VCS MN neural net module identifies micronucleated cells.

(A) Diagrams of constructs transduced into RFP703/Dendra cells and how these constructs localize in micronucleated cells before and following MN rupture. (B) Image of H2B-emiRFP703 in RFP703/Dendra cells at 20x and overlaid with MN+ and MN-cell classification results. Bottom: visual depiction of classification pipeline. (C) Recall and positive predictive value (PPV) of MN- and MN+ cells. N=2, n=328, 186. (D) Recall and PPV of MN classification within image crops. MN were manually scored as intact or ruptured by Dendra2 signal. N=5, n=264, 158, 365, 283, 249. (E) Same as for (d), except analyzed images are of U2OS cells expressing H2B-emiRFP703 and NLS-3xDendra2. N = 1, n = 85, 95. (F) RFP703/Dendra cells with ruptured (arrows) and intact (arrow-head) MN showing a loss of NLS-3xDendra signal in ruptured MN. Scale bar = 10 µm. (G) Recall and PPV for rupture-based cell classification. N=3, n=120, 91, 82.

MNFinder module robustly segments MN across cell types and imaging conditions.

(A) Representative image showing undersegmentation by the VCS MN neural net. (B) Overview of MNFinder module for classifying and segmenting MN and nuclei. Images are tiled by a sliding window and processed by 2 neural nets in parallel: one for classifying regions as nuclei or MN (Nuc/MN) and one for classifying cells. Nuc/MN classifier results are post-processed to correct MN that were misclassified as small nuclei and to expand MN masks. Cell classifier gradient map outputs are used to define cells through watershed-based post-processing. Nuc/MN results are then integrated with cell results to produce final labels of individual MN, cells, and MN. Image crops are reassembled onto the final image by linear blending. C-D) Example images and MN pixel predictions using MNFinder on multiple cell types (RPE1 H2B-emiRFP703, U2OS, HeLa H2B-GFP, and HFF), chromatin labels (DAPI, Hoechst, H2B-FP), and magnifications (20x, 40x). MN recall, PPV, and mean intersection-over-union (mIOU) per object were quantified across conditions. Dotted line = performance of the VCS neural net on RPE1 H2B-emiRFP703 live 20x images. Performance is similar across conditions except H2B-GFP in fixed HeLa cells (teal squares). N = 1. n (on graph) = cells.

VCS can isolate RPE1 RFP703/Dendra micronucleated cells

(A) Overview of VCS protocol. Cells are plated in multi-well plates, during imaging cellular phenotypes are quantified, VCS MN is deployed, and specified classes are photocoverted for either 200 or 800 ms, yielding two different ratios of red:green fluorescence. These differences are quantified by FACs and gated for cell sorting. Graphic created with BioRender.com. (B) Quantification of nuclear red:green ratios from images of the same field taken 0, 4, and 8 h after photoconversion displayed as histograms. Representative images from each time point pseudo-colored by log10 Dendra red:green ratio (below). N = 1, n = 82, 353, 285, 313. (C) Experimental design of MN cell isolation validation. Classifier PPV calculated on images acquired during activation and frequency of MN- or MN+ cells manually quantified in cells plated and fixed post-sorting. Pre-FACS: N = 2, n = 328, 186 post-FACS N=1. n = 338, 353.

VCS pipeline identifies Mps1i transcriptional response.

(A) Timeline of experiment. (B) PCA plot showing clustering of Mps1i-treated and DMSO-treated cells by treatment (major) and by replicate (minor). Each experimental replicate represents 2 technical replicates. (C) MA plot. Differentially expressed genes (FDR adjusted p-value < 0.05) are in green. Gray lines represent 1.5 fold-change in expression. (D) Heatmap of Hallmark pathway enrichment between VCS data and data from (Santaguida et al., 2017) and (He et al., 2019) analyses of RPE1 cells after induction of chromosome missegregation. Hallmark pathways (bottom) were clustered based on manually annotated categories (top).

