Chromatin dynamics upon serum starvation with CRISPR-Sirius.

a, Representative image of a nucleus in fixed sample with chromosome loci visualized using CRISPR-Sirius. Minimal distance to edge indicated in white boxes, microns. Scale bar 6 microns. b, Kernel density estimate distributions showing the minimal distance to the nuclear edge for different chromosome loci in serum (10%) or after 1h serum starvation (0.3%). c, Fractions of chromosome 1 loci detected in shells of equal volume derived from the division of the nuclear area by sequential morphological erosion. d, e Representative images of U2OS cells expressing CRISPR-Sirius plasmids with the gRNA targeting chromosome 1 (d) and telomeres (e) in green, and LaminA-mCherry for nucleus segmentation in magenta. f, Minimal distance to edge of visualized loci in serum (10%) and under starvation (0.3%). Each dot in a swarmplot, upper panel, represents individual locus. Lower panel represents the distributions of bootstrapped delta mean values for datasets. Error bars show 95% confidence intervals in both upper and lower panels. Statistical significance with Mann-Whitney U rank test. g, Chromosome 1 loci are observed to localize closer to the nuclear periphery under starvation, also when measured in the same cells before and after (within 40 min) medium change. Every line represents the change in minimal distance to the nuclear periphery for a particular locus; left scale. The distribution shows bootstrapped delta mean values for paired dataset; right scale. Error bars show 95% confidence intervals. The difference is significant, paired two-tailed t-test, normality Shapiro-Wilk (α=0.05), p<0.01.

Features extracted from the imaging data collected in this study.

Machine learning-assisted dataset exploration.

a, Information flow from raw pixel intensities to inference and explanations. (b-d), Decrease of accuracy estimation variance in repeated cross-validation of gradient boosting classifier (GBC) (b), multilayer perceptron (MLP) (c) and InceptionTime (d). Shaded area shows 95% confidence interval. e, Hyperparameter tuning for GBC. f, The accuracy of multilayer perceptron converges after approximately 50 epochs of training at validation accuracy of 0.64; g, InceptionTime demonstrated performance comparable to GBC, but without access to manually engineered features. (f,g) Shaded area shows 95% confidence interval.

Explaining machine learning models’ decision using Shapley additive explanations (SHAP).

a,b Global summary of SHAP showing feature impacts towards one of the classes, high serum or starvation, for gradient boosting classifier (GBC) (a) and multilayer perceptron (MLP) (b). Feature values are color-mapped from blue (low) to red (high). c,d Heatmap of feature interactions depicting the components of synergistic effect of features, after accounting for the direct impact. Interactions from GBC seem to be more uniform and show less usage of binary and discrete features (c), compared to the interactions from MLP (d). e, Change in feature importance rank with the growth of model complexity from the baseline single level decision tree (designated as Tree) to GBC and MLP. Binary and discrete features are not shown for the Tree set, as the model is trained on a single feature at a time to estimate baseline performance, thus, feature interactions are not used by definition. f, Global summary of SHAP showing feature impacts towards class serum or starvation, InceptionTime.

Changes in nuclear morphology upon serum starvation.

a, Nuclear area (MA) and its dynamics (sA) in high serum medium and under starvation. The marginal distributions show corresponding kernel density estimations. b-e, Changes in nuclear morphology, particularly, area (b) area dynamics (c,d) and circularity. Statistical significance with Mann-Whitney U rank test. (e). Each dot in the swarmplot represents an individual nucleus. Panels on the right represent the distributions of bootstrapped delta mean values for datasets. Error bars show 95% confidence intervals. d, Impact of nuclear area dynamics (as SHAP values, shown in color) on model decision depending on nuclear area and dynamics of its change. f, Locus speed affects Shapley values for nuclear area, which is clear from vertical color gradient, where the locus speed is color-mapped, see legend on the right. MLP SHAP explanation; outliers (less than 2% of data) are not shown for clarity.

Specific features of chromosome 1 dynamics.

a-d, Changes in loci motility as measured by mean displacement (MD) and total distance traveled (TD). Mean loci displacement did not change for chromosome 1 under starvation (a), but showed a decrease in telomeres (b). Total distance traveled by a locus for chromosome 1 grew insignificantly, in a very similar manner to its mean displacement (c), while for telomeres stays stable despite the decrease in speed (d). Each dot in the swarmplot represents an individual nucleus. Panels on the bottom represent the distributions of bootstrapped delta mean values for datasets. Error bars show 95% confidence intervals. e-g, SHAP analysis of features describing persistence and directionality of loci movement. e, dynamics of the change of nuclear area (sA) affects the persistance (Pers) SHAP values. f, locus mean displacement (MD) affects the SHAP values calculated for total traveled distance (TD). g, Persistence (Pers) is negatively correlated with SHAP values of feature which detects long spontaneous displacements (out3sd). h, Long spontaneous displacements (out3sd = 1) of chromosome 1 loci are more likely to occur in short distances from nuclear periphery under serum starvation.

Distinct dynamic behavior of chromosome 1 homologs upon serum starvation.

a-d, Chromosome 1 radial organization as explored using time series data analysis. a, Nuclear area (MA) and distance to nuclear edge (MDist) in high serum medium and under starvation. The marginal distributions show corresponding kernel density estimations. b,c Timelapse-derived data confirm chromatin radial reorganization under serum starvation. The dynamic range of radial movement changed insignificantly for chromosome 1, but showed substantial growth in telomeres (b). Total length of radial relocalization (TDist), showing not only scale but also the direction of relocalization, decreased in chromosome 1 but had only insignificant change in telomeres (c) Statistical significance (b,c) with Mann-Whitney U rank test. d, Total distance traveled by locus (TD) affects distance to edge (MDist) SHAP values. e-i, Differential behavior of Chromosome 1 homologs. e-g, Correlation of mean loci displacement, MD (e), total distance traveled by locus, TD (f), and total radial displacement, TDist (g) between chromosome 1 homologs. h, Relationship between categorical features subgrouping chromosome 1 homologs based on speed (ifFast) and radial position (ifCentr). i, Representative image of chromosome 1 loci (in yellow), where one of the homologs is associated with a nucleoli-like subcompartment. The LaminA channel is shown in cyan.

Mean locus distance to nuclear periphery for chromosome 1 loci under starvation, measured in the same cell before and after medium change.

a, Raw data plotted without time-related subsampling, each line represents a nucleus. b, Cells sampled within the first 20 min after medium change). Every line represents the change in minimal distance in particular locus, left panel. The distribution shows bootstrapped delta mean values for paired dataset, right scale. Error bars show 95% confidence intervals. c, Representative image of U2OS cells overexpressing Sirius plasmids with guide for chromosome 1 locus, in 10% serum and under starvation. The descriptive statistics for cells imaged before and after the medium change is shown in Figure 1h. Scale bar = 6μm.

Changes in nuclear morphology under starvation, particularly, area (MA, a) and perimeter (MP, b).

Each dot in the swarmplot represents an individual nucleus. Panels on the right represent the distributions of bootstrapped delta mean values for datasets. Error bars show 95% confidence intervals. c, Correlation between nuclear area (MA) and the length of perimeter (MP) in 10% FBS and in starvation (0.3% FBS). Shaded area shows 95% confidence interval. d-f, Locus speed as measured by mean locus displacement (MD) affects Shapley values for nuclear area (MA), which is clear from vertical color gradient on SHAP dependence plot, where the locus speed is color-mapped, see the legend on the right. SHAP was used to explain the decisions of MLP (d, repeated from Fig 4f for comparison), GBC (e), and InceptionTime (f) classifiers.