A computational platform for robotized fluorescence microscopy (I): High-content image-based cell-cycle analysis


  • Laura Furia,

    1. Department of Experimental Oncology, European Institute of Oncology, IFOM-IEO Campus for Oncogenomics, Milano 20139, Italy
    Search for more papers by this author
  • Pier Giuseppe Pelicci,

    1. Department of Experimental Oncology, European Institute of Oncology, IFOM-IEO Campus for Oncogenomics, Milano 20139, Italy
    Search for more papers by this author
  • Mario Faretta

    Corresponding author
    1. Department of Experimental Oncology, European Institute of Oncology, IFOM-IEO Campus for Oncogenomics, Milano 20139, Italy
    • Department of Experimental Oncology, European Institute of Oncology, IFOM-IEO Campus for Oncogenomics, via Adamello 16, 20139 Milano, Italy
    Search for more papers by this author


Hardware automation and software development have allowed a dramatic increase of throughput in both acquisition and analysis of images by associating an optimized statistical significance with fluorescence microscopy. Despite the numerous common points between fluorescence microscopy and flow cytometry (FCM), the enormous amount of applications developed for the latter have found relatively low space among the modern high-resolution imaging techniques. With the aim to fulfill this gap, we developed a novel computational platform named A.M.I.CO. (Automated Microscopy for Image-Cytometry) for the quantitative analysis of images from widefield and confocal robotized microscopes. Thanks to the setting up of both staining protocols and analysis procedures, we were able to recapitulate many FCM assays. In particular, we focused on the measurement of DNA content and the reconstruction of cell-cycle profiles with optimal parameters. Standard automated microscopes were employed at the highest optical resolution (200 nm), and white-light sources made it possible to perform an efficient multiparameter analysis. DNA- and protein-content measurements were complemented with image-derived information on their intracellular spatial distribution. Notably, the developed tools create a direct link between image-analysis and acquisition. It is therefore possible to isolate target populations according to a definite quantitative profile, and to relocate physically them for diffraction-limited data acquisition. Thanks to its flexibility and analysis-driven acquisition, A.M.I.CO. can integrate flow, image-stream and laser-scanning cytometry analysis, providing high-resolution intracellular analysis with a previously unreached statistical relevance. © 2013 International Society for Advancement of Cytometry © 2012 International Society for Advancement of Cytometry

Limited sample availability, as for material of clinical origin, and a poor statistical representation of targeted cell populations, for example, stem cells, can make the comprehension of biological heterogeneity a very challenging task (1–4). High-content approaches have been consequently developed to extend the number of simultaneously analyzable parameters (5–7).

Flow cytometry (FCM) is a powerful technique that provides statistically relevant measurements. Quantification at single-cell level, elevated content, fast acquisition, and contained analysis time make it a very valuable tool for the identification of rare cell populations. In addition, enormous advances in limiting the required amount of sample per analysis have also been achieved by system miniaturization (7–11).

This excellent “statistical resolution” is, however, coupled to a complete absence of an intracellular view and, to date, fluorescence microscopy maintains its leadership in providing a high-resolution description of the inner cell compartments. Nevertheless, technological efforts have tried to fill the imaging gap between FCM and fluorescence microscopy, leading to the development of image-streaming cytometry (ISC) and laser-scanning cytometry (LSC; 12–23). Developments based on time-delay integration, extended field-of-view for in focus imaging of the entire cell, and a wide range of detectable fluorophores, made ISC an excellent tool for statistically relevant imaging at a good resolution (21, 22, 24–30). However, the in-flow approach can represent an obstacle for some applications (31, 32) since cells are detached from their substrate, with the consequent loss of information about intercellular communication and/or interactions with the surrounding environment. However, slide-based LSC shares some of the features of ISC (e.g., spectral imaging) but has gained big advantages from the preservation of tissue-sample architecture, acquiring the possibility to execute multiple measurements (1, 3, 33, 34).

Despite these technological advances, the high optical resolution granted by oil-immersion microscope objectives, is not currently achievable either by ISC or by LSC.

An improvement in the spatial resolution of cytometry approaches is highly desirable to analyze proteins and molecules in their natural environment. This step requires the validation of high-resolution imaging measurements with a statistically relevant number of analyzed events. Microscopy for life sciences has been considered for years a merely descriptive tool, due to the extremely low throughput that is typical of human-driven approaches. The removal of this barrier thanks to the microscope robotization gave rise to a plethora of applications (35–38). New instruments capable of unprecedented high-content and throughput have been developed, employing the high-sensitivity and speed of the widefield microscope and/or the optical-sectioning ability of the confocal microscope (37, 39). Excellent image-analysis tools have been created (40), and new technological solutions have contributed to the foundation of tissomics (41–46). Moreover, quantitative in-situ analysis of the human genome has strongly stimulated the development of ad-hoc high-resolution cytometry approaches (47–49).

