High-speed fluorescence image-enabled cell sorting

Abstract

Fast and selective isolation of single cells with unique spatial and morphological traits remains a technical challenge. Here we address this by establishing high speed image-enabled cell sorting (ICS), which records multicolor fluorescence images, and sorts cells based on measurements from image data at speeds up to 15,000 events per second. We show that ICS quantifies cell morphology and localization of labeled proteins, and increases the resolution of cell cycle analyses by separating mitotic stages. We combine ICS with CRISPR-pooled screens to identify regulators of the NF-κB pathway, enabling the completion of genome-wide image-based screens in around nine hours of run-time. By assessing complex cellular phenotypes, ICS substantially expands the phenotypic space accessible to cell sorting applications and pooled genetic screening.

Fluorescence microscopy and flow cytometry are instrumental technologies used in almost all areas of biological and biomedical research. While flow cytometric cell sorting simplifies the isolation of cells in a rapid, sensitive and high-throughput manner, it is limited to a low dimensional parameter space and lacks subcellular resolution (1). It is therefore unable to capture phenotypes associated with processes involving varying signal localization, such as protein trafficking, cellular signaling, or protein mislocalization during disease (2, 3). Fluorescence microscopy, on the other hand, enables high-resolution readouts of cellular morphology and protein localization, but lacks the ability to isolate cells with specific phenotypes at high speed (4). Combining the spatial resolution of fluorescence microscopy with flow cytometric cell sorting has broad implications, and would inspire novel experimental strategies through rapid identification and isolation of cells with specific (sub)cellular phenotypes.

While flow- and microfluidics-based cytometers with imaging capabilities have been developed, these approaches were either unable to sort cells, came with drastically reduced throughput, or depended on non-human interpretable pattern recognition from raw data without image reconstruction (5–14). Furthermore, image-enabled cell sorting has so far relied on technically challenging and custom-built solutions. To date, no system has been developed that integrates traditional flow cytometry and microscopy, operates at speeds compatible with genetic screening approaches and short-lived dynamic phenotypes, and can be operated in non-specialized laboratories.

Here, we present a fully integrated image-enabled cell sorter (ICS) by combining (i) fluorescence imaging using radiofrequency tagged emission (FIRE), a fast fluorescence imaging technique (15) with (ii) a traditional cuvette-based droplet sorter, and (iii) new low-latency signal processing and sorting electronics (Fig. 1A, B, see Methods and fig. S1 for detailed description and characterization of ICS technology, and supplementary text for performance attributes of ICS). To enable blur-free imaging at a high nominal flow speed of 1.1 meters per second, ICS uses the FIRE approach to produce an array of 104 laser spots across 60 μm within the core stream of the sorter cuvette, each modulated at a unique radiofrequency (Fig. 1A). The array of spots excites modulated fluorescent and scattered light from particles or cells as they flow through the optical interrogation region in the cuvette. Emitted light is collected and the signal output is digitized and processed using low-latency field programmable gate arrays, allowing real-time image analysis and image-derived sort decisions. This is different to other image-enabled flow cytometers without cell sorting capabilities (5–8,11–13) (see supplementary text for a comparison between technologies). To reconstruct a row of pixels from the FIRE signal for visualization of the event, the amplitude of the signal at a unique modulation frequency is assigned to a pixel value in a specific horizontal coordinate in the cuvette; in the direction of flow, the pixels are assigned a vertical location based on their temporal value, which forms a two-dimensional image of an event (Fig. 1A). The system collects scatter and fluorescent signals, as well as a light loss signal (analogous to brightfield images produced by traditional microscopes), which allow to visualize events in real-time. This contrasts with Ghost Cytometry, which is unable to reconstruct images from raw data (10). The combination of FIRE with a cuvette-based droplet sorter design, along with the integrated low-latency electronics, enable sorting rates at speeds of up to 15,000 events per second (fig. S1A-C), which is comparable to traditional cell sorters, and at least one order of magnitude faster than IACS (9, 14). Notably, image acquisition and high sorting rates allow immediate human interpretation of the generated data, the capture of dynamic short-lived spatial phenotypes, and the retrieval of sufficient cell numbers for downstream assays, such as genome-scale screens.

Fig. 1. Functionality of the high speed image-enabled cell sorter (ICS).

