Fig. 1: CNN models predict developmental time from nuclear histone images of live embryos.

a Still frames from time-lapse images of a his2av-mrfp1 embryo during nc10–14 (top) and within nc12 (bottom). Scale bar, 50 μm. b Schematics of the time prediction workflow (top) and model architecture (bottom). Scale bars, 50 μm (left) and 5 μm (right). c Grad-CAM visualization of the attention from three independent models. Left: representative histone images from an nc13 embryo at different spatial scales as inputs for each model. Right: Grad-CAM maps (green) from the final convolutional layer in each model, overlaid on the original images. Scale bar, 5 μm. Model 1 focuses on the intra-nuclear region, implying that features such as nuclear texture and size are key to its predictions. Models 2 and 3 attend to inter-nuclear regions surrounding multiple nuclei, implying that broader-scale features, such as nuclear density, play an important role. d Histograms of residuals from individual models (left) and the overall time predictor (right) in nc13. e Evaluation of developmental time prediction across multiple nuclear cycles. The predicted time for each time point of each embryo was compared to ground truth. Insets: histograms of residuals from the overall time predictor. n = 4, 6, 4, and 3 embryos for nc11–14, respectively. Source data are provided as a Source Data file.