non-small cell lung cancer (NSCLC) datasets revealed distinct microenvironments associated with disease severity and progression. The framework identified multiple biologically relevant niches, including hypoxia-, epithelial-mesenchymal transition (EMT)-, apoptosis-, interferon alpha-, interferon gamma-, TNF-alpha-, and E2F-associated tumor regions. These niches exhibited molecular signatures linked to tumor aggressiveness, immune regulation, metastasis, and therapeutic resistance. In addition, tumor-stroma interfaces showed enrichment of inflammatory, immune, coagulation, and MAPK-related pathways.

By integrating metabolomics with spatially resolved transcriptomic and imaging data, TG-ME-Met provides a systems-level understanding of tumor ecosystem organization and function. This approach enables the identification of molecular and metabolic programs associated with specific tumor niches and disease progression. Future applications will focus on uncovering metabolomics-driven microenvironments and biomarkers associated with prognosis and treatment response, advancing precision oncology through comprehensive multi-modal characterization of the TME.