(A) Schematic representation of the ICS optical and flow hardware components. Excitation beam path: The acousto-optic deflector (AOD) splits a single laser beam (λ = 488 nm) into an array of beamlets, each having different optical frequency and angle. A second AOD tunes the optical frequency of a reference beam, which is then overlapped with the array of beamlets. The overlapping beams intersect the flow cell (FC) of a cuvette sorter. Inset left side: The array of FIRE-beams (dark cyan) is shown overlapping with the reference beam (light cyan). Due to their differing optical frequencies, the overlapping beams exhibit a beating behavior, which causes each beamlet to carry a sinusoidal modulation at a distinct frequency f_1-n. Emission beam path: Images are generated from digitized signals on a per event basis and include light loss, forward scatter (FSC) and side scatter (SSC) images, and four different fluorescent channels. Example images: Hela cells expressing the Golgi marker GalNAcT2-GFP (green) were stained with cell surface marker CD147 PE-CF594 (orange) and DRAQ5 nuclear dye (red). FSC, SSC and light loss images are shown in grayscale. BS, beam splitter; M, mirror; Obj, objective; DP, deflection plates; OB, obscuration bar;P, pinhole; L, lens; BP, band pass; PMT, photomultiplier tube; PD, photodiode. Scale bar represents 20 μm. (B) Overview of the ICS low-latency data processing pipeline. Each photodetector produces a pulse with high-frequency modulations encoding the image (waveform). Fourier analysis is performed to reconstruct the image from the modulated pulse. An image processing pipeline produces a set of image features (image analysis), which are combined with features derived from a pulse processing pipeline (event packet). Real-time sort classification electronics then classify the particle based on image features, producing a sort decision that is used to selectively charge the droplets (dotted gray line in panel (A)). (C) ICS-based imaging of HeLa cells expressing GFP- or mNeonGreen-tagged fluorescent proteins or stained with organellespecific green fluorescent dyes. One representative image is shown per organelle, the full datasets containing 10,000 images each are shared as described in the data and materials availability section. The following dyes or protein fusions were used: cell membrane (Cellmask dye), cytoplasm (GFP fused to HIV Rev nuclear export sequence), mitochondria (Mitotracker dye), nucleus (H2B-mNeonGreen), Golgi apparatus (GalNAcT2-GFP), endoplasmic reticulum (ER, ERtracker dye), nucleolus (eGFP-Ki-67), nuclear envelope (LamB1-GFP), P-bodies (eGFP-DDX6), Cajal bodies (eGFP-COIL), and centrosomes (anti-pericentrin antibody). P-bodies and Cajal bodies were recorded from fixed cells, centrosomes from fixed and metaphase-stalled cells; fixation resulted in decreased contrast in the light loss (LL) image. Scale bar represents 20 μm.

To illustrate the utility of ICS for blur-free visualization of fast-flowing cells and subcellular protein distribution, we imaged a range of well-known organelles and structures of different sizes, shapes, and distributions. We were able to visualize the cell membrane and cytoplasm, membrane-enclosed organelles (nucleus, endoplasmic reticulum, Golgi apparatus, mitochondria), and small membrane-less organelles (P-bodies, Cajal-bodies, centrosomes) (Fig. 1C and fig. S2). We further demonstrated imaging of 13 cell lines of variable size and origin (fig. S3), showing the broad applicability of ICS.

For cell sorting, a set of intuitive spatial image parameters are extracted in real-time from each image channel (Fig. 1B, see Methods and fig. S4A for details of the image parameters). Image parameters are treated identically to conventional pulse parameters (area, width, height) by the sorting electronics, allowing the combination of spatial information and traditional flow cytometry features for analysis and sorting. We demonstrate the ability of ICS parameters to quantify spatial features and to differentiate cells in a variety of use cases, which so far can only be distinguished using microscopy. We were able to separate cells with single or multiple/enlarged nucleoli (Fig. 2A and fig. S4C), single or multiple nuclei (Fig. 2B and fig. S4D), and distinguish cells based on cellular shape (Fig. 2C and fig. S4E). We also demonstrate the ability of ICS to reveal drug-induced organelle responses, such as the effect of brefeldin A on Golgi integrity (Fig. 2D and fig. S4F). Finally, we demonstrate the advantage of multicolor fluorescence imaging for quantification of protein localization via spatial correlation of two signals. We quantified the translocation of the NF-κB pathway component RelA from the cytoplasm to the nucleus upon TNFα induced pathway activation, as detected by an increase in correlation between RelA and the nuclear dye DRAQ5 (Fig. 2E and fig. S4G). These experiments illustrate the utility of ICS parameters for quantification, and ultimately sorting, of a broad spectrum of phenotypes.

