ABSTRACT
Q fever, caused by Coxiella burnetii, is a worldwide zoonotic disease that may cause severe forms in humans and requires a specific and prolonged antibiotic treatment. Although current serological and molecular detection tools allow a reliable diagnosis of the disease, culture of C. burnetii strains is mandatory to assess their susceptibility to antibiotics and sequence their genome in order to optimize patient management and epidemiological studies. However, cultivating this fastidious microorganism is difficult and restricted to reference centers, as it requires biosafety level 3 laboratories and relies on cell culture performed by experienced technicians. In addition, the culture yield is low, which results in a small number of isolates being available. In this work, we developed a novel high-content screening (HCS) isolation strategy based on optimized high-throughput cell culture and automated microscopic detection of infected cells with specifically designed algorithms targeting cytopathic effects. This method was more efficient than the shell vial assay, at the level of time dependency, when applied to both frozen specimens (7 isolates recovered by HCS only, sensitivity 91% versus 78% for shell vial) and fresh samples (1 additional isolate using HCS, sensitivity 7% versus 5% for shell vial), for which most strains were recovered more rapidly with the new technique. In addition, detecting positive cultures by an automated microscope reduced the need for expertise and saved 24% of technician working time. Application of HCS to antibiotic susceptibility testing of 12 strains demonstrated that it was as efficient as the standard procedure that combines shell vial culture and quantitative PCR.
INTRODUCTION
Coxiella burnetii is the causal agent of Q fever, a polymorphic disease that may occur as acute, mostly mild, and self-limiting forms or potentially severe persistent focalized infections, the main presentations being endocarditis and vascular infections (1–5). Although a large number of animals are able to carry C. burnetii, the main reservoirs of the bacterium are sheep, goats, and cattle (6). Q fever may cause outbreaks (7, 8), the largest to date having been recorded in the Netherlands (9). The laboratory diagnosis of C. burnetii infections relies mainly on serology and molecular biology (10). Such tools have greatly improved the diagnosis and management of patients, especially those developing a persistent focalized infection such as blood culture-negative endocarditis (1). Molecular detection assays can notably detect C. burnetii in clinical samples before seroconversion occurs. However, these methods cannot overcome the need for culture. Culturing C. burnetii is restricted to reference laboratories, as this bacterium requires cell culture, is classified as a risk group 3 microorganism, and must therefore be manipulated in biosafety level 3 laboratories. Moreover, it is highly contagious and can be infectious at the unit level. For these reasons, cultured strains remain scarce, limiting the access to antibiotic susceptibility tests, modern whole-genome sequencing for epidemiological studies, and research on virulence (11). However, the genome availability of isolated strains allowed the development of more efficient molecular detection tools (12). Therefore, isolating more strains remains crucial, and many strategies have been developed over time to cultivate C. burnetii (13). Nowadays, coculture remains the key tool for isolation. Reference centers and diagnostic laboratories have adopted the shell vial assay for the coculture of intracellular bacteria (14). C. burnetii was typically cultured in shell vials on human embryonic lung (HEL) cells, and detection was monitored every 10 days by immunofluorescence, Gimenez staining, and specific PCR (14, 15). However, this strategy remains subjective, tedious, time consuming, and operator dependent and has poor yield (29).
In this work, we revisited the isolation strategy of C. burnetii and brought improvements to two main axes: coculture and detection. We started by standardizing the coculture process of susceptible cell lines at many levels, such as culture medium, temperature monitoring, contamination control, cell viability, and proliferation monitoring. The detection process was then optimized using a fully automated system for high-content screening that was used in a previous study for the detection of giant viruses in protozoa (16). This new-generation microscope allowed the live monitoring of cocultures and large-scale image analysis, where specific algorithms were applied to detect any potential signs of infection, including cytopathic effects, morphological modifications, and vacuoles induced by C. burnetii, and predict cell phenotypes as being infected or healthy cells. After validating the proof of concept, a large-scale comparative screening of clinical samples from patients with acute or chronic Q fever was performed with both conventional shell vial and high-content screening strategies for C. burnetii isolation. Finally, this strategy was adapted for antibiotic susceptibility testing. This new strategy has proven to be more efficient and sensitive than the shell vial assay for the isolation of C. burnetii from clinical samples, with easier and quick manipulations associated with reduced subjectivity and thus a reduced need for highly experienced technicians.
