Benchmarking Cell-Type-Specific Spatially Variable Gene Detection Methods Using a Realistic and Decomposable Simulation Framework

Identifying spatially variable genes within individual cell types is essential for characterizing spatially organized cell states and microenvironments from spatial transcriptomics data. Several computational methods have been developed for identifying cell-type-specific spatially variable genes (ctSVGs), but their relative performance and practical utility under realistic biological complexity remain largely unknown. To address this gap, we present the first systematic bench-mark study of all five existing ctSVG detection methods--CELINA, STANCE, C-SIDE, CTSV and spVC--using an integrated evaluation framework that combines idealized simulations, Xenium-based realistic simulations, and a decomposition-based diagnostic analysis. We compared the methods in terms of detection accuracy, scalability, and usability. Across realistic datasets generated on various tissue types, all methods experienced sharp declines in detection accuracy and substantial inflation of false discoveries compared to idealized simulations. To explain this failure, we developed a new simulation framework that decomposes the "realness" of the realistic simulation into interpretable biological and technical components, enabling us to attribute method-specific performance losses to specific components, including realistic diversity of cell types, heterogeneous cell layouts, null gene distributions, capture efficiency and realistic intra-cell-type spatial patterns. Together, our results show that no single method dominates across detection accuracy, scalability and usability, and we further clarify why current ctSVG methods fall short in realistic settings. We summarize these tradeoffs into a practical user guide to support method selection and highlight key challenges in developing robust, scalable ctSVG detection tools for real spatial transcriptomics data.