Fig. 2. ICS measurements quantify spatial cellular processes and isolate phenotypes of interest.

(A) HeLa cells expressing eGFP-Ki-67 were gated for singlets and live cells, and the ICS size parameter of the eGFP-Ki-67 signal was used to distinguish between cells with single small nucleoli and those with multiple or large nucleoli. Size is defined by the number of pixels above a user-defined threshold. n, nucleolus. Scale bar represents 20 μm. (B) HeLa cells stained with the nuclear dye DyeCycle Green were gated for singlets and live cells, and the radial moment of DyeCycle Green was used to differentiate cells with single or multiple nuclei. Radial moment is the mean-square distance of the signal from the centroid. n., nucleus. Scale bar represents 20 μm. (C) HeLa cells were gated for singlets and live cells, and the eccentricity calculated from the side scatter image was used to distinguish round from elongated cells. Eccentricity is computed by first finding the magnitudes of the spread along the two principal components of the image, then taking their ratio. Scale bar represents 20 μm. (D) HeLa cells expressing the Golgi marker GalNAcT2-GFP were gated for singlets and live cells, and treated with brefeldin A (BFA) or left untreated. The maximum intensity of the GalNAcT2-GFP channel was used to distinguish treated from untreated cells, while the overall GFP intensity (y-axis) was largely unaffected by the treatment. Maximum intensity is the value of the brightest pixel. A, area. Scale bar represents 20 μm. (E) HeLa cells expressing RelA-mNeonGreen (mNG) were treated with TNFα or left untreated, and stained with the cell permeable nuclear dye DRAQ5. Cells were then gated for singlets and live cells, and the correlation between RelA-mNG and DRAQ5 was used to differentiate between the treated (nuclear RelA) and untreated (cytoplasmic RelA) conditions. Correlation is the Pearson’s correlation score between the intensities of the pixel values from two imaging channels. Scale bar represents 20 μm. (F) HeLa cells expressing H2B-mNeonGreen (mNG) were synchronized to increase the frequency of rare mitotic stages, and released into mitosis without chemical perturbation. Then, cells were fixed for labeling with an antibody recognizing phosphorylated serine 10 on histone H3 (pS10H3), and to allow microscopic validation after sorting. Samples were stained with DAPI for univariate cell cycle analysis. Representative images of individual cells within the G2/M population reveal captures of major mitotic stages. LL, light loss. Scale bar represents 20 μm. (G) A decision tree model was trained to distinguish the mitotic stages of manually classified datasets (n = 100 per stage, three replicate recordings and classifications). Shown are the results of a feature importance analysis of ICS measurements, that represents the summarized reduction in the loss function attributed to each feature at each split in the tree. A, area; RM, radial moment; ecc, eccentricity; MI, maximum intensity. (H) Feature values from panel (G) were standardized, and median values for cells and from three replicates of classified datasets are shown as a heatmap. Only features that vary between the mitotic stages are shown (variable importance greater than 0 in panel (G)). (I) Based on the identified features in panel (H), a hierarchical gating strategy was built that enriches for interphase, prometaphase, metaphase, anaphase, and telophase stages. A, area; RM, radial moment; MI, maximum intensity. (J) 5,000 cells were sorted for microscopic validation based on the gating strategy established in panel (I), and manual classification from confocal z-stacks of the sorted cells was performed. Shown are mean percentages of three independent replicates. Prometaphase cells were generated by two consecutive sorts (see Methods). interph., interphase. (K) Representative single-slice confocal fluorescence microscopy images from sorted cells from panel (J) with brightfield/H2B-mNG overlays as inlays. Scale bar represents 50 μm.