However, even if significant efforts have been made to give a robust statistical validity to high-resolution and high-content imaging, the large repertoire of assays provided by FCM has been only marginally exploited by fluorescence microcopy and image-based approaches.

With the aim to fulfill this gap, we developed an image-cytometry protocol combined with a series of computational tools (A.M.I.CO., Automated Microscopy for Image-Cytometry) able to confer to different microscopes a high quantitative and statistical potential. Both conventional and confocal automated microscopes have been employed, working at variable spatial resolutions thanks to a platform-independent image-analysis procedure. The developed tools inherited from FCM the ability to perform multiparameter analysis, targeting of specific cell subpopulations, and cell-cycle analysis based on DNA content evaluation. We validated the approach on the cell-cycle distribution of well known targets in FCM, by associating the maximum optical resolution (about 200 nm) to high statistical sampling (from 5,000 to 20,000 events analyzed). We also demonstrated how an optimal choice of the acquisition setting (i.e., magnification, numerical aperture (NA), pinhole aperture, pixel dimension, and section spacing) provided good quality data in accordance with the evaluation criteria of FCM analysis (Coefficient of Variation, CV, and linearity). Simultaneously, we reported how our quantitative and qualitative analysis can be employed to complement cell-cycle-related measurements with information typical of a high-resolution, image-based approach.

To remove limitations imposed by the large amount of data required for a high-content, diffraction-limited, and statistically significant analysis, we created a direct connection between data collection and processing, two processes that have been frequently considered as sequential and independent steps. A valuable feature of the developed approach is, indeed, the possibility to perform multiple acquisitions on the same sample. The ability to relocate events based on a first analysis has been already validated in both slide-based LSC and high-resolution microscopy (1,3,14,50-56). We demonstrated that, in our approach, a first acquisition could be exploited to target in-silico identified subpopulations. In this step, spatial resolution is limited in favor of statistical sampling, minimizing computational and storage efforts. Optimal acquisition conditions, which maximize spatial resolution, sensitivity, and content, can then be selectively applied to the targeted fraction of events by a computer-assisted high-resolution 3D acquisition. Notably, A.M.I.CO. differs from many commercialized softwares, as it is not restricted to a specific platform but is compatible with commercially available instruments and easily adaptable to custom solutions.

Materials and Methods

Cell Culture

MCF10A cells were grown in DMEM + Ham's F12 Medium (1:1) containing 5% FBS, 2 mM glutamine, 50 ng/ml penicillin/streptomycin (all from Lonza, Switzerland), cholera toxin (Sigma-Aldrich, MO), 10 μg/ml insulin (Roche, Switzerland), 100 μg/ml hydrocortisone (Sigma-Aldrich), and 20 ng/ml EGF (PeproTech, NJ) at 37°C in 5% CO2. Cells were grown on glass coverslips coated with 0.5% gelatin (wt/vol) in PBS.

To detect actively DNA-replicating cells, Ethinyl-deoxyuridine, EdU, (57,58; Life Technologies, CA), was added to the culture media (final concentration 10 μM) 40 min before fixation. Exponentially growing cells were fixed for 10 min in 4% paraformaldehyde (wt/vol) at room temperature (RT).

