key: cord-0801365-scd3f8vk
authors: Pape, Constantin; Remme, Roman; Wolny, Adrian; Olberg, Sylvia; Wolf, Steffen; Cerrone, Lorenzo; Cortese, Mirko; Klaus, Severina; Lucic, Bojana; Ullrich, Stephanie; Anders-Össwein, Maria; Wolf, Stefanie; Cerikan, Berati; Neufeldt, Christopher J.; Ganter, Markus; Schnitzler, Paul; Merle, Uta; Lusic, Marina; Boulant, Steeve; Stanifer, Megan; Bartenschlager, Ralf; Hamprecht, Fred A.; Kreshuk, Anna; Tischer, Christian; Kräusslich, Hans-Georg; Müller, Barbara; Laketa, Vibor
title: Microscopy-based assay for semi-quantitative detection of SARS-CoV-2 specific antibodies in human sera
date: 2020-10-07
journal: bioRxiv
DOI: 10.1101/2020.06.15.152587
sha: 09bffdda6d662e06607c033c457996bf703231fc
doc_id: 801365
cord_uid: scd3f8vk

Emergence of the novel pathogenic coronavirus SARS-CoV-2 and its rapid pandemic spread presents numerous questions and challenges that demand immediate attention. Among these is the urgent need for a better understanding of humoral immune response against the virus as a basis for developing public health strategies to control viral spread. For this, sensitive, specific and quantitative serological assays are required. Here we describe the development of a semi-quantitative high-content microscopy-based assay for detection of three major classes (IgG, IgA and IgM) of SARS-CoV-2 specific antibodies in human samples. The possibility to detect antibodies against the entire viral proteome together with a robust semi-automated image analysis workflow resulted in specific, sensitive and unbiased assay which complements the portfolio of SARS-CoV-2 serological assays. The procedure described here has been used for clinical studies and provides a general framework for the application of quantitative high-throughput microscopy to rapidly develop serological assays for emerging virus infections.

The recent emergence of the novel pathogenic coronavirus SARS-CoV-2 [1] [2] [3] and the 55 rapid pandemic spread of the virus has dramatic consequences in all affected countries. In the 56 absence of a protective vaccine or a causative antiviral therapy for COVID-19 patients, testing 57 for SARS-CoV-2 infection and tracking of transmission and outbreak events are of paramount 58 importance to control viral spread and avoid the overload of healthcare systems. The sequence 59 of the viral genome became publicly available only weeks after the initial reports on COVID-19 60 witnessed in the early phases of the ongoing SARS-CoV-2 pandemic. Thus, complementary 94 strategies to test for antiviral antibodies that can be rapidly deployed in situations where 95 commercially available kits are either not yet developed or not available are an important addition 96 to the diagnostic toolkit. 97 Immunofluorescence (IF) using virus infected cells as a specimen is a classical 98 serological approach in virus diagnostics and has been applied to coronavirus infections, 99 including the closely related virus SARS-CoV [14] [15] [16] . The advantages of IF are (i) that it does not 100 depend on specific diagnostic reagent kits or instruments, (ii) that the specimen contains all viral 101 antigens expressed in the cellular context and (iii) that the method has the potential to provide 102 high information content (differentiation of staining patterns and intensities due to reactivity 103 against various viral proteins). A mayor disadvantage of the IF approach as it is typically used in 104 serological testing is its limited throughput capacity due to the involvement of manual microscopy 105 handling steps and sample evaluation based on visual inspection of micrographs. Furthermore, 106 visual classification is subjective and thus not well standardized and yields only binary results. 107

Here, we address those limitations, making use of advanced automated microscopy and image 108 analysis strategies developed for basic research. We present the establishment and validation 109 of a semi-quantitative, semi-automated workflow for SARS-CoV-2 specific antibody detection. 110

With its 96-well format, semi-automated microscopy and automated image analysis workflow it 111 combines advantages of IF with a reliable and objective semi-quantitative readout and high 112 throughput compatibility. The protocol described here was developed in response to the 113 emergence of SARS-CoV-2, but it represents a general approach that can be adapted for the 114 study of other viral infections and is suitable for rapid deployment to support diagnostics of 115 emerging viral infections in the future. 116 2. Results 117 2.1 Setup of the IF assay for SARS-CoV-2 antibody detection 118 We decided to use cells infected with SARS-CoV-2 as samples for our IF analyses, 119 since this setup provides the best chance for detection of antibodies targeted at the different 120 viral proteins expressed in the host cell context. African green monkey kidney epithelial cells 121 (VeroE6 cell line) have been used for infection with SARS-CoV-2, virus production and IF [3, 17] . 122

In preparation for our analyses we compared different cell lines for use in infection and IF 123 experiments, but all tested cell lines were found to be inferior to VeroE6 cells for our purposes 124 (see Materials and Methods and Fig. S1 ). All following experiments were thus carried out using 125 the VeroE6 cell line. 126