To demonstrate cell sorting functionality of the ICS, we applied it to the mitotic cell cycle, a dynamic process associated with multiple complex phenotypic changes. Traditional flow cytometry can separate three cell cycle stages (G1, G2/mitosis and S phase), but fails to distinguish cells in different mitotic stages. While chemicals that block mitosis can be used to enrich certain stages (notably excluding anaphase and telophase) (16–18), these approaches can alter gene expression and posttranslational modifications. We demonstrate that ICS can isolate mitotic stages of HeLa cells, using H2B-mNeonGreen (mNG) to visualize chromatin and the intensity of phosphorylated Ser10 on histone H3 (pS10H3) as a marker associated with mitotic chromatin condensation (19). We investigated cells from the G2/mitosis phases of the univariate cell cycle and created a training dataset by manually classifying 100 cells from each stage throughout mitosis (Fig. 2F, see Methods for criteria to distinguish mitotic stages). Classified events were organized on a trajectory in chronological order (fig. S5A). We used this training dataset to identify the most differing image-, scatter- and intensity-based parameters between stages by fitting a decision tree model and performing feature importance analysis (20) (Fig. 2G and fig. S5B). Image-derived parameters dominated the most differentiating parameters, such as maximum intensity, radial moment and eccentricity of the H2B-mNG signal that differentiated between metaphase, anaphase and telophase cells (Fig. 2H). We used these features to establish a hierarchical gating strategy for cell sorting, and performed independent microscopic validation of the isolated populations (Fig. 2I). We found that ICS isolated highly pure populations, including G2 interphase (96% purity), prometaphase (64%), metaphase (78%), anaphase (94%), and telophase (93%) (Fig. 2J, K and fig. S5C-E). With these advances, we increased the resolution of flow cytometric cell cycle analyses to the level of distinguishing individual mitotic stages (including the thus-far inaccessible anaphase and telophase stages), yielding a method for robust enrichment of high numbers of cells in the absence of chemical blockers, and from the same source sample. Isolated cells can be utilized in numerous downstream applications, such as the comparison of stage-specific changes in transcriptome, chromatin architecture, or protein modifications.

Pooled functional genomic screens with microscopic readouts have so far been limited in throughput and depended on technically challenging methods (21–25). ICS allows high speed cell isolation based on fluorescence spatial information, and therefore has the potential to increase scale and speed of microscopy-based screens, as well as reducing technical complexity, duration, and cost. We tested the compatibility of ICS with pooled CRISPR screens by examining the nuclear translocation of RelA upon NF-κB pathway activation - a process that is invisible to traditional flow cytometry. To measure RelA translocation upon CRISPR-mediated perturbation, we quantified RelA-mNG/DRAQ5 spatial correlation (see also Fig. 2E) in HeLa cells expressing Tet-inducible Cas9 and fluorescently-tagged RelA (23) (HeLa RelA) (fig. S6A-C). We validated the approach using individual CRISPR knockouts of three core NF-κB pathway components (IKBKA, IKBKG, MAP3K7) and found consistent defects in RelA nuclear translocation upon gene knockout, demonstrating that ICS sensitively captures the effects of those perturbations (Fig. 3A and fig. S6D). Next, we proceeded with a pooled screen, in which a population of Cas9 expressing cells is transduced with a mixture of guide RNAs (gRNAs). We transduced HeLa RelA cells with an NF-κB pathway-focused library targeting 1,068 genes including 37 NF-κB core canonical pathway components. Cells were then treated with TNFα and we isolated the 5% lower (cytoplasmic RelA) and upper (nuclear RelA) bins of the RelA-mNG/DRAQ5 correlation parameter (Fig. 3B and fig. S7A). Sorting was conducted with an average event rate of 4,000 events per second, a comparable speed to current flow-based technology for large cells like HeLa, enabling a 100x coverage of a 1,000 gRNA library in less than 9 minutes. Bulk sorts were performed at different library coverage to determine optimal library coverage and gRNA number per gene. We generated a “ground-truth” high-coverage (359-fold) dataset by pooling all reads from the differently sized samples followed by gRNA hit calling (26). Among the most significant hits, we identified known NF-κB pathway components, demonstrating that ICS can identify bona fide regulators of the NF-κB pathway (Fig. 3C and table S1). We found strong correlation between the individual and pooled perturbations, indicating that both perturbation strategies rank genes similarly (Fig. 3D and fig. S7B). Next, we investigated how the number of gRNAs per gene and library coverage affect hit-calling performance. High performance (area under the precision recall curve AUPRC > 0.7; 70% of hits detected at < 1% FDR) was achieved with only 100 cells per gRNA, and three gRNAs per gene (Fig. 3E and fig. S7C, D). Performance increased with library coverage and number of gRNAs per gene, as sporadic false hits caused by gRNA dropouts in the low coverage samples decreased (fig. S7E, F). Independent screen replicates showed high reproducibility (0.77 ≤ r ≤ 0.87, fig. S7G).

