Abstract
Non-small cell lung carcinoma (NSCLC) exhibits complex intratumoral microenvironment heterogeneity that severely impacts clinical outcomes and therapeutic responses. Although single-cell RNA sequencing has illuminated transcriptomic cell states, it forfeits crucial spatial context. Conversely, sequencing-based spatial transcriptomics preserves histological architecture but often suffers from cellular mixing within multi-cell capture spots. In this study, we present an unsupervised Bayesian probabilistic topic modeling framework tailored for spatial transcriptomic deconvolution in NSCLC tissue sections. By treating spatial capture spots as documents and gene expression counts as word tokens, our model decomposes spot-level profiles into latent transcriptomic 'topics' representing discrete cell-type states and functional microenvironmental niches without requiring matched single-cell reference panels. Applied to 10x Visium spatial transcriptomic profiles from primary NSCLC resections, our method resolved five primary spatial topics: malignant epithelial core, tertiary lymphoid structure (TLS)-rich, desmoplastic stroma, myeloid-suppressive interface, and hypoxic/vascular niches. Quantitative benchmarking demonstrated superior deconvolution accuracy and robustness against noise compared to classical non-negative matrix factorization and reference-based algorithms. Crucially, spatial proportions of the myeloid-suppressive and TLS-associated topics strongly stratified progression-free survival in independent clinical cohorts. This probabilistic framework offers a robust, reference-free strategy for mapping spatial microarchitecture, providing pivotal insights into NSCLC tumor biology and immune evasion.