MATERIALS AND METHODS
In the developmental stage, we targeted two main axes: the coculture process and the detection process. At the axis of coculture, modifications took place at many levels: cell line culture, microplates, cell concentrations, proliferation monitoring, and the coculture process. At the axis of detection, improvements were made at the levels of cell staining, the screening protocol, and data analysis to detect cytopathic effects induced by C. burnetii (vacuoles or cell burst), and finally, automation was introduced.
Coculture standardization.(i) Cell line selection. Two cell lines were used as cellular supports for coculture: human embryonic lung fibroblast MRC5 cells (RD-Biotech, Besançon, France) and mouse fibroblast L929 cells (ATCC CCL-1). The cell lines were cultured to confluence at 37°C under 5% CO2 in minimal essential medium (MEM) supplemented with 2 mM l-glutamine per liter and 10% or 4% heat-inactivated fetal bovine serum (FBS) for MRC5 and L929 cells, respectively (15). Cells were then harvested using a phenol red-free MEM culture medium, transferred into 96-well microplates at a volume of 200 μl per well, and incubated for 24 h to allow cell adhesion.
(ii) Choice of microplates. Different 96-well microplates were compared using the two cell lines mentioned above. The objective was to choose appropriate plates with the best compromise between image resolution, cell adhesion, and confluence as well as cell viability. We tested clear transparent plates with a thick polymer bottom (167008; Thermo Scientific), black plates with an optical bottom and cover glass base (164588; Thermo Scientific), and black plates with an optical bottom and polymer base (165305; Thermo Scientific).
(iii) Optimal cell concentrations. Different cell concentrations (105, 2 × 105, 4 × 105, and 106 cells/ml) were tested to determine the optimal concentration granting a confluent monolayer suitable for inoculation and monitoring. Cell viability and growth were monitored for 30 days, with the culture medium being changed every 10 days for MRC5 cells and every 7 days for L929 cells.
(iv) Cell proliferation monitoring. L929 cells have uncontrolled growth, which could interfere with the visualization of cytopathic effects. Therefore, different attempts were made to control their proliferation and maintain a single monolayer for 30 days. In contrast, MRC5 cells have contact inhibition of proliferation; therefore, no further controls were required.
(a) Culture medium composition. The first attempt to control the growth of L929 cells was to reduce the percentage of FBS added to the culture medium. Cell proliferation was monitored using a culture medium supplemented with 2% or 4% FBS.
(b) Cycloheximide addition. Another attempt was the addition of cycloheximide to the L929 cell monolayer, which inhibits the synthesis of nucleic acids and proteins of eukaryotic cells (17). Different concentrations, ranging from 0.05 to 1 μg/ml, were tested. Cell viability and proliferation rates were monitored for 30 days, and cycloheximide was added after every culture medium renewal. The effects of cycloheximide on C. burnetii infectivity and cell susceptibility to infection were assessed by quantitative PCR, Gimenez staining, and immunofluorescence to search for any inhibition or improved infection related to cycloheximide. We compared the results from cells infected with C. burnetii with and without cycloheximide addition to the coculture.
(v) Coxiella burnetii strains. Three strains of C. burnetii were used for infection: C. burnetii strain Nine Mile phase II (CB NMII), C. burnetii strain 223 (CB 223) and C. burnetii strain 227 (CB 227). All strains used in this study were obtained from CSUR-IHU (Méditerranée Infection, Marseille, France). Bacterial cells were produced on L929 cells, and quantification was performed by endpoint titration (50% tissue culture infective dose [TCID50]) on L929 and MRC5 cells 15 days postinfection.
(vi) Coculture process. We kept the same coculture strategy used in the shell vial assay for the isolation of C. burnetii (14) and we optimized it at many levels. Figure 1 summarizes the steps of the isolation strategy in the conventional shell vial assay and the new developed high-content screening (HCS) assay. Regarding the coculture process, cells were transferred to 96-well microplates and were incubated for 24 h prior to infection. Supernatant was then removed, and infection was carried out with 50 μl of CB NMII, CB 223, and CB 227 diluted up to 10−10 (the initial bacterial load was at 107 particles/ml). Similarly to the shell vial assay, low-speed centrifugation of plates (700 × g for 1 h at 22°C) was performed to enhance the attachment and the penetration of the bacteria inside the cells. The final volume was adjusted to 250 μl with culture medium. Uninfected cells were considered a negative control.
Traditional and improved coculture and detection processes adopted for the isolation of C. burnetii. This figure details all the steps involved in processing a clinical sample using conventional shell vial analysis and the new optimized high-content screening technique.