Fig. 3. ICS detects the effects of CRISPR perturbations and enables pooled genetic screens of protein localization.

(A) Effects of individual CRISPR perturbations on RelA nuclear translocation. HeLa cells with Tet-inducible Cas9 and stably expressing RelA-mNeonGreen (mNG) were transduced with guide RNAs (gRNA-1, -2, -3) targeting core NF-κB pathway proteins IKBKG, IKBKA and MAP3K6 or non-targeting (nt) control gRNAs. gRNA expression was induced with Doxycycline (Dox) or left uninduced. Correlation between RelA-mNG and DRAQ5 was quantified using ICS as a measurement for RelA nuclear translocation in the presence or absence of TNFα. (B) Overview of the pooled CRISPR screening setup and readout using ICS. Positive regulators of RelA nuclear translocation are enriched in the lower bin and depleted from the upper bin. Dox, doxycycline; Tet::Cas9, Tetracycline/Doxycycline-inducible Cas9. (C-E) Results of the ICS-based CRISPR screen using an NF-κB pathway-focused library (n = 1,068 genes). (C) The screen was performed at different library coverages, and reads from collected samples were combined in silico to a high-coverage (359 cells per gRNA per sorted bin) dataset. Hits were called using the software MAUDE (26). Genes are ranked by their statistical significance and selected positive/negative regulators are highlighted. The horizontal dashed lines indicate a false discovery rate (FDR) of 1%, while genes with FDR < 1% were marked in cyan and orange, respectively. (D) Comparison of phenotypes measured in individual perturbation experiments from panel (A) (x-axis) or the pooled screen (y-axis) using the same gRNAs. For the pooled screen, differences in gRNA abundance in the upper (top panel) and lower (bottom panel) sorted bins compared to the input sample were determined from the high-coverage dataset in panel (C). R-values represent Pearson correlation coefficients. FC, fold change. (E) Screen hits as determined at different library coverages (12 to 155 cells per gRNA per sorted bin) using between one and six gRNAs per gene were compared to a high coverage reference sample (359x, six gRNAs per gene) by precision-recall analysis. Heat map shows area under the precision-recall curve (AUPRC) values for different levels of library coverage and different numbers of gRNAs per gene. (F-J) Results of the ICS-based genome-wide screen (n = 18,408 genes). (F) Scatter plot of fold changes visualizing gRNA abundance changes in upper (x-axis) and lower (y-axis) sorted bins compared to the plasmid library. Blue and yellow dots indicate statistically significant positive and negative regulators (FDR < 1% according to MAUDE). (G) Genome-wide CRISPR screen identified core canonical NF-κB pathway components. Left panel: Schematic of the core canonical NF-κB signaling pathway. Right top panel: Distribution of the gRNA Z-score for the whole genome-wide library. Right panels: gRNA Z-score for individual gRNAs per gene, overlayed with a gradient (grayscale) depicting overall Z-score distribution. Right bar chart: Gene essentiality as determined by the log2 fold change (FC) of the gRNA abundance in the unsorted cell population compared to the plasmid library. (H) GO network of hits with FDR < 1%, colored by modules identified from protein–protein interactions using STRING-db (45). Gray lines connect associated GO terms, edges represent GO terms. Names of individual edges were omitted, clusters that were not associated with immune signaling or chromatin modification were collected in a third class called “others”. (I) Screen results for SAGA and INO80 protein complex components. Left panel: Schematic illustration of the SAGA and INO80 protein complexes. Right panels: as described in panel (G). (J) Selected hits from the genome-wide screen (1 gRNA per gene, we picked the gRNA that showed the strongest z-score in the pooled genetic screen) were validated using two orthologous methods (individual validation using ICS, individual validation using microscopy). The top row in the heatmap shows the phenotypes measured in the genome-wide screen (MAUDE Z-score). The phenotype in the second and third rows of the heatmap represents the standardized difference in signal medians between the knockout and control gRNA cell populations. Nuclear RelA abundance was quantified using microscopy by measuring the correlation between RelA-mNeonGreen and DRAQ5.