EdU Staining and Immunofluorescence of MCF10A Cells

Fixed MCF10A cells were washed and permeabilized for 10 min in a permeabilization buffer containing 0.1% Triton X-100 (vol/vol) in PBS. EdU incorporation into DNA was detected using the Click-iT™ EdU Imaging kit (Life Technologies), according to the manufacturer instruction. All steps of the Click-iT™ reaction were performed at RT. For cyclin detection, a biotin-conjugated azide (B10184, Life Technologies) was added to the reaction cocktail. After EdU reaction, coverslips were immersed for 30 min in a blocking solution, 5% BSA (wt/vol) in PBS, then incubated for 1 h at RT with primary antibodies in blocking solution. To measure the cell-cycle distribution of cyclin proteins, the following antibodies were employed: rabbit anti-cyclin A (H432, Santa Cruz Biotechnology, Germany) and mouse anti-cyclin E (HE12, BD Biosciences, NJ). The cells were rinsed three times in PBS and incubated for 1 h at RT with anti-rabbit Pacific Orange-conjugated IgGs (Life Technologies) and anti-mouse Cy3-conjugated IgGs (Jackson Immuno-Research, UK). To detect the biotin azide, an anti-biotin CW800-conjugated antibody (600-132-098, Rockland, PA) was incubated together with the secondary antibodies. The cells were again rinsed three times in PBS, briefly refixed in 4% paraformaldehyde (wt/vol), and blocked with 5% BSA (wt/vol) containing mouse IgGs (Jackson Immuno-Research). After washing, the cells were incubated for 1 h with anti-KI67 HorizonV450-conjugated (clone B56, BD Pharmingen) monoclonal Antibody (mAb). DNA was counterstained with 4′,6-diamidino-2-phenylindole (DAPI) 10 μg/ml for 1 h. Slides were mounted in Mowiol-containing mounting media for widefield fluorescence microscopy analysis. For the KI67 analysis, EdU incorporation into DNA was detected using the Click-iT™ EdU Alexa 555 Imaging kit (Life Technologies). An Alexa 647-conjugated mouse mAb (B56, BD Pharmingen) was used to detect KI67. For confocal analysis, DAPI was replaced by Chromomycin A3 (10 μM in PBS containing 70 mM MgCl2). This dye grants stoichiometric binding for a correct DNA content estimation [(59) and references therein]. Since it is excited in the visible range (458 nm, Argon laser line), it provided a more uniform field of excitation and maximized chromatic-aberration correction. Slides were then mounted in glycerol-based mounting media containing 1,4-diazabicyclo(2.2.2)octane (DABCO) to preserve the cell three-dimensional structure.

Antibody specificity was checked by a blank control sample in which cells were labeled with secondary antibodies only. In addition, cells were stained separately for each of the target proteins and analyzed to verify the absence of crosstalk and cross-reactivity.

Automated Microscopy and Image Acquisition

All widefield images were collected by an Olympus BX61 fully motorized fluorescence microscope controlled by the image screening Scan∧R software (version 2.2.09, Olympus, Germany) and equipped with an ORCA ER Hamamatsu Camera (Hamamatsu, Germany). The fluorescence-filter configuration was designed to minimize crosstalk among channels. The following objectives were employed for widefield acquisition: 20× 0.75 NA, 40× 0.90 NA, and 60× 0.90 NA dry objectives; 40× 1.3 NA and 60× 1.35 NA oil-immersion objectives. Exposure times were calculated in order to ensure optimal sensitivity and minimal photobleaching. No binning has been employed for any acquisitions (CCD camera Pixel Linear Size: 6.45 μm; Image pixel size: 6.45 μm/objective magnification) to obtain the correct spatial sampling. Focus was determined by an algorithm for the recognition of target objects (cells).

Confocal microscopy data were collected with oil-immersion 40× 1.25 NA and 63× 1.4 NA objectives by a SP5 laser-scanning spectral confocal microscope, equipped with a resonance scanning unit and controlled by the Matrix routine for high-content microscopy included in the LAS software package (Leica Microsystems, Germany). To maximize the speed in data collection and to avoid laser-induced photobleaching, a scanning frequency of 8000 Hz was selected for image recording. An optimized slit aperture in the spectral detection unit and a sequential acquisition were employed to avoid both channels bleedthrough and crosstalk (60–62). For image-analysis-driven acquisition, an undersampled large field-of-view (40×, about 200 μm) was initially defined, and data collection was limited to the parameters required for the identification of the target subpopulation (DNA and KI67 fluorescence). To avoid photobleaching and to minimize Z-slices collection, the pinhole aperture was increased to nonconfocal imaging modality (>2 airy units, see Results; 63,64). A number of adjacent stage positions, ranging from 900 to 1600, were collected, to reach a cell population of at least 5,000 events. Once the position list of the selected cells was produced by the image analysis, the second round of acquisition started. Images obtained from all the parameters of interest were collected under Nyquist oversampling conditions (63× 1.4 NA, optical zoom 4.9, 50 nm XY pixel size, slice spacing <200 nm; 64). The systematic shift due to repositioning was corrected by panning of the scanning mirrors.

Removal of Image-Acquisition Induced Artifacts

Quantitative image analysis requires a preprocessing procedure to subtract the background and to correct illumination disparities. In addition to software-calculated background, empty coverslips were employed as a reference for the experimental determination of the background level (Reference-based background). The creation of even illumination conditions was based on imaging of fluorescent slides (Chroma Technologies) and/or thin fluorescent layers (65). Flat-field corrected images were normalized to their mean value, after background removal, and used as input for the image analysis routine.

