Abstract
Sequencing-based spatial transcriptomics technologies have revolutionized our understanding of complex biological systems by enabling transcriptome profiling while preserving spatial context. However, spot-level expression measurements often amalgamate signals from diverse cells, obscuring potential heterogeneity. Existing methods aim to deconvolute spatial transcriptomics data into cell type proportions for each spot using single-cell RNA sequencing references but overlook cell-type-specific gene expression, essential for uncovering intra-type heterogeneity. We present PANDA (ProbAbilistic-based decoNvolution with spot-aDaptive cell type signAtures), a novel method that concurrently deciphers spot-level gene expression into both cell type proportions and cell-type-specific gene expression. PANDA integrates archetypal analysis to capture within-cell-type heterogeneity and dynamically learns cell type signatures for each spot during deconvolution. Simulations demonstrate PANDA's superior performance. Applied to real spatial transcriptomics data from diverse tissues, including tumor, brain, and developing heart, PANDA reconstructs spatial structures and reveals subtle transcriptional variations within specific cell types, offering a comprehensive understanding of tissue dynamics.