Detection process optimization.The workflow for the detection process development is summarized in Fig. 2.
Workflow of the detection process development and optimization at the software level. (A) Steps performed in the HCS Studio software for image and data generation. (B) Main steps in the data analysis script development in R Studio. (C) Minimized application for automated analysis.
(i) DNA staining of cells. NucBlue Live ReadyProbes reagent (Molecular Probes, Life Technologies, USA) was used as a live-cell DNA stain. Staining was performed by the direct addition of NucBlue to the cells without washing. Different concentrations were tested, and the minimal concentration granting sufficient staining and lowest cell toxicity was adopted for each cell line. Cell viability and aspects were monitored for any cytotoxic effects related to staining.
(ii) Screening protocol. Image acquisition and analysis were performed at 10, 20, and 30 days postinfection using the automated CellInsight CX7 High-Content Analysis platform (Thermo Scientific) allowing real-time acquisition and on-the-fly multiparametric analysis. Acquisition parameters were defined in the HCS Studio 3.1 software (Thermo Scientific) using the Morphology Explorer Bio-Application. This later provides quantitative measurement of morphological and texture-related features at the single-cell level, intracellular level, and multicellular level. Autofocus parameters and exposure times were adjusted so that the fluorescent or optical density signal reached 50% of the dynamic range of the 14-bit camera. The nuclear fluorescent probe NucBlue (386 nm) was used to perform software-based autofocus and served as a primary mask for single-cell detection and quantification. The secondary channels consisted of brightfield images with adapted z offset to collect morphological information. The region of interest (ROI) consisted of an enlarged Voronoi diagram derived from the nuclear mask to include the cytoplasm, thus limiting possible overlap with neighboring cells. Cell aggregates and debris were excluded from analysis using area cutoffs. Image acquisition was performed using a 20× lens objective (0.45 numerical aperture [NA]), and 80 images or fields were generated per well in a way to cover 90% of the well surface. We extracted intensity-, texture-, and morphology-related information from the region of interest and exported a data set of 148 features for every cell.
(iii) Data analysis. The developed algorithms described in this article were carried out on MRC5 cells infected with 3 strains of C. burnetii: CB NMII, CB 223, and CB 227. Cells infected with C. burnetii exhibit a particular phenotype due to vacuole formation. Therefore, we used the extracted data set to detect cytopathic effects and differentiate between infected and uninfected cells. The exported file was uploaded in a dedicated application developed in R Studio using the user interface ShinyR. We generated a database of labeled cell data coming from uninfected and infected cells. Outliers were removed from this data set, and 6,000 cells were kept as training data. Using radar graphs and principal-component analysis (PCA), we screened all 148 features and identified key features that distinguish infected cells (positive control) and uninfected cells (negative control). Two K-means clusters, representing infected and uninfected cells, were calculated using the generated training data and key features. These clusters were then used to predict the phenotype of the experimental data set using a semisupervised K-mean clustering algorithm (from flexclust R package, kccaFamily “k-median”). A preliminary data sorting was performed based on the total cell count per well in order to detect wells showing cell burst and prevent false-negative results. All wells with fewer than 5,000 cells were excluded from the assay and marked as “not applicable” (NA) in the final result. The prediction algorithm was then applied to the remaining data set. This algorithm associates a phenotype for each cell depending on the cluster it falls in; therefore, percentages of infected and uninfected cells per well can be calculated. Finally, we defined a threshold for positivity based on the percentage of infected cells per well.
(iv) System automation. After validation, we coupled the CellInsight CX7 microscope with an automation system consisting of a robotized incubator Cytomat 2C-LIN (Thermo Scientific) and a plate handler Orbitor RS microplate mover (Thermo Scientific). Incubation times, plate handling, and acquisition protocols were monitored through the Thermo Scientific Momentum 5.0.5 software. The script for data analysis was also automated using the user interface ShinyR in order to make the application user friendly and interactive. An automated application, HCS Data Clustering, was created, where all the steps (cell data import, data filtering and normalization, clustering of training data, prediction of experimental data, and results generation) were automatically performed.
Comparison of the detection process with gold standard methods.The immunofluorescence assay and the manual quantification of infected cells were adopted as reference methods in order to validate the results.
(i) Immunofluorescence assay. We kept the immunofluorescence assay as a gold standard method for the validation of results. The same protocol previously described for the detection of C. burnetii (14) was optimized for microplates. Imaging was performed on the CellInsight CX7 microscope, and the entire well was screened at ×20 magnification.