In order to allow for clear identification of positive reactivity in spite of a variable and 127 sometimes high nonspecific background from human sera, our strategy involves a direct 128 comparison of the IF signal from infected and non-infected cells in the same sample. Preferential 129 antibody binding to infected compared to non-infected cells indicates the presence of specific 130 SARS-CoV-2 antibodies in the examined serum. Under our conditions, infection rates of ~40-131 80% of the cell population were achieved, allowing for a comparison of infected and non-infected 132 cells in the same well of the test plate. An antibody that detects dsRNA produced during viral 133 replication was used to distinguish infected from non-infected cells within the same field of view 134 ( Fig. 1A) . 135

In order to define the conditions for immunostaining using human serum, we selected a 136 small panel of negative and positive control sera. Four sera from healthy donors collected before 137

November 2019 were chosen as negative controls, and eight sera from PCR confirmed COVID-138 19 inpatients collected at day 14 or later post symptom onset were employed as positive controls. 139 Sera from this test cohort were used for primary staining, and bound antibodies were detected 140 using fluorophore-coupled secondary antibodies against human IgG, IgA or IgM. 141

No difference between infected and non-infected cells in serum IgG antibody binding was 142 observed when sera collected before the onset of the SARS-CoV-2 pandemic were examined 143 ( Fig. 1B, Fig. S2 ). In contrast, COVID-19 patient sera were clearly characterized by higher serum 144

IgG antibody binding to infected compared to non-infected cells (Fig. 1B) . All eight COVID-19 145 patient serum samples yielded higher IgG binding to infected compared to non-infected cells as 146 assessed by visual inspection (Fig. S2 ). Similar results were obtained when an IgA or IgM 147 specific secondary antibody was used for detection (Fig. S3) . In order to allow for the parallel 148 assessment of IgG and IgA or IgM antibodies, we established conditions for the parallel detection 149 of anti-IgG coupled to AlexaFluor488 and anti-IgA or anti-IgM coupled to DyLight650 or 150 AlexaFluor647 secondary antibodies, respectively, without signal bleedthrough. Using this 151 approach, it was possible to implement detection of SARS-CoV-2 specific IgG and IgA or IgM 152 antibodies in a single experimental setup (Fig. S4) . 153

Titration experiments were performed with positive control sera to determine the optimal 154 range of serum concentration in the IF experiments. All eight positive control samples showed 155 visually detectable specific labelling of infected cells over the range of 1:10 2 and 1:10 5 , 156 demonstrating robustness of the assay (Fig. S5 ). Serum concentrations of less than 1:10 5 did 157 not yield detectable signals in all cases. We decided to employ a dilution of 1:10 2 in the further 158 experiments 159 2.2 Image analysis 160 Our next aim was to establish a semi-automated analysis workflow for image acquisition and 161 analysis for a medium to high throughput setting. VeroE6 cells were seeded into 96-well plates 162 infected and immunostained using anti-dsRNA antibody and patient serum, followed by indirect 163 detection using a mixture of anti-IgG and anti-IgA/IgM secondary antibodies. Images were 164 acquired using an automated widefield microscope (see Materials and Methods section for more 165 detail). 166

To obtain a measure for specific antibody binding we performed automated 167 segmentation of cells and classified them into infected and non-infected cells based on the 168 dsRNA staining. We then measured fluorescence intensities in the serum channel per cell as a 169 proxy for the amount of bound antibodies for both infected and non-infected cells and calculated 170 the ratio between these values for infected and non-infected cells in a given specimen. To enable 171 training of a machine learning approach for cell segmentation and to directly evaluate infected 172 cell classification, we manually labelled cells and annotated them as infected/non-infected in 10 173 images chosen from 5 positive and 5 control specimens. Fig. 2 presents a graphical overview of 174 all analysis steps; the full description of every step can be found in Materials and Methods. 175

Briefly, our approach works as follows: 176

First, we manually discarded all images that contained obvious artefacts such as large 177 dust particles or dirt and out-of-focus images. Then, images were processed to correct for the 178 uneven illumination profile in each channel. Next, we segmented individual cells with a seeded 179 watershed algorithm [18] , using nuclei segmented via StarDist [19] as seeds and boundary 180 predictions from a U-Net [20, 21] as a heightmap. We evaluated this approach using leave-one-181 image-out cross-validation on the manual annotations and measured an average precision [22] of 182 0.77 +-0.08 (i.e., on average 77% of segmented cells are matched correctly to the 183 corresponding cell in the annotations). Combined with extensive automatic quality control which 184 discards outliers in the results, the segmentation was found to be of sufficient quality for our 185 analysis, especially since robust intensity measurements were used to reduce the effect of 186 remaining errors. 187

We then classified the segmented cells into infected and non-infected, by measuring 188 the 95th percentile intensities in the dsRNA channel and classifying cells as infected if this value 189 exceeded 4.8 times the noise level, determined by the mean absolute deviation. This factor and 190 the percentile were determined empirically using grid search on the manually annotated images 191 (see above). Using leave-one-out cross validation on the image level, we found that this 192 approach yields an average F1-score of 84. 3%. 193 In order to make our final measurement more reliable, we then discarded whole wells, To score each sample, we computed the intensity ratio : 199

