DANST enables cell-type deconvolution in spatial transcriptomics using deep domain adversarial neural networks

Spatial transcriptomics is an emerging technology that can analyze gene expression profiles of tissues while preserving spatial location information. To restore cell type proportions from mixed gene expression data, here we present DANST, a deconvolution framework based on deep domain adversarial neural networks. By integrating single-cell RNA sequencing (scRNA-seq) with inferred spatial coordinates, we construct pseudo-spatial data. DANST utilizes a variational autoencoder to learn refined feature representations and introduces a domain adversarial architecture to align feature distributions between pseudo and real data, enabling accurate label transfer. Benchmarking on human and mouse datasets shows that DANST achieves superior deconvolution accuracy compared with existing methods. These findings highlight its effectiveness for tumor microenvironment analysis and potential clinical utility. DANST uses deep domain adversarial neural networks to align single-cell and spatial transcriptomics, enabling accurate cell-type deconvolution and revealing complex cellular architectures in tumor microenvironments.