Micronucleation and rupture transcriptional changes largely overlap with aneuploidy response.

(A) Timeline of experiment for MN+ and MN- cell isolation from RFP703/Dendra cells. (B) PCA plot showing clustering of MN+ and MN-cells by replicate (major) and condition (minor). (C) MA plot. Of identified DEGs, only 2 have fold-changes larger than 1.5. Both, TNFAIP3 and EGR1, are also significantly upregulated in Mps1i treated cells. (D) Timeline of experiment for rupture+ and rupturecell isolation. (E) PCA plot showing clustering of intact MN and ruptured MN cells by condition and replicate. (F) MA plot. Three highly differently expressed genes unique to this dataset are indicated on plot. (G) Heatmap of Hallmark pathway enrichment in datasets of DMSO vs Mps1i, Mps1i-treated cells with and without MN, and synchronized, Mps1i-treated, MN+ cells with and without MN rupture. Pathways are grouped based on manual annotation (left) and show substantial overlap between categories enriched in Mps1i+ cells versus the MN+ and rupture+ subsets.

Micronucleation and rupture do not significantly contribute to the aneuploidy transcription response.

(A) Examples of DNA FISH for chromosomes 1, 11, and 18 and H3K27Ac identification of intact MN. Arrows = ruptured MN, arrowheads = intact MN. (B) Quantification of aneuploidy frequency (foci ≠ 2) per chromosome. Cells manually classified as MN- or MN+, and rupture- or rupture+. MN: Chr 1: N=2, n=429, 158; Chr 11: N=3, n=406, 313, 160; Chr 18: N=3, n=425, 202, 230. Rupture: Chr 1: N=2, n=187, 74; Chr 11: N=3, n=190, 108, 71; Chr 18: N=3, n=186, 102, 101. (C) Heatmap of highly different Mps1i+ DEGs (cutoff = absolute FC 1.5) compared to MN+ and rupture+ replicates. Euclidean distances calculated for features and samples and clustered by complete-linkage. Genes with lacking values for at least one class were excluded. Line = gene cluster upregulated in rupture+ cells. (D-E) Representative images of ATF3 and EGR1 labeling in RPE1 2xRFP-NLS cells after Mps1i incubation. Arrows = ruptured MN cell, arrowheads = intact MN cell (top). Quantification of normalized ATF3 and EGR1 mean nuclear intensity in manually classified cells (bottom). N = 2 (graph colors), n = on graph, p: ns > 0.05, * ≤ 0.05, *** < 0.001 by GEE. Scale bar = 20 µm.

Supporting data for figure 1.

(A) Micronucleation frequencies in hTERT RPE-1s treated with either DMSO (Mps1-) or the Mps1i-inhibitor BAY-1217389 (Mps1i+) for 20 hrs. N = 7, n = 328, 186, 175, 344, 228, 237, 262. **** = p < 0.0001 by GEE. (B) Histogram of number of MN/cell in Mps1i+ RFP703/Dendra cells. N = 5, n = 1323. (C) Quantification of proportion of MN assigned to the correct nucleus by proximity alone. N = 5, n = 264, 158, 365, 283, 249. (D)Distributions of MN/nucleus Dendra2 intensity ratio for intact and ruptured MN. Solid gray line = calculated threshold. N=3, n=179, 113, 105. E, F) Recall and rupture frequency in MN+ cells by # MN. Cells manually classified. N=2, n=328, 186.

Details of UNet architectures and output post-processing in MNFinder module.

