iSORT: An Integrative Method for Reconstructing Spatial Organization of Cells using Transfer Learning

Understanding the cellular spatial organization is a paramountly important direction of exploring the intricate functionalities of tissues and organs. However, conventional single-cell RNA sequencing (scRNA-seq) technologies often lose cells’ spatial information. Recent spatial transcriptomics (ST) techniques aim to preserve locations but easily come with either low resolution or high costs. Here, we introduce iSORT (an i ntegrative method for reconstructing S patial OR ganization of cells via T ransfer learning), which can recover the spatial organization of large-scale multi-source scRNA-seq dataset with only one ST reference. iSORT employs a transfer learning approach to assign coordinates for each single cell with scRNA-seq. iSORT also integrates multiple ST tissue slices as references to enhance accuracy and capture multi-slice biological heterogeneity. Furthermore, iSORT conduces to the precise identification of spatially variable genes (SVGs). We demonstrated the effectiveness of iSORT by applying it to four benchmark datasets, including those from Drosophila embryo, mouse embryo, mouse brain, and human brain tissues, and by a comparative analysis with five other algorithms as well. Additionally, we collected scRNA-seq data from the patients with atherosclerosis to reconstruct the structure of arteries for the first time, using ST data from other individuals as references. Atherosclerosis-related SVGs were identified and verified by the gene set enrichment analysis. As a promising tool, iSORT provides deep insights into the single-cell omics data analysis and biomedical research.