(ii) Manual quantification of infected cells. We then performed a manual quantification of infected cells using brightfield images. We quantified the number of vacuoles visible to the naked eye and then calculated the percentage of infected cells per well for each strain as follows: percentage of infected cells = (number of vacuoles/total cell count) × 100. Results were then compared to the HCS results. Quantifications were performed by 2 different operators.
Proof-of-concept validation: artificial samples.To test the system’s efficacy toward clinical samples, we artificially spiked one blood sample and one serum sample having negative PCR toward C. burnetii, with CB NMII at different concentrations (stock, 10−3 and 10−6 dilutions). Coculture was then performed on the MRC5 cell line as described above. Fifty microliters was used for inoculation, and cells were rinsed twice with culture medium after the centrifugation step. Two negative controls were included: uninfected cells and cells cocultured with the nonspiked samples. Cocultures were then monitored every 10 days on the automated CellInsight CX7 microscope, and results were validated by immunofluorescence.
Applicative stage: high-content screening assay versus shell vial assay.(i) Screening of clinical samples. For the applicative stage, we performed a comparative study on 90 clinical samples from our sample collection using the traditional shell vial assay technique (14) and the new optimized HCS technique. A large variety of samples was tested, including 13 blood samples, 24 valve samples, 7 biopsy specimens, 3 thrombus samples, 6 aneurysm samples, 4 abscesses, 29 serum samples, 3 articular fluid samples, and 1 tick sample. The first group consisted of 47 frozen samples from which different strains of C. burnetii were previously isolated after primary inoculation upon reception at the laboratory. The second group consisted of 43 fresh samples tested positive by PCR as previously described (12) and inoculated prospectively. Cocultures were performed on the MRC5 cell line, as summarized in Fig. 1, simultaneously in shell vials (14) and in microplates for HCS. Regarding the shell vial assay, 3 shell vials were inoculated for each sample and monitored under a light microscope for cytopathic effect detection at 10, 20, and 30 days postinfection. The results were then validated by immunofluorescence, Gimenez staining, and specific PCR (14, 15). Subcultures were then performed as previously described by Raoult et al. (14). As for the HCS strategy, 5 wells were inoculated for each sample and monitored at the same time points on the CellInsight CX7 microscope. The results were validated by immunofluorescence. After 30 days, negative cocultures were subcultured in 96 -well microplates containing a fresh monolayer of cells and then monitored weekly using the same strategy. We then compared the results from both strategies regarding isolation rate and culture delay.
(ii) Application for MIC testing. We adopted the same principle developed by Angelakis et al. (18) for antimicrobial susceptibility testing. We tested the MIC of two antibiotics used in the treatment of C. burnetii: doxycycline and hydroxychloroquine (18), using both the conventional shell vial strategy and the new HCS technique. Twelve strains of C. burnetii were tested: CB 109S, CB 196, CB 226, CB 228, CB 242, CB 244A, CB 248, CB 249A, CB 250, CB 252, CB 260, and CB Henzerling. Regarding the HCS strategy, strains were cultured in 96-well microplates containing a monolayer of MRC5 cells with serial 2-fold dilutions of doxycycline (0.25 to 8 μg/ml) and hydroxychloroquine (0.25 to 4 μg/ml). Uninfected cells treated or not with the highest concentrations of antibiotic were used as negative controls, and the positive control consisted of infected cells without any antibiotic treatment. Each test was performed in quadruplicates, and results were assessed 15 days postinfection by HCS for cytopathic effect detection. In parallel, the standard MIC testing was performed in shell vials, and results were assessed by quantitative PCR as previously described (18).
Statistical analysis.The R Studio software was used to perform all statistical tests included in strategy development and data analysis. P values were calculated to search for significant differences between the two isolation strategies.
Ethical statement.According to the procedures of the French Commission for Data Protection (Commission Nationale de l’Informatique et des Libertés), collected data were anonymized. The study was approved by the local ethics committee of IHU (Institut Hospitalo-Universitaire)—Méditerranée Infection.
