Fig. 2: Thor accurately predicts single-cell spatial gene expression in human breast cancer.

a Spatial gene expression of in silico cells inferred from the Visium data and the H&E staining image of a breast cancer tissue by Thor align closely with Xenium data from the adjacent tissue section. The numbers on the H&E staining image mark DCIS regions of interest. b Thor-inferred spatial transcriptome of in silico cells demonstrates consistent cell clusters with Xenium using scRNA-seq clustering. The cluster annotations were adapted from the original study of the dataset²⁸. The mean expression levels (normalized) of differentially expressed genes in each cluster were visualized using heatmaps. c Thor outperforms other methods in spatial gene expression prediction. The box plots summarize the similarity across 306 genes included in the Xenium panel. The middle line in the box plot, median; box boundary, interquartile range; whiskers, 5–95 percentile; minimum and maximum, not indicated in the box plot. One-sided Mann–Whitney U tests are used to compare Thor with the two next-best-performing tools; corresponding p-values are shown in the plot. d Spatial expression profiles of representative genes at the region of interest level are compared between Thor, iStar, and Xenium. Thor-inferred spatial gene expression closely aligns with the Xenium data, while iStar introduces artifacts at segment boundaries (the red arrows) and in regions with sparse cells (the blue arrow). Source data are provided in a Source Data file.