Fig. 2: Intral- and cross-modality detection of anomalous tissue domains (ATDs) in single 10x Visium datasets.

a Identification of ATDs in a 10x Visium human breast cancer dataset. True anomalous spots and those identified by the tested methods are both indicated in blue in the ground truth and respective method panels. b Performances of the tested methods are quantified using accuracies and F1-scores, PR curves, SGD_degree and SGD_cc scores. Accuracies and F1-scores are calculated in 5 independent experiments and presented as mean ± SD. c Top scatter plot maps each spot’s distance to the nearest carcinogenic region (x-axis) against its anomaly score (y-axis). Spots in the green circle are true anomalies. Yellow spots denote STANDS-identified anomalies. Red line is a regression line representing the correlation between anomaly scores and distances. Middle spatial map shows the spatial distribution of anomaly scores, with a color gradient from blue to red indicating lower to higher scores. Bottom histogram illustrates the probability density distributions of normalized anomaly scores in the reference and target datasets. d STANDS pinpoints emerging cancers adjacent to known cancerous regions. Spots within cancerous regions are in red, while those within normal tissue regions are in cyan. Notably, normal spots in orange, which are adjacent to cancerous regions and identified as anomalies, potentially represent developing cancer. e Expression levels of breast cancer marker genes ACTB and TMSB10 within the normal (n = 238), emerging cancerous (n = 69), and known cancerous regions (n = 160). In the boxplot, the center line denotes the median, box limits denote the upper and lower quartiles, and whiskers denote the 1.5× interquartile range. f Cross-modality identification of ATDs within a 10x Visium pancreatic ductal adenocarcinomas dataset, using a healthy pancreatic duct scRNA-seq dataset as reference. True anomalous spots and those identified by the tested methods are both indicated in green in the ground truth and respective method panels. g Performances of the tested methods are quantified using accuracy and F1-scores, PR curve, and a scatterplot of SGD_degree vs SGD_cc scores. Accuracies and F1-scores are also calculated in 5 independent experiments and presented as mean ± SD. Source data are provided in this paper.