RESULTS
Coculture standardization.(i) Culture medium and microplates selection. The use of a transparent culture medium without phenol red indicator minimized the autofluorescence that could interfere with the imaging process. All 96-well microplates tested showed adequate cell adhesion. However, plates with a cover glass base were less suitable for prolonged culture durations (see Fig. S1 in the supplemental material). On the other hand, black plates with an optical bottom were better for imaging than clear plates with a thick polymer bottom, as photobleaching was minimal and a better resolution was obtained, especially on brightfield images. Therefore, we adopted the black plates with the optical bottom and polymer base to be used for coculture.
(ii) Optimal cell concentrations. Different cell concentrations were tested for each cell line to determine an optimal concentration granting a confluent monolayer for 30 days without cell overgrowth. The optimal concentrations were at 4 × 105 cells/ml and 2 × 105 cells/ml for MRC5 and L929 cells, respectively (Fig. 3).
Cell densities of L929 and MRC5 cell lines at different initial concentrations 24 h into culture. (a to d) Brightfield images of L929 cells at 106, 4 × 105, 2 × 105, and 105 cells/ml, respectively. (e to h) Respective concentrations for MRC5 cells. Bars, 100 μm.
(iii) Cell proliferation monitoring. Contrary to that for MRC5 cells, which have contact inhibition of proliferation, L929 cells showed uncontrolled cell proliferation, and aggregates started forming 3 days into culture (see Fig. S2d to f). Moreover, when infected with C. burnetii, cytopathic effects or vacuoles were difficult to visualize due to the high cell density (Fig. S4b). Therefore, control of cell overgrowth was a must to maintain a single monolayer of cells for the longest period.
(a) Culture medium composition. We started by reducing the percentage of FBS added to the culture medium from 4% to 2% to check if cell proliferation would be slower. However, no significant change in the proliferation rate was observed, and cells became very dense starting 3 days into culture (Fig. S3).
(b) Cycloheximide addition. We tested a wide range of cycloheximide concentrations and searched for cytotoxicity or cell mortality while monitoring the proliferation rate. High concentrations exhibited extensive cytotoxic effects on cells and induced rapid cell mortality. The optimal concentration was 0.25 μg/ml for an initial cell concentration of 2 × 105 cells/ml. It is important to note that cycloheximide was only added to cells after the 24-h period allowing cell adhesion. However, although the proliferation rate was lower, it was not completely inhibited; therefore, we found it necessary to increase the cycloheximide concentration to 0.5 μg/ml after culture medium renewal at days 7, 14, and 21. This strategy allowed us to maintain a single monolayer of L929 cells for 30 days with no significant toxicity or mortality (Fig. S4).
Moreover, no significant difference in infectivity was observed between cells treated or not with cycloheximide in terms of C. burnetii infectivity and/or L929 cell susceptibility to infection (Fig. S4 and S5). Immunofluorescence and Gimenez images showed similar infection states in cells treated or not with cycloheximide. The same was observed by PCR quantification, where the bacterial multiplication rate appeared to be the same. However, cytopathic effect visualization was not possible in the absence of cycloheximide, where high cell density masked the vacuoles formed by C. burnetii (Fig. S4b and c).
Optimized detection process.(i) Cell DNA staining. Regarding DNA staining, several concentrations of NucBlue were tested, and the optimal concentrations granting sufficient staining were 4 ng/ml and 2 ng/ml for MRC5 and L929 cells, respectively, for the predetermined cell concentrations. This corresponds to 10 and 5 μl per well, respectively, directly added from the stock solution. Note that NucBlue is a live-cell stain and was directly added to culture without cell wash. Cell aspects and viability were monitored by microscopy to search for any cytotoxicity related to staining. We noticed that prolonged contact with cells induced nuclear fragmentation, morphological modifications, and, eventually, cell mortality (see Fig. S6). To overcome this problem, staining was performed a few hours before screening, and stained wells were only considered exploitable during the following 24 h.
(ii) Screening protocol. The image acquisition and analysis protocol was developed in the HCS Studio software to extract the maximum data from the region of interest in brightfield images. MRC5 cells infected with serial dilutions of CB NMII, CB 223, and CB 227 were then screened at 10, 20, and 30 days postinfection using the CellInsight CX7 microscope. Screening time was found to be around 3 min per well, where autofocus, image acquisition, and algorithm application on generated images were simultaneously performed. Eighty fields were screened per well, and 4 images were generated in each field, the first consisting of the nucleus fluorescence image, followed by 2 brightfield images, and finally, the overlay image. Cell data containing intensity, texture, and morphology information were extracted from generated images as a .csv file and used in the following step of the analysis.