To fully exploit the high speed capabilities of ICS, we next sought to identify NF-κB pathway regulators globally in a genome-wide screen. We generated a novel genome-wide CRISPR/Cas9 library targeting 18,408 protein-coding genes with fully adjustable numbers of gRNAs per gene (fig. S8 and supplementary text). Using six gRNAs per gene and a 100x library coverage, we identified 169 hits (FDR < 1%), encompassing 133 positive and 36 negative regulators (Fig. 3F and fig. S9A, table S2, and supplementary text). A down-sampling based analysis confirmed that three gRNAs per gene ranked genes similarly to the full library of six gRNAs (fig. S9B, C). Among these hits, we identified all core canonical NF-κB pathway components, except for three pathway genes (TRAF5, TAB1, NFKB1), consistent with previous reports of these genes not being essential for pathway functionality (23, 27–29) (Fig. 3G). To identify potential novel regulators, we performed a GO-term-based network analysis, which showed striking enrichment of a cluster of processes centered around chromatin modification (Fig. 3H). Among the underlying genes, we identified the histone deacetylase HDAC3, which induces RelA nuclear export during pathway shutdown (30). We also found previously unknown regulators including multiple components of the SAGA chromatin-acetylation complex (31) and the INO80 chromatin-remodeling complex (32), indicating a previously unknown role of these complexes in NF-κB pathway regulation (Fig. 3I, supplementary text). For hit validation, we assessed the top ten previously unknown positive and negative candidates, the ten identified members of the SAGA and INO80 complexes, and three known NF-κB pathway components. Individual CRISPR knockouts followed by quantification of RelA nuclear translocation using both ICS and microscopy revealed strong agreement (0.857 ⩽ R ⩽ 0,908) between these measurements, and confirmed the observations from the pooled genetic screen (Fig. 3J and fig. S9D, E). In addition, our validation experiments indicate that ICS can reach similar accuracy as, and ranks genes similar to fluorescence microscopy (Fig. 3J and fig. S9D). With the applied event rate of 4,000 events per second, ICS is significantly faster compared to recently developed microscopy-based methods for pooled genetic screens (22–25) (comparison in supplementary text), and enabled the completion of a genome-wide screen (three gRNAs per gene, 100x coverage) within only nine hours of run time.

In conclusion, ICS substantially expands the phenotypic space accessible to cell sorting applications and functional genomic screening. ICS meets the requirements of high speed cell sorting, multicolor fluorescence imaging, and full integration into a device that can be operated in non-specialized laboratories. This will ensure broad availability and inspire new experimental strategies in diverse areas, including basic research, cell-based diagnostics, cell atlas efforts (3), and high-content image-based screening (2, 33, 34). With the potential to include downstream (multi)omics readouts (35–41), ICS provides a fundamentally new capability for probing deep into the molecular mechanisms underlying cell physiology and protein localization.

One-Sentence Summary.

A high speed cell sorter uses fluorescence imaging to enable genome-scale studies of complex phenotypes.

Data and materials availability

NGS data from gRNA library and targeted genome sequencing was deposited at Gene Expression Omnibus GEO (GSE167944). Documented code to reproduce all analyses and figures is deposited at GitHub (https://github.com/benediktrauscher/ICS) and Zenodo (42). Fiji tools and source code thereof are deposited at GitHub (https://github.com/embl-cba/ICS) and Zenodo (42). ICS image data was deposited on BioImage Archive (43) (S-BSST644, available via https://ebi.ac.uk/biostudies/). Flow cytometry data was deposited at flowrepository.org (44) (FR-FCM-Z4M5). Metadata and archiving information for ICS data are provided in table S3. Plasmid "phage UbiC tagRFP-T-DDX6” is available under a material transfer agreement from Addgene. Cell line "HeLa Tet::Cas9 RelA-mNeonGreen” is available under a material transfer agreement from Broad Institute (Paul Blainey).