Systematic evaluation of spatial transcriptomic annotation methods reveals conserved tumor microenvironment programs...

Spatial transcriptomics (ST) enables high-resolution mapping of cellular heterogeneity in tumor microenvironments, but downstream cell-type annotation remains highly method-dependent.

Here, we systematically evaluated four representative annotation frameworks-Seurat, CARD, RCTD, and SPOTlight-across 45 non-small cell lung cancer (NSCLC) ST samples from 10×Visium comprising 328,561 spots. We performed a comprehensive cross-method assessment across multiple analytical dimensions, including cell-type composition, inter-method agreement, spatial architecture, cell-cell interaction patterns, and functional pathway activity.

While all methods captured broad tissue organization, substantial variability emerged in inferred cellular composition, spatial clustering, and functional states. CARD and RCTD demonstrated the highest overall concordance across abundance estimation and spatial structure, whereas Seurat and SPOTlight displayed greater method-specific divergence. Functional analyses further revealed method-specific biological programs, with CARD and RCTD better capturing stromal and myeloid identity. Importantly, beyond these differences, cross-method consensus analysis identified robust spatially conserved biological programs across multiple layers, including gene-level signatures, transcription factor activity, and pathway-level reprogramming characterized by enhanced metabolic and translational activity in tumor-proximal regions. These transcriptional signatures were further supported by independent single-cell RNA sequencing datasets, suggesting their biological relevance.

Overall, our study demonstrates that while no single annotation method is universally optimal, multi-method integration enables the identification of robust multi-layer spatial regulatory programs and provides a more reliable framework for interpreting ST data in complex tumor ecosystems.