(iii) Data analysis. Cell data extracted from MRC5 cells infected with CB NMII, CB 223, and CB 227 were used for data analysis and database generation. A database of negative and positive controls was generated to be used as training data. We selected data from different time points of the infection. Positive controls were selected from wells where approximately 50% of the cells were infected. Highly infected cultures are often in advanced states of cell death and do not resemble early stages of infection; thus, it was better to use data from images with a moderate infection rate (∼50%) and to find a clustering that meets this value while leaving the negative train data as close to zero as possible. We then identified 4 key features that distinguish well between the negative and the positive controls using radar graphs and principal-component analysis: nuclear average fluorescence intensity per cell, skewness (the levels of asymmetry of the brightfield intensity distribution around the mean within the region of interest), kurtosis (the levels of peakedness or flatness of the brightfield intensity distribution within the region of interest), and finally, the ratio of the variation intensity over the average intensity of the brightfield within the region of interest (ObjectAvgIntenCh1, ROI_SkewIntenCh3, ROI_KurtIntenCh3, and Var_Avg.IntensityRatio, respectively). The latter feature was calculated to compensate for the loss of illumination at the well edges, a phenomenon known as the vignette effect, observed in ROI_VarIntensityCh3 (the standard deviation of intensities in the region of interest). Training data normality was assessed in a QQ plot. Data were then filtered accordingly, and outliers were removed to ensure a normal distribution. Using the 4 key features and the training data, 2 clusters were generated using the K-means clustering algorithm. These clusters represent uninfected and infected cells Due to experimental variability, training data were rescaled so the mean values of negative training data equaled the mean values of untreated cells for each feature (untreated cells being the negative control of the experimental data set to be predicted). The clustering algorithm predicted a baseline of 6% to 7% infected cells in the negative training data and a value of 50% to 60% infected cells in the positive training data. These values are the agnostic result generated by the clustering algorithm. The predicted baseline of infected cells in the negative control is due to the presence of artifacts, such as debris or dead cells, which could interfere with the results. Therefore, any prediction below this baseline was considered a negative result. We empirically determined a threshold of 10% for positivity. However, any data between 7% and 10% were systematically checked for possible false negativity. The percentages of infected and uninfected cells per well of the experimental data set were then predicted using the generated clusters. Results were then represented as a color-coded heat map showing the percentage of infected cells (Fig. 4A). All values below the baseline of 7% are represented in white (negative result), values between 7% and 10% are represented in green (suspected), and values >10% are represented in red (positive result). However, wells showing cell burst are prone to false-negative results. Therefore, wells with fewer than 5,000 cells were excluded from analysis and marked as NA in the heat map (gray color).
Results from the prediction algorithm and validation references of MRC5 cells infected with CB NMII at 10, 20, and 30 days postinfection. (A) The heat map represents the percentages of infected cells obtained with the prediction algorithm of MRC5 cells infected with serial dilutions of CB NMII. (B) Respective immunofluorescence images of the negative control (a) (well B1 in the heat map), the false-positive result (b) (well D9), a positive result (c) (well B3), and a false-negative result (d) (well B5). (C) The table represents the manual quantification results from the same experiment. (D) Respective fluorescence and brightfield images of the negative control (e, f), cells at an advanced stage of infection (g, h), and slightly infected cells (vacuoles are indicated with red arrows) (i, j). Bars, 100 μm.
(iv) Automated system. The automation system allowed the systematic handling of several plates, where plates are transported by the robotic arm from the incubator to the microscope and vice versa (https://www.mediterranee-infection.com/acces-ressources/donnees-pour-articles/plate-handler-orbitor-rs-microplate-mover). Momentum software supervised the incubation time of each plate and synchronized the appropriate acquisition protocols developed in the HCS Studio software. This fully automated system allows minimal handling of plates by the operator and thus reduces the risk of cross contamination. Regarding data analysis, all steps were automatically performed in the automated application HCS Data Clustering. The time required for analysis was less than 1 min from data import to result generation.
