SlotDeconv: spatial transcriptomics deconvolution via diversity-constrained prototype learning and spatial refinement

Spatial transcriptomics (ST) measures gene expression in intact tissues. In spot-based ST assays, each spot can contain mixtures of multiple cell types. Deconvolution is particularly challenging when closely related cell subtypes share highly similar expression profiles and when spatial context is underutilized during proportion estimation.

We present SlotDeconv, a method for ST deconvolution consisting of a single-cell reference module and a spatial inference module. The reference module learns discriminative cell-type signatures using slot-based prototype vectors decoded into a reference matrix, trained with a negative binomial reconstruction loss and a max-margin diversity constraint that discourages similar cell-type signatures. Ablation studies confirm that both components are essential: removing the diversity constraint reduces spot-wise Pearson correlation by 41%, and replacing learned prototypes with cell-type mean expression reduces it to near zero. The spatial inference module initializes spot-level proportions via gene-weighted nonnegative least squares (NNLS), then refines them by minimizing Kullback-Leibler (KL) divergence between observed and reconstructed spot expression under a spatial neighborhood consistency regularizer. Benchmarked against CARD, RCTD, Cell2location, and Spotiphy on a 27 cell type mouse brain dataset, SlotDeconv achieves the highest spot wise Pearson correlation (approximately 0.56) and cosine similarity (0.633), outperforming competing methods in spot-wise correlation, with particularly strong gains on transcriptionally similar cortical neuronal subtypes. Biological validation on human pancreatic cancer and mouse olfactory bulb datasets further confirms spatial specificity.

Source code is available at https://github.com/HannahNJIT/SlotDeconv.