Image Analysis

The entire A.M.I.CO. analysis package has been developed within the open-source ImageJ platform (66), exploiting its macro programming language. A.M.I.CO. is composed of two separated modules.

The first module is devoted to data browsing, parameter setting and image-analysis execution (Fig. 1, Panels I and II). The definition of the required experimental parameters, when not read in ad-hoc stored files, was created by the user allowing for correct channel recombination and browsing through the stored data. Image-screening softwares (e.g., Scan∧R, Metamorph Multi Dimensional Acquisition Module, Micro-Manager, Leica Matrix) store experimental parameters in the filenames. The data-loading interface handles specific acquisition formats thus deciphering the employed fluorophores, the acquisition time interval, and the spatial sampling (both Z-slices numbering and stage positions). Once the number of acquired channels is defined, the image browsing only requires tiff-images as input. Since browsing is based on the sequential name ordering, it is also compatible with any acquisition software able to save or export the tiff format. Image analysis was preceded by subtraction of a calculated region-of-interest (ROI) background, a Rolling-ball algorithm background, or a Reference-based background and flat-field correction (see the previous section; Fig. 1, Panel III). Segmentation threshold can be computed according to the different methods available in the ImageJ software (http://fiji.sc/wiki/index.php/Auto_Threshold). The Watershed segmentation algorithm present in ImageJ is then applied to separate nuclei clusters. Geometrical parameters (area, circularity, and physical position in the field-of-view) are calculated for each identified cell. Our experimental conditions (type of cells, DNA dye concentration, employed magnification, and exposure time) allowed an efficient decomposition of cell clusters. Artifacts created by nuclei over-segmentation were eliminated by the analysis of their size and circularity.

Figure 1.

Computational platform: schematic representation. The developed software, fed by images acquired by automated microscopes (I), asks for a user-guided definition of basic parameters to determine the type of analysis (II), corrects for imaging artifacts (III), executes image-analysis creating a single-cell data repository (IV), and creates a flow-cytometry like data-analysis interface to target cell populations and statistical measurements (V). Every event can then be physically retrieved (VI) to be re-analyzed adopting appropriate image-acquisition parameters for single-cell high-resolution analysis. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Mean and maximum values of fluorescence per pixel are calculated, together with the integrated fluorescence intensity per cell (Fig. 1, Panel IV). For every analyzed channel, subcompartments can also be defined for the measurement of intracellular details. The segmentation analysis of subcompartments includes ad-hoc image filters to enhance their detection, for example, Laplace-of-Gaussian filtering or Top-Hat transformation, based on the available ImageJ plugin (G. Landini, Grayscale Morphology; http://www.dentistry.bham.ac.uk/landinig/software/software.html).

Physical and fluorescence parameters are re-calculated for each detected intracellular event, and summary values are assigned to the pertaining cell of origin (number of detected spots, average size per spot, total area of the cell covered by the spots, ratio between spots and cell area (fractional occupancy), mean fluorescence value per spot, total fluorescence related to the subcompartment, ratio between confined subcompartment, and total cell fluorescence for every analyzed channel). The software can also process confocal stacks; as a general approach, their analysis is dimensionally reduced by an additive projection for the measurement of the global fluorescence values per cell. We also implemented a real 3D approach to the evaluation of subcompartment features, using the publicly available 3D particle count plugin (67).

To exploit the multi-step acquisition procedure for an optimized single cell analysis with correct oversampling better, we also decided to implement the following optional routines for the processing of high-resolution microscopy images:

The analysis can be performed either on single-experiments or in batch on different image collections, for example, different time points and/or drug treatments. An online analysis routine was implemented to speed up the entire process. Data can be processed in a coordinated manner with respect to image acquisition: image analysis can be delayed for a defined time-interval according to the adopted acquisition scheme (an online analysis compatible with both the adopted control softwares, Scan∧R and Matrix, has been successfully implemented).The execution time of the analysis varies according to the number of acquired images, parameters, and subcompartments, ranging from about one up to several hours. The analysis procedure stores data in a tab-delimited text file containing the calculated measurements, which are classified in cell- or sub-compartment-related values and indexed to the corresponding image-file and cell of origin (in case of sub-compartments). This allows the compatibility with statistical and/or FCM analysis packages (e.g., R, BioConductor, Cytospec).

