ST-deconv: an accurate deconvolution approach for spatial transcriptome data utilizing self-encoding and contrastive learning

Single-cell RNA sequencing (scRNA-seq) has significantly deepened our understanding of cellular heterogeneity and cell type interactions, providing insights into how cell populations adapt to environmental variability. However, its lack of spatial context limits intercellular analysis. Similarly, existing spatial transcriptomics (ST) data often lack single-cell resolution, restricting cellular mapping. To address these limitations, we introduce ST-deconv, a deep learning-based deconvolution model that integrates spatial information. ST-deconv leverages contrastive learning to enhance the spatial representation of adjacent spots, improving spatial relationship inference. It also employs domain-adversarial networks to improve generalization and deconvolution across diverse datasets. Moreover, ST-deconv can generate large-scale, high-resolution spatial transcriptomic data with cell type labels from single-cell input, facilitating the learning of spatial cell type composition. In benchmarking experiments, ST-deconv outperforms traditional methods, reducing the root mean square error (RMSE) by 13% to 60%, with an RMSE as low as 0.03 for high spatial correlation datasets and 0.07 for low spatial correlation datasets across different transcriptomic contexts. Reconstructing real tissue structure, a purity of 0.68 on mouse olfactory bulb (MOB) and a cell type correlation of 0.76 on human pancreatic ductal adenocarcinoma (PDAC) were achieved. These advancements make ST-deconv a powerful tool for enhancing spatial transcriptomics and downstream analyses of intercellular interactions.