Detection process validation.Plates infected with 3 strains of C. burnetii were used in the developmental stage for algorithm optimization. Prediction results were generated as color-coded heat maps and were validated by immunofluorescence as well as by a manual quantification of infected cells (Fig. 4 and Table S1). Figure 4A shows an example of the generated heat map of MRC5 cells infected with CB NMII at 10, 20, and 30 days postinfection. Regarding manual quantification, the number of vacuoles visible to the naked eye was quantified on brightfield images, and the percentage of infected cells was then calculated. Wells showing advanced stages of infection were difficult to quantify and were noted as uncountable (unc). Note that manual quantification was very difficult to perform and was highly time consuming. We obtained values close to the expected percentages but not as accurate due to quantification difficulties. However, manual quantification helped detect false-positive and false-negative results. We detected 0.68% false-negative results and 7.08% false-positive results. False-negative results were predicted below the 7% baseline, and further investigation showed a very low infection rate, where only a few vacuoles were detected on the generated images and by immunofluorescence. Furthermore, false-positive results were predicted as suspected or positive (>7%) when no vacuoles were detectable. False positivity can be due to several reasons, such as high cell density, dead cells, or debris. Table S1 summarizes the results of the prediction algorithm by HCS versus the manual quantification for each strain. Different infection profiles were observed for each strain, where the infection was positive up to 10−7 dilution for CB NMII, 10−5 dilution for CB 223, and 10−6 dilution for CB 227 30 days postinfection.
Proof-of-concept validation: artificial samples.The prediction algorithm was as efficient with clinical samples as with the pure bacterial culture in the detection of cytopathic effects. However, more false-positive results were observed due to debris coming from samples (Fig. S7).
Applicative stage: high-content screening assay versus shell vial assay.(i) Screening of clinical samples. Among the first group of 47 frozen samples from which C. burnetii was previously isolated, the isolation ratios were 37/47 (78.7%) for the conventional shell vial assay and 43/47 (91.5%) for the HCS assay. Regarding the second group of 43 prospectively inoculated specimens, isolation ratios were 2/43 (4.7%) and 3/43 (7%), respectively. Results are shown in Fig. 5 and Table S2. Overall results for the conventional shell vial assay and the HCS assay were of 39/90 (43%) and 46/90 (51%), respectively. The majority of the isolated strains originated from valves samples. Moreover, 32% of the strains were isolated faster with the HCS than the shell vial technique, where 4 strains were recovered at 10 days postinoculation, 4 strains after 20 days, 7 strains after 30 days, and 1 strain after 40 days, in addition to 6 strains solely isolated by HCS at 30 days and 1 strain at 90 days. Furthermore, 28% were isolated at the same time points with both techniques (12 strains), and only 23% were isolated faster with the shell vial technique (11 strains, including 7 isolated after 10 days and 4 isolated after 20 days) (Fig. 5). Note that 95% of strains isolated with the HCS assay were recovered before 30 days postinfection compared to 65% with the shell vial assay. These results show a significantly higher time-dependent efficiency with the new HCS strategy than with conventional methods (P value of 0.002), where 7 strains were recovered from different clinical samples solely using the HCS assay, and most strains were isolated faster with the HCS strategy. Furthermore, we compared the operating time required for each step of the process with both strategies on 20 clinical samples with 5% positivity rate; we observed a greater time consumption (24%) during manipulations with the shell vial assay (25 h) than with the HCS assay (19 h) (see Table S3).
Comparative results between the shell vial assay and the HCS assay for the isolation of C. burnetii from clinical samples. This interaction scheme shows all positive samples isolated using the two strategies, the samples’ origins (nodes), and their coculture delay (lines).
(ii) Application for MIC testing. We obtained similar MIC results for the 12 C. burnetii strains tested with both techniques, where the HCS strategy detected cytopathic effects due to C. burnetii multiplication and the conventional technique quantified bacterial multiplication by quantitative PCR. All strains were found to be sensitive to doxycycline, with an MIC of 0.25 μg/ml, except the CB Henzerling strain, which was resistant up to a concentration of 4 μg/ml. Regarding hydroxychloroquine, all strains were resistant, with small variations in MIC values, with strains CB 226 and CB 242 having an MIC of 4 μg/ml, and strains CB 109S, CB 196, CB 228, CB 244A, CB 248, CB 249A, CB 250, CB 252, CB 260, and CB Henzerling having an MIC of >4 μg/ml.
DISCUSSION
Over the past decades, major questions regarding intracellular bacteria such as C. burnetii started to be resolved after the isolation and the proper identification of many strains (1, 3, 19, 20). Currently, the rapid diagnosis of Q fever is made possible by various culture-independent tools such as serology, molecular biology, and histology (10, 21, 22). However, culturing the bacterium remains crucial, as assessments of its infectivity, tropism, and virulence may only be obtained from isolates (23). Recently, many attempts have been made for axenic culture of C. burnetii on agar plates or in cell-free liquid medium (24, 25). Although this approach was successfully used for primo isolation and to propagate established strains, cell culture currently remains the reference method for isolation (14, 15). Therefore, updates and improvements to cell culture by introducing novel technologies were mandatory. To this end, we developed a new isolation strategy starting from culture standardization to detection process optimization through an automated imaging platform and data analysis for cell phenotype prediction.