The second data-analysis module presents a FCM-like interface with the possibility of creating histograms and 2D multicolor dot-plots (Fig. 1, Panel V). As previously mentioned, the software allows to define specific regions for statistical measurements and logical gates for selective targeting of events. Cells and subcompartments can be analyzed either according to a hierarchical view (compartments as a property of the event “cell“), or considering them as separated statistical populations. The gates work independently for cells or compartments, and their effects in the resulting statistical filtering can be specified by the user. For every targeted population, a statistics summary can be obtained, reporting the corresponding number of events, global or relative (to gate) percentage representation, and the average value per cell for every measured parameter. Additionally, for every region, it is possible to calculate the Pearson correlation coefficient and a linear interpolation for a two-parameter correlation analysis. Every object (cell or subcompartment) can be visualized thanks to the indexing to the image file of origin. Each event can be physically positioned on the microscope stage, correlating its file of origin to the stage position list generated by the employed acquisition software. The A.M.I.CO. analysis software can read the calibration and acquisition parameters available in the data-repository specific for each machine (Fig. 1, Panel VI). A position list is therefore generated to center targeted objects (cells or subcompartments), recalculating the centered position from the image of origin. An a posteriori shift along the three spatial dimensions can be introduced by the user before a second round of data collection; this procedure allows to correct any systematic errors introduced by stage positioning and/or objective change. The produced list of stage coordinates allows object repositioning to perform a single-cell, high-resolution data collection. Further examples of applications are reported in the article accompanying this technical report. The software is freely available for the scientific community upon request.


Correct Background-Subtraction and Excitation Normalization Improves the Quality of DNA Histograms

The correction of image acquisition artifacts is fundamental for data analysis and interpretation, particularly for the evaluation of the DNA content distribution of a cell population. As described in the “Materials and Methods” section, A.M.I.CO. provides different tools for background subtraction and illumination inhomogeneities removal. We compared the effects of different preprocessing steps on the reconstruction of monoparametric DNA profiles. DNA content was obtained by measuring the total DAPI fluorescence intensity per nucleus. Specifically we tested: (i) Simple background subtraction based on a fixed value (determined by a user-defined ROI), (ii) Rolling-ball subtraction algorithm (included in ImageJ), and (iii) Reference-based background subtraction coupled to flat field correction; results are reported in Figure 2.

Figure 2.

Influence of image pre-processing on the estimation of DNA content distribution in high-resolution widefield image-cytometry. DNA content distribution (first row) calculated without (RAW column) and after removal of background by different procedures, i.e. estimation from a preset region of interest (ROI column), Rolling-ball algorithm included in ImageJ software (Rolling Ball column) and background subtraction and flat-field illumination correction estimated from reference samples (Reference column). Bivariate analysis of DNA-content versus cell area (second row) provided correct distribution with optimal linearity. CV of the histograms was dramatically reduced upon flat-field correction due to the removal of artifacts induced by uneven illumination, as demonstrated by the plot of DNA-content versus physical coordinates of the cell in the field-of-view (third and fourth rows). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

Clearly, background removal was crucial for the correct reconstruction of signal distribution, as shown by the DNA histograms in Figure 2, first row. In absence of any background subtraction, DNA-content estimation was affected by aspecific signals, as demonstrated by the direct correlation between DNA fluorescence and cell area (Fig. 2, second row, RAW column). The use of a spatially invariant background, as the one set according to the mean value of an empty ROI, allowed the reconstruction of a DNA profile with optimal linearity between G1 and G2 peaks (Fig. 2, first row, ROI column) and the removal of the linear correlation with the cell size. Moreover, G1 and G2 populations presented a distribution of DNA content versus cell area that is comparable to the one detected in standard FCM (Fig. 2, second row, ROI column). However, suboptimal CVs were measured. Peak spreading was related to an uneven illumination as shown in the dot plots reporting DNA content versus spatial position of the cell in the field-of-view (Fig. 2, third and fourth rows, ROI column). A decreased illumination at the image borders caused a DNA content underestimation and, consequently, an increase in the standard deviation of the Gaussian distribution. The introduction of a local background, based on the Rolling-ball algorithm, provided comparable results (Fig. 2, Rolling-Ball column). The Rolling-ball algorithm estimates the background by calculating circles of size in the order of the objects present in the image. Figure 2 reports data obtained assuming a radius comparable to the one of the smallest detectable cell. In addition, in this case, the correction of uneven illumination was marginal, as shown by the dot plots reporting the DNA content versus the cell spatial coordinates (Fig. 2, third and fourth rows, Rolling-Ball column). The low efficiency of the Rolling-ball algorithm in our experimental conditions may also be attributed to the high cell density, as neighboring cells could contribute to overestimate the background level.

