Spatial transcriptomics technologies profile gene expression across tissue sections while retaining spatial information, yet most platforms capture signals from multiple cells per measurement location, requiring computational methods to determine the underlying cellular composition. Current deconvolution approaches either depend on single-cell reference atlases-which may be unavailable for rare tissues or specific disease contexts-or employ unsupervised methods that struggle with complex spatial patterns and typically need manual specification of how many cell types to identify. We developed Attention-Guided Enhanced Deconvolution (AGED), a two-stage framework combining probabilistic modeling with neural attention architectures for reference-free analysis. The first stage uses a Performer-based network with linear-complexity attention to systematically evaluate models with different numbers of cell types, automatically selecting the optimal configuration through composite scoring of reconstruction quality, cluster separation, and topic diversity. The second stage, Attention-Guide, stars with hierarchical probabilistic initialization. It progressively refines cell type features through multiple attention mechanisms: cross-attention based on statistical priors, spatial attention aggregating neighborhood information, and collaborative attention capturing theme-gene associations. A dynamic gating mechanism enables the model to balance global statistical patterns with locally data-driven features, while regularization promotes sparse solutions consistent with biological reality-each position containing only a few dominant cell types. Testing on Mouse Olfactory Bulb (MOB) tissue, AGED automatically identified four anatomical structures and achieved superior reconstruction performance (r=0.86) with characteristic sparsity patterns. When similarly applied to human pancreatic ductal adenocarcinoma (PDAC) and human thymus tissue, it revealed detailed relationships between dissected structures and cell types. The learned cell type distributions aligned well with established neuroanatomical boundaries and known molecular markers, indicating the attention-based refinement maintains biological interpretability throughout training. This framework offers a practical solution for spatial transcriptomics analysis across diverse experimental systems without requiring matched single-cell data.
Yang et al. (Tue,) studied this question.