Extended Data Fig. 8: Clustering results for a human breast cancer spatial gene and protein expression dataset.
From: Resolving tissue complexity by multimodal spatial omics modeling with MISO
Image features used as input for each method were extracted from an immunofluorescence image with 3 channels (DAPI, Vimentin, and PCNA) using a pre-trained InceptionV3 model. Patches corresponding to the omics spots were extracted from the immunofluorescence image, resized to 299×299 pixels, and normalized prior to extracting features for each patch using InceptionV3. a, Clustering results from MISO, MUSE, and SpatialGlue when taking RNA and image data as input. b, Clustering results from MISO, MUSE, and SpatialGlue when taking RNA and protein data as input. c, Clustering results from MISO, MUSE, and SpatialGlue when taking protein and image data as input. d, Clustering results from MISO when taking RNA, image, and protein data as input. e, Eosin-stained histology image of analyzed tissue section. f, RNA, image, and protein ICC distributions across all clusters and features for each method when taking each possible combination of modalities as input (n = 500 ICC values for each group). For each method, the mean ICC for each modality is printed on the corresponding box plot for its top-performing combination of modalities. Test statistics and p-values were obtained using one-sided (<,>) or two-sided (≈) t-tests when comparing each method’s top-performing results for a given modality. Box plots: center line, median; box limits, upper and lower quartiles; whiskers, 1.5x interquartile range; points, outliers.