The use of a slide-based background and a flat-field illumination correction in the Reference-based background subtraction (see “Materials and Methods”) provided an optimal linearity with good CVs (Fig. 2, Reference Column) and required minimal computing times with respect to the software-generated background.

Optimization of Acquisition Parameters for the Evaluation of DNA Content in High-Resolution Cytometry

High-resolution analysis in LSC and ISC is heavily limited by the maximal NA available for dry objectives. The use of standard fluorescence microscopes for cytometry can take advantage of a wide range of immersion objectives, with a resulting dramatic increase in the achievable optical resolution and an enhanced contrast thanks to the large amount of collected photons. However, the thin focal plane typical of a high NA objective can limit the correct measurement of DNA- or protein-content per cell. At the same time, an optimal spatial sampling is required for the acquisition of a diffraction-limited image; the correct pixel dimension, which is determined by the objective magnification, has to be set according to the Nyquist sampling theorem (64).

We thus compared linearity and CV of DNA distributions obtained by employing different magnifications and NA. For all the tested objectives, at fixed scanned area, the efficiency in cell recognition and in cluster resolution was comparable, as showed by the constant number of detected cells (data not shown). All the objectives provided DNA profiles with almost perfect G1/G2-peaks linearity, thus demonstrating a correct evaluation of the DNA content regardless of the depth-of-focus (Fig. 3). CV behavior presented dependence on both magnification and NA. It decreased when the increase in magnification was paralleled by an increased NA (by passing from the 20× to the 40× objective). When NA was maintained (40× and 60× 0.9 NA dry objectives), an increased magnification produced unchanged or even worse results. The exposure time was corrected to maintain comparable raw fluorescence intensities; however, the decreased contrast affected the CVs of the DNA distribution. A reduced contrast can decrease the efficiency of autofocus procedure, signal detection, and cell segmentation, with an impact on the correct DNA-histogram calculation.

Figure 3.

Influence of objective parameters on DNA distribution quality in high-resolution widefield image-cytometry. DNA profiles calculated from images collected employing different objectives showing how optical resolution and magnification influenced linearity and CV.

Strikingly, an increase in magnification and resolution, even at limited depth-of-focus, produced optimal results thanks to the higher contrast and number of pixels for each cell nucleus. The possibility to operate with oil-immersion objectives, maintaining a statistically valid sample size (at least 10,000 events per coverslip), introduces enormous advantages such as the possibility to employ fluorescence confocal microscopes at the maximal 3D resolution. Considering the large amount of generated data due to the axial sectioning procedure and the need of a correct evaluation of the integrated signal per cell, we analyzed the effect of Z-sampling and resolution on the DNA-content distribution of a cell population (Supporting Information Fig. 1). The pinhole aperture was varied maintaining the Z-slice spacing equal to a half of the theoretical slice thickness. The obtained profiles demonstrated that a reduced axial resolution and a low number of slices were still sufficient to reconstruct the DNA distribution. On the contrary, confocality conditions (pinhole aperture = 1 airy unit) could be challenging in maintaining optimal parameters for DNA-content measurements.

High-Resolution Cytometry Analysis of the Cell-Cycle Complements a Standard FCM Analysis

The increase in sensitivity and resolution, together with the improved correction of optical aberrations, allow exploiting the full range of wavelengths typical of white-light excitation (400-800 nm). To check for detection efficiency over different spectral ranges and to validate the calculated protein- and DNA-content distributions, we employed high-resolution image-cytometry to measure well-characterized cell-cycle-modulated targets.

Figure 4 shows the bivariate distribution of two cyclin proteins, namely cyclin E and A, during the cell-cycle. MCF10A exponentially growing cells were pulsed with EdU prior to fixation to mark actively replicating cells. The reported data demonstrated the possibility: (a) To excite infrared fluorochromes (CW800, see “Materials and Methods”) without employing tandem dyes; (b) to target each cell-cycle phase as occurring in FCM BrdU-based assays; (c) to gain intracellular imaging ability at a previously unreached resolution with such a statistical sampling. The multicolor gating procedure introduced in the software provided the expected reconstruction of the cell-cycle regulation for both cyclins (69). Cyclin E expression (Fig. 4, upper panel, second row) peaked at the G1-S transition, when the origins of replication were fired, as demonstrated by the discrete punctuate pattern of replication factories visible in the EdU channel (Fig. 4, lower panel, EdU row, G1/EarlyS and S/3 columns). As DNA replication progressed, cyclin E expression was switched-off with a concomitant rising of cyclin A (Fig. 4, upper panel, third row) which peaked during the G2M phase.

