Fig. 4: O2-VAE representations identify meaningful biological subgroups of cells and nuclei.

a We train O2-VAE on mouse embryonic fibroblasts⁷ cultured on three micropatterned substrates: circular, triangular, or control (no micropattern); and two treatment groups: wild-type (LMNA^+/+) and lamin-deficient (LMNA^−/−). (Top chart) GMM clustering with k = 10: (top, left) prototypes of cluster centroid reconstructions; (top, middle) cluster samples; (top, right) heatmap of relative cluster frequencies per group. (Bottom) highlighting two group differences. Loss of LMNA (^−/−) correlates with reduced filopodia-like structures, a novel finding not identified by prior methods to the best of our knowledge. Low LMNA (^−/−) groups have lower prevalence of triangular classes that have sharp edges. b (Left) example cell+nucleus multistructure hiPSCs²³, and (right) UMAP of embedding space coloured by mitosis class. Mitosis cells separate from normal cells: this is an unsupervised detection method. c Interpolations of cell and multistructure hiPSCs are candidate shape deformations. Images at row edges are real cells. We sample points between their embeddings and reconstruct them to form the other images. d Fitting a GMM to learned representations enables detection of outliers for cells and nuclei in hiPSCs; they are likely bad segmentations or in mitosis and can be filtered in preprocessing. e Mitochondria in Allen Cell collection data clustered with GMM, k = 14. (Left) cluster prototypes, (middle) samples, and (right) prevalence, where `area' is the percent of mitochondrial area in that cluster and `frequency' is the percentage count of mitochondria objects in that cluster. f Example cell with pseudo-coloured segmentation masks with other organelles, that have many contacts (overlapping organelle pixels in white). g For hiPSC data, clustering with GMM and k = 14 (cluster samples and prevalence in Supplementary Fig. 27). Contact rates of each mitochondrial shape group with each organelle, which is the percentage of mitochondria in the group in contact with that organelle. `Aggregate' is the contact rate over all clusters. Supplementary Note 3c shows results per sub-experiment.