KGLAR: Deconvoluting Spatial Transcriptomics Data with Single-cell Transcriptomes through Knowledge-guided NMF and Least Angle Regression.

The advance in spatial transcriptomics (ST) technologies has brought great opportunities to unlock cellular heterogeneity in cancers. However, most of current spatially resolved transcriptomic technologies provide the mean gene expression of spots and do not still have single-cell resolution in breadth and depth. The deconvolution technique can efficiently integrate gene expression and spatial location to annotate cell types. Here, we report KGLAR, a novel deconvolution framework for cell type annotation by integrating single-cell RNA sequencing and ST data. KGLAR incorporates topic identification through knowledge-guided non-negative matrix factorization, topic distribution analysis with least angle regression, cell type annotation based on non-negative least squares, and label refinement by combining spatial information. Using 4 different evaluation metrics, KGLAR is compared with other 10 state-of-the-art deconvolution methods on 2 gridded and 17 pseudo-ST datasets. KGLAR surpassed the 10 cell-type deconvolution methods under different sequencing depths, different sequencing platforms and different tissues within a relatively short period of time. Furthermore, KGLAR is capable of precisely mapping subtle cortical layers and the hippocampus architectures in mouse brain, and accurately annotating tumor-specific immune cell types with different spatial distributions in the pancreatic ductal adenocarcinoma tissue. KGLAR provides a publicly available R package and facilitates to dissect spatial and cellular heterogeneity.