Fig. 3: Rotation-invariant point cloud representations recover the cell-cycle-dependent spatial pattern of DNA replication foci.
From: Interpretable representation learning for 3D multi-piece intracellular structures using point clouds
a, Dataset of DNA replication foci in hIPS cells expressing monomeric enhanced green fluorescent protein (mEGFP)-tagged PCNA. DNA replication foci have a stereotypical cell-cycle-dependent localization pattern. Shown are examples of image and sampled point cloud center slices with adjusted contrast for eight expert-annotated cell-cycle stages. b, Benchmarking unsupervised representations across different models and metrics. Left, polar plot showing performance of classical and rotation-invariant image and point cloud models across efficiency metrics (model size (n = 1), inference time (n = 40) and emissions (n = 40)), generative metrics (reconstruction (n = 122) and evolution energy (n = 180)) and representation expressivity metrics (compactness (n = 5), classification of cell cycle via top-2 classification accuracy (n = 5), rotation invariance error (n = 488) and average interpolate distance (n = 180)). Metrics are z-scored and scaled such that larger is better. Right, bar plots showing raw metric values across models for each metric. Error bars are standard deviations. The best model for each metric is indicated. c, Eight archetypes identified using rotation-invariant point cloud representations. Each archetype corresponds to one of the eight expert-annotated cell-cycle stages. d, PC1 for each cell-cycle stage using rotation-invariant point cloud model. PCA is fit to representations of each cell-cycle stage separately. Shown are normalized PCs (s.d./σ) sampled at three map points (−2σ to 2σ in steps of σ). e, Average canonical reconstructions across five bins of nuclear volume (Supplementary Note 5.3). All reconstructions shown are center slices.
