Journal: bioRxiv

Article Title: HEDeST: An Integrative Approach to Enhance Spatial Transcriptomic Deconvolution with Histology

doi: 10.64898/2026.01.06.697922

Figure Lengend Snippet: Overview of the HEDeST framework. A Schematic overview of the HEDeST workflow. The method integrates histology-derived morphological information with declevel proportions as supervision. In parallel, this step enabled onvolution results from ST to predict single-cell types across the tissue. HEDeST outputs cell-level annotations and enables downstream analyses such as spatial co-localization, morphological profiling, and neighborhood composition studies. B Inputs to the model. Nuclei are first segmented from the H&E image, and cell-centered patches are extracted to compute morphological embeddings. In parallel, spot-level cell-type proportions are obtained from a deconvolution of the ST data. Each cell within a spot is weakly labeled by the corresponding proportion vector. C Training and inference strategy. A classifier is trained under a LLP paradigm to align aggregated cell-level predictions with spot-level proportions. After training, the model predicts cell-type probabilities for all nuclei across the slide, including those outside Visium spots. A PPSA step then refines predictions by incorporating local priors. The architecture of the classifier is presented in Supp. Fig. 1 .

Article Snippet: Here, we present HEDeST (Histology-based Enhancement of Deconvolution for Spatial Transcriptomics), a weakly supervised framework that jointly leverages deconvolution-derived ST data and morphological features from histology to predict cell types at single-cell resolution.

Techniques: Derivative Assay, Labeling, Plasmid Preparation

Journal: bioRxiv

Article Title: HEDeST: An Integrative Approach to Enhance Spatial Transcriptomic Deconvolution with Histology

doi: 10.64898/2026.01.06.697922

Figure Lengend Snippet: Spatial organization analysis of a real breast cancer sample using HEDeST. CAFs, Cancer-Associated Fibroblasts; PVL, Perivascular-like cells. A Output of cell-type deconvolution using DestVI after preprocessing. B Spatial map of predicted single-cell types obtained with HEDeST, revealing fine-grained tissue organization. C Representative mosaics of predicted cell types, showing characteristic morphologies consistent with known cellular phenotypes. D Distribution of nuclear areas across predicted cell types (amongst a total of 45,237 cells). E Spot-level co-localization matrix showing Pearson correlations between spot-wise cell-type proportions. F Mean inter-cell-type distance matrix derived from neighboring cells identified through Delaunay triangulation. G Cell-level co-localization graph summarizing neighborhood composition; arrow width indicates the average frequency of one cell type among another’s neighbors, highlighting strong co-localization among immune populations. H Neighborhood composition clustering. Cells were grouped based on the composition of their local neighborhoods, revealing biologically interpretable clusters. The resulting clusters correspond to coherent and biologically interpretable tissue structures, consistent with observable histological organization.

Article Snippet: Here, we present HEDeST (Histology-based Enhancement of Deconvolution for Spatial Transcriptomics), a weakly supervised framework that jointly leverages deconvolution-derived ST data and morphological features from histology to predict cell types at single-cell resolution.

Techniques: Derivative Assay