ARCADE disentangles cell abundance from continuous transcriptional state in spatial transcriptomics

Spatial transcriptomics is widely used to infer both cell-type abundance and within-type transcriptional variation, but these signals are difficult to separate in mixed spot-level measurements. Here, we present ARCADE, a computational framework that addresses two major sources of uncertainty in spatial analysis: unstable single-cell references and entanglement between abundance and cell state. ARCADE first builds a robust reference through hyperparameter optimization and consensus filtering, then performs spatial inference in two stages by estimating cell-type composition before modeling continuous intratype transcriptional programs around this compositional anchor. Across simulated and semi-synthetic benchmarks, ARCADE improves deconvolution accuracy and yields more stable state inference than existing approaches. In matched human dorsolateral prefrontal cortex datasets with spatial proteogenomic readouts, ARCADE distinguishes localized cell enrichment from astrocyte and microglial activation programs. These results show that decoupled modeling can improve the interpretability of spatial transcriptomic signals in complex tissues.