Abstract
Deconvolution of spatial transcriptomics data returns, for every spot, a full vector of cell type proportions, yet standard spatial domain identification keeps only the dominant cell type assignment (argmax) and discards the remaining composition signal. We present stPularity, a post-deconvolution clustering tool that operates directly on the proportion vectors a deconvolution pipeline already produces: it builds a k-nearest neighbor graph on spatially smoothed cell type proportions and applies a modified Louvain procedure whose objective trades off modularity against a normalized-entropy purity term, automatically determining the number of clusters without ground-truth labels. Its few hyperparameters are set automatically by Bayesian optimization over label-free internal criteria, so the pipeline runs without manual resolution selection. Because stPularity reads only the proportion matrix and spatial coordinates, it is agnostic to the deconvolution method and operates in the low dimensionality of the cell type panel (4–38 cell types). Across four benchmarks, stPularity improves over the dominant-cell-type baseline on three of four datasets, raising the adjusted Rand index (ARI) over argmax by 0.14–0.26. On three of the four datasets, this advantage over argmax persists or widens as the deconvolution proportions are corrupted with increasing noise. An ablation shows that proportion vectors and gene-expression principal component analysis (PCA) embeddings are complementary rather than redundant, their relative strength depending on whether spatial structure is compositional or transcriptomic. Expression-based spatial methods attain higher absolute ARI by reprocessing the full transcriptome; stPularity, instead, targets a distinct goal: recovering interpretable spatial structure from the composition vectors that the argmax workflow would otherwise discard.