We observed that uncontrolled cells can affect susceptibility to infection and complicate the detection of the pathogen. As cell proliferation is not controlled, the multilayers will mask the detection of cytopathic effects or vacuoles. Many samples were falsely negative, with positive PCR results but no signs of infection detectable under the microscope. This is common and usually associated with poor sample transport and conservation, susceptibility of the bacterium, and previous antibiotic therapy. Therefore, monitoring cell concentrations and proliferation and choosing an adequate culture medium are critical factors for a more efficient coculture. It is also important to avoid temperature fluctuations that can stress the cells and cause their derivation, i.e., cancer cell lines are not the best choice for optimal culture; therefore, primary cell lines would be a better choice for future applications. In addition, the use of microplates instead of shell vials for coculture had many advantages, as coculture and immunofluorescence were performed in sealed microplates, which protected the culture as well as the manipulator from contamination. Another advantage was the analysis of the immunofluorescence in wells, as we can overlay results with brightfield images, which is more quantitative and less risky than with shell vials. Moreover, the management of samples in microplates is better, since we can culture up to 18 samples in a single plate, which is equivalent to 57 shell vials. Manipulations are therefore easier and incubators are less crowded.
Regarding the axis of detection, scanning in microplates did not change the area of screening corresponding to that in shell vials. In addition, we noticed that screening of the whole well by the robotic microscope is robust and explores the totality of the surface, whereas observation of the shell vial under an inverted microscope is more difficult, requires expertise, and covers only a small part of the surface with less resolution. Therefore, our new HCS system showed a higher isolation rate in fewer time points compared to that with the conventional shell vial assay, in addition to higher sensitivity and specificity, as well as reduced subjectivity. It is important to note that the choice of samples was dependent on their availability, and we managed to isolate C. burnetii even though samples were frozen, which could result in the loss of bacteria. A small rate of false-negative results (0.68%) was observed in cases where the infection was very low, and false-positive results (7.08%) depended on the cell status and the amount of debris present in the well. Nevertheless, this risk is easily corrected by immunofluorescence and specific PCR in suspected samples.
The introduction of a diverse panel of cell lines for the isolation of C. burnetii could increase the efficiency of the system by making it possible to isolate strains with different affinities for different cell lines. Such variations in affinity for C. burnetii have already been described in the literature (26, 27). This system would also be applicable for tropism and virulence assessment.
Using a panel of cell lines for isolation would be easy with this automated system, which allows the processing of several plates at the same time under incubation with reduced cross-contamination risks. In addition, this method does not require any expertise other than carrying out the coculture, since the selection process and the extraction of the results are automated, and any biologist, student, or technician can retrieve the data.
We successfully applied this new strategy to study the antimicrobial susceptibility of C. burnetii which can replace the conventional PCR technique, since this method is more feasible, economic, and faster than PCR. In addition, PCR results do not always reflect the number of infectious particles. Note that the prevalence of antibiotic resistance among C. burnetii strains is low, with only one strain so far reported to be doxycycline resistant among the 270 human isolates in our collection (28), and the assessment of antibiotic susceptibility of newly isolated strains therefore remains a prevention method and a surveillance tool to detect any potential emerging resistance.
Finally, this high-content screening method was based on semisupervised deep learning and algorithms, which can be updated and optimized for the detection of other intracellular bacteria as well as human viruses.
ACKNOWLEDGMENTS
This work was supported by a grant from the French State managed by the National Research Agency under the Investissements d’avenir (Investments for the Future) program under reference no. ANR-10-IAHU-03 (Méditerranée Infection) and by the Région Provence-Alpes-Côte-d’Azur and the European funding FEDER PRIMI.
We thank the Thermo Fisher team for their technical support on the high-content analysis platform and the automation system.
Maxime Mioulane is an employee of Thermo Fisher Scientific.
FOOTNOTES
- Received 17 December 2019.
- Returned for modification 18 January 2020.
- Accepted 26 February 2020.
- Accepted manuscript posted online 4 March 2020.
Supplemental material is available online only.
- Copyright © 2020 American Society for Microbiology.
