A scalable workflow for in silico TMA construction and groupwise comparison of spatial transcriptomic data for analyzing tumor microenvironment

Background: Spatial transcriptomics provides high-resolution characterization of the tumor microenvironment (TME), offering insights into cellular architecture, spatial interactions, and molecular heterogeneity. However, integrating data across multiple tissue sections or multiple samples from tissue microarrays (TMAs) remains a substantial challenge, particularly for bioinformatics workflows requiring scalability and flexible cohort design. The difficulty is amplified in high-density platforms such as Visium HD, where datasets contain large numbers of cells and preserving spatial relationships during multi-slide integration is critical. Consequently, there is a growing need for a scalable computational framework that enables efficient in silico TMA construction and supports groupwise comparisons for robust TME analysis. Methods: We developed a flexible and modular computational workflow enabling in silico TMA construction and groupwise comparison of gastric cancer Visium HD data. From a total of 48 samples (2 mm core), yielding over 3.1 million 8-µm bins of spatially resolved expression data, was used for the integration and scalable group-wise comparison. Selected cores from different slides were digitally reassembled into new TMA-like layouts with automatically re-registered H&E images reflecting updated spatial coordinates. For batch correction and large-scale integration, the workflow employs stratified subsampling and the scArches model, allowing efficient harmonization of datasets containing millions of spatial bins. Interactive visualization modules support region-level comparison and cross-core analysis of cellular composition. Results: Preliminary groupwise analyses of the reconstructed TMA revealed reproducible spatial patterns across tumor regions. Areas with dense stromal activation exhibited enrichment of fibroblast and endothelial-related signatures, whereas immune hotspots showed localized CD8⁺ T-cell and macrophage co-accumulation near the tumor invasive front. Gene expression-based metabolic pathway analysis revealed subtle gradients (glycolytic vs oxidative) between core groups with distinct histopathologic features. The entire integration pipeline processed 3.1 million spatial bins in under 2hours on a single NVIDIA A6000 GPU(48GB VRAM), demonstrating the framework's potential to detect spatially resolved tumor-stroma interactions and immune heterogeneity at scale. Conclusions: This workflow provides a scalable approach for constructing virtual TMAs and comparing spatial transcriptomic data. By integrating >3 million Visium HD bins from gastric cancer slides, it enables systematic TME profiling and identification of spatially resolved biomarkers relevant to tumor progression and therapeutic response.