Figure 4.

Widefield high-resolution image-cytometry identifies DNA replication progression through analysis of incorporated EdU and correctly reproduces cell-cycle expression profile of cyclin E and A. Bivariate distribution of DNA content versus incorporated EdU signal was employed to target G1 (magenta), G1/early S (yellow, reduced EdU incorporation corresponding to transit from G1 to S and/or origin firing), first (S/3, green), second (2S/3, cyan) and third (3S/3, blue) stage of established S phase, lateS/G2 (violet, marked by reduced incorporation upon exit into G2) and G2M (red) stages of the cell cycle. Multicolor tagging of the selected populations in the cyclin E versus DNA content (second row) and cyclin A versus DNA content (third row) distributions revealed the cell cycle modulation of the expression of the selected cyclins. Representative images (lower panel) from the targeted cell subpopulations revealed the correlation between the S phase replicating pattern and the mean fluorescence of the analysed proteins. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

To demonstrate how image-retrieved information can complement a classic FCM assay, we analyzed the KI67 content and its intracellular distribution in dependence on the cell-cycle stage (Fig. 5). The measurement of the total amount of KI67 in relation to the DNA content (Fig. 5, first row) reproduced the standard FCM profiles (Supporting Information Fig. 2).

Figure 5.

Widefield high-resolution image-cytometry analysis complements classic cytometry analysis coupling expression level and intracellular distribution of KI67 proliferation antigen. Cell-cycle phases identification through DNA content versus incorporated EdU analysis (color coding is the one adopted in Fig. 4) has been employed to analyze distribution and localization of the KI67 antigen. Total expression level (second row) was compared to the number of detected intracellular regions (third row), size of the KI67 regions (fourth row) and EdU fluorescence concentrated in the detected areas (fifth row) demonstrating that cell-cycle modulation acted both at content and spatial localization level. Particularly, the peak of the replicated DNA contained in the KI67 regions in the late stages of the S-phase underscores their heterochromatic nature. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

A previous work by Endl and Gerdes (70) has shown that KI67 is mainly localized in the nucleoli during the interphase, progressively accumulates at chromosomes in mitosis and then moves back to the original compartment as soon as the nucleus is reformed. We measured the “KI67-spots” distribution and quantitatively described their spatial localization in thousands of cells simultaneously. The KI67-spots number peaked sharply (mean number 12.8) in G0/G1 and then reached a steady-state passing through the S up to the G2 phase; a second increase was detected in a less relevant fraction of G2 cells (Fig. 5, third row). The average physical size of KI67-spots per cell (obtained considering all the regions contained in a nucleus) revealed an opposite trend, with a progressive increase up to G2 (Fig. 5, fourth row). To confirm these observations, we analyzed the size distribution of KI67-spots considering them as an independent statistical population. KI67-spots were arbitrary classified into three classes of growing surface (Supporting Information Fig. 3, first row), and then the cells containing the gated spots were analyzed (since the software stored the cell of origin for every stored spot). All the cells contained KI67-spots of smaller to medium dimensions (from 100 to 1000 pixels corresponding to 1-10 μm2), regardless of the cell-cycle phase (Supporting Information Fig. 3, first and second columns). However, only 25% of G1 cells showed KI67-spots of larger average size (about 25 μm2), as the presence of an enlarged area was detected in the late cell-cycle stages (80% of G2 cells; Supporting Information Fig. 3, third column). These data, showing increasing average-size per cell and decreasing number of spots per cell, suggest the occurrence of a KI67-spots aggregation as the cell-cycle proceeds.

The identification of the KI67-subcompartment also allowed highlighting the heterochromatic nature of the nucleolar area. The plot of EdU fluorescence intensity localized in the intracellular KI67-positive regions showed a progressive increase towards the late stages of DNA synthesis (Fig. 5, blue population, fifth row). Thus, early replication of euchromatic regions only marginally overlapped the KI67 spatial distribution. However, when heterochromatin was opened for replication, the compartment dramatically enriched in EdU signals, corresponding to newly replicated DNA.