(A) The Nuc/MN ensemble classifier takes a single channel input image of chromatin and feeds it into two parallel, attention-gated UNets, one of which also has multiscale downsamplers (yellow). In these blocks input is fed into three parallel, differently-sized convolution operations that are then concatenated. The nucleus weights from the basic UNet are retained and both sets of MN weights are fed to a third UNet for ensembling to produce the final predictions. (B) Results from the Nuc/MN UNet are further processed to improve accuracy. To limit misclassification of large MN as small nuclei, nuclei under a user defined area threshold are reclassified as MN. To limit MN undersegmentation, MN pixel groups are expanded by transforming each into their convex hulls. (C) Example of how a “cell” is generated from existing training data by defining a concave hull that groups a nucleus and any associated MN. Distance and proximity maps are used to define cell boundaries and are derived from convex hulls for training as described in Methods. (D) Diagram of the triple decoder cell segmenter UNet. Two of the decoders have a UNet3+-like architecture with multiple skip connections and deep supervision during training. Feature depths are kept constant and most concatenation/max-pooling operations are replaced with addition to reduce training overhead. One decoder generates distance maps of a concave hull containing each nucleus and any associated MN (a “cell”) and the other generates a proximity map of each cell’s distance to any other. The output of a third decoder that uses a standard UNet with attention gates to segment foreground pixels (nuclei or MN) is used as input into every level of the distance- and proximity-map decoders via an integration block (magenta). (E) Resulting distance and proximity maps from the cell segmenter UNet are combined to generate seeds for watershed segmentation. To correct for oversegmentation, only labels with boundaries that intersect a skeletonized proximity map or border background pixels are retained.

Controls for VCS MN isolation experiments.

(A) Outline of RFP703/Dendra VCS validation experiment using CellTrace labeling as the activation trigger. Cells were incubated with CellTrace far-red and mixed with unlabeled cells at a 1:1 ratio. Nuclei were classified based on CellTrace fluorescence intensity and converted with either an 800 ms (CellTrace+) or 200 ms (CellTrace-) UV pulse. The well was only partially converted prior to FACs analysis and sorting. Representative image of the mixed population prior to photoconversion is shown. Scale bar = 10 µm. (B) FACS plot of Dendra2 red:green ratio versus CellTrace fluorescence. Colored bars represent gates. Values are percentage of negative and positive CellTrace cells present in 200 ms and 800 ms gate, respectively. (C) Histogram of CellTrace fluorescence in cells sorted by Dendra2 ratio after re-analysis by FACs. (D) Predicted classifier PPV (population purity) for untreated low MN frequency U2OS cells (U2OS Broad). We observe a lower but still substantial enrichment of micronucleated cells in the MN+ population compared in a high MN frequency population (Fig. 3C). N = 1, n = 17 cells.

Differential UV pulses do not induce substantial transcriptional changes.

(A) PCA plot of cells treated with DMSO and exposed to 800 ms or 200 ms UV. (B) MA plot of the data in A). Only 6 differentially expressed genes were identified in cells exposed to 800 ms vs 200 ms UV and only 3 were downregulated over 1.5 fold: DDX39B, FASN, RGPD6.

Cell synchronization reduces loss of intact MN cell population purity.

(A) Change in rupture frequency over time in asynchronous and synchronized cells treated with Cdk1i. Other = mitotic, MN-, or Dendra2-cells. N=1, n=∼200 cells per time point. (B) Change in MN rupture frequency between the start and end of a VCS experiment (4h) and predicted change in classifier PPV due to ongoing rupture of intact MN based on values in A).

Controls related to Figure 6.

(A) Quantification of chromosome 1, 11, and 18 micronucleation rates in RPE1 Mps1i cells, grouped by chromosome ploidy. N=2, 3, 3. n = 587, 879, 857. p: **** ≤ 0.0001. (B) Same analysis as Fig. 6b, but with chromosomes in ruptured MN excluded from the foci count. Similar levels of aneuploidy were observed between groups as in Fig. 6b. (C) Representative images and quantification of ATF3 nuclear mean fluorescence intensity in cells treated with DMSO or doxorubicin (Doxo.). N = 2 (colors on graph), n = on graph. (D) Representative images and quantification of EGR1 nuclear mean fluorescence intensity in cells treated with DMSO or hEGF. N = 2 (colors on graph), n = on graph. p: *** ≤ 0.001, by GEE. Scale bar = 20 µm.