Among image-derived measurements, KI67 mean intensity per pixel per cell completed the total expression analysis, providing an estimate of the local concentration of protein per nucleus. The highest values of KI67 mean intensity per pixel per cell were detected in correspondence of two subpopulations with 2N and 4N ploidy (Supporting Information Fig. 4). The analysis of the cell area revealed that the tagged subpopulations were mainly formed by small cells. Moreover, the correlation between mean fluorescence per pixel per cell of both KI67 and DAPI revealed a concomitant up-regulation in the two channels, suggesting a tight link with DNA chromosome condensation. Hence, all the observed features described a phenotype that well correlates with the onset of mitosis. Since cell recognition was limited to the nuclear compartment, mitosis were distributed between 4N (till metaphase and first stages of anaphase) and 2N (from anaphase on) DNA-content. To validate this result, we employed the relocation ability of our system. The subpopulation with a high KI67 mean intensity per pixel per cell was identified after a first low-resolution acquisition. Physical coordinates were retrieved and every event was re-acquired at maximal 3D resolution, thus reconstructing the KI67 mitotic localization. Upon nuclear lamina destruction and chromosome condensation, KI67 decorated the surface of condensed chromosomes during all stages of mitosis, as shown in Supporting Information Figure 5. The high-resolution 3D data were supported by a robust statistical sampling.


Despite the continuous growth of automated solutions for high-resolution fluorescence microscopy, the application of this approach is frequently limited to the analysis of a small number of cells and the resulting low statistical sampling is not able to reproduce the biological heterogeneity evidenced by FCM, ISC, and LSC. However, the coupling of image-cytometry with high-resolution imaging can provide cytomics with new information about the inner cell compartments at and beyond the diffraction limit. We translated part of the vast repertoire of FCM analysis techniques into a novel image-cytometry approach. In particular, we generated novel applications to analyze cell-cycle events through widefield and/or confocal quantitative fluorescence microscopy. Two key strengths of our system are as follows: (i) The choice of the open-source software ImageJ as a development platform, which warrants a robust and continuously evolving environment for microscopy-related analysis, and (ii) the direct interface to different hardwares, which reduces the total time required for the procedure, making it possible to perform data acquisition and analysis in parallel. This feature also allows creating a communication interface with robotized fluorescence microscopes, enabling repeated acquisitions.

The high spatial sampling requested by the Nyquist sampling theorem imposes a restricted field-of-view, which is not compatible with a cell population analysis at an optically limited resolution. This limitation becomes dramatically relevant in the context of the “super-resolution” fluorescence microscopy (71–73). A step-by-step, image-analysis-driven acquisition procedure, can provide an important tool to progress towards an improvement in the spatial resolution for image-cytometry techniques.

In our approach, in a first round, data can be collected with a reduced resolution and/or content, thus preserving both storage and time, in order to identify the target events by a parallel image analysis. The throughput can be optimized according to the representative frequency of the targeted population. This initial selection has no limitation in terms of analysis, and can be even extended to extremely complex phenotypes (e.g., combined multiple fluorescence positivity, presence of specific intracellular structures); however, the higher the number of analyzed parameters are, the heavier the storage and computational efforts, thus increasing the execution time of the entire protocol.

Thanks to this in-silico identification, single events can be relocated, allowing the modulation of the achievable spatial-resolution. In this way, a diffraction-limited high-content image collection, that requires stringent conditions, can be selectively applied to the targeted population.

It is worth mentioning that many slide-based cytometry systems sacrifice spatial resolution to maximize the detection of other parameters. On the contrary our approach maintains the highest resolution, optimal sensitivity, and content by choosing a time-consuming acquisition, which is absolutely not comparable to FCM or ISC “exposure time” (e.g., the detection of a single event is in the microsecond range in FCM and ISC, while for the collection of a fluorescence image it spans from tens of milliseconds up to seconds). In our experience, data-collection can require up to several hours, according to the target spatial resolution and statistical sampling but gaining increased sensitivity and excellent spatial resolution.

Our approach can be easily extended by making available a large repertoire of specialized optical techniques (two-photon microscopy, in vivo live cell imaging, spectral and lifetime imaging, optical nanoscopy). Moreover, thanks to the wide compatibility with multiple platforms and the possibility to extend its features to specific hardware configurations, A.M.I.CO. can thus provide the trait-d'union between the development of high-resolution microscopy and image-cytometry.


The authors thank Massimo Scauso and Paolo Sapuppo for their support in system set-up and helpful discussions. MF is grateful to Wayne Rasband for the support in the software generation. They are indebted to Stefano Campaner, Ivan Dellino, Daniel Krueger, and Dario Parazzoli for critical discussions. They also thank Francesca Ballarini and Roberta Aina for critical revision of the manuscript.