Here, is the median serum intensity of infected cells and the median serum intensity of 200 non-infected cells. For each cell, we compute its intensity by computing the mean pixel intensity 201 in the serum channel (excluding the nucleus area where we typically did not observe serum 202 binding) and then subtracting the background intensity, which is measured on two control wells 203 that did not contain any serum. 204

We used efficient implementations for all processing steps and deployed the analysis 205 software on a computer cluster in order to enhance the speed of imaging data processing. For 206 visual inspection, we have further developed an open-source software tool (PlateViewer) for 207 interactive visualization of high-throughput microscopy data [23] . PlateViewer was used in a final 208 quality control step to visually inspect positive hits. For example, PlateViewer inspection allowed 209 identifying a characteristic spotted pattern co-localizing with the dsRNA staining ( Fig. S6) that 210 was sometimes observed in the IgA channel upon staining with negative control serum. In 211 contrast, sera from COVID-19 patients typically displayed cytosol, ER-like and plasma 212 membrane staining patterns in this channel (Fig. 1B, Fig. S3 ). The dsRNA co-localizing pattern 213 observed for sera from the negative control cohort is by definition non-specific for SARS-CoV-2, 214 but would be classified as a positive hit based on staining intensity alone. Using PlateViewer, 215

we performed a quality control on all IgA positive hits and removed those displaying the spotted 216 pattern colocalising with the dsRNA signal from further analysis. 217

With the immunofluorescence protocol and automated image analysis in place we 219 proceeded to test a larger number of control samples in a high throughput compatible manner 220 for assay validation. All samples were processed for IF as described above, and in parallel 221 analysed by a commercially available semi-quantitative SARS-CoV-2 ELISA approved for 222 diagnostic use (Euroimmun, Lübeck, Germany) for the presence of SARS-CoV-2 specific IgG 223 and IgA antibodies. 224

As outlined above, a main concern regarding serological assays for SARS-CoV-2 225 antibody detection is the occurrence of false positive results. A particular concern in this case is 226 cross-reactivity of antibodies that originated from infection with any of the four types of common 227 cold Corona viruses (ccCoV) circulating in the population. The highly immunogenic major 228 structural proteins of SARS-CoV-2 nucleocapsid (N) and spike (S) protein, have an overall 229 homology of ~30% [3] to their counterparts in ccCoV and subdomains of these proteins display a 230 higher degree homology; cross-reactivity with ccCoV has been discussed as the major reason 231 for false positive detection in serological tests for closely related SARS-CoV and MERS-CoV [12] . 232

Also, acute infection with Epstein-Barr virus (EBV) or cytomegalovirus (CMV) may result in 233 unspecific reactivity of human sera [24, 25] . We therefore selected a negative control panel 234 consisting of 218 sera collected before the fall of 2019, comprising samples from healthy donors 235 (n=105, cohort B), patients that tested positive for ccCoV several months before the blood 236 sample was taken (n=34, all four types of ccCoV represented; cohort A), as well as patients with 237

diagnosed Mycoplasma pneumoniae (n=22; cohort Z), EBV or CMV infection (n=57, cohort E). 238

We further selected a panel of 57 sera from 29 RT-PCR confirmed COVID-19 patients collected 239 at different days post symptom onset as a positive sample set (cohort C, see below). 240

Sera were employed as primary antisera for IF staining using IgM, IgA or IgG specific 241 secondary antibodies, and samples were imaged and analysed as described above. This 242 procedure yielded a ratiometric intensity score for each serum sample. Based on the scores 243 obtained for the negative control cohort and the patient sera, we defined the threshold separating 244 negative from positive scores for each of the antibody channels. For this, we performed ROC 245 curve analysis [26] [27] [28] on a subset of the data (cohorts A, B, C, Z). Using this approach, it is 246 possible to take the relative importance of sensitivity versus specificity as well as seroprevalence 247 in the population (if known) into account for optimal threshold definition. By giving more weight 248 to false positive or false negative results, one can adjust the threshold dependent on the context 249 of the study. Whereas high sensitivity is of importance for e.g. monitoring seroconversion of a 250 patient known to be infected, high specificity is crucial for population based screening 251 approaches, where large study cohorts characterized by low seroprevalence are tested. Since 252 we envision the use of the assay for screening approaches, we decided to assign more weight 253 to specificity at the cost of sensitivity for our analyses (see Materials and Methods for an in-depth 254 description of the analysis). Optimal separation in this case was given using threshold values of 255 1.39, 1.31 and 1.27 for IgA, IgG and IgM channels respectively (Fig. S7 ). We validated the 256 classification performance on negative control cohort E (n=57) which was not seen during 257 threshold selection, and detected no positive scores. Results from the analysis of the negative 258 control sera are presented in Fig. 4 and Table 1 . 259

While the majority of these samples tested negative in ELISA measurements as well 260 as in the IF analyses, some positive readings were obtained in each of the assays, in particular 261 in the IgA specific analyses ( Fig. 4 and Table 1 ). Since samples from these cohorts were 262 collected between 2015 and 2019, and donors were therefore not exposed to SARS-CoV-2 263 before sampling, these readings represent false positives. Of note, negative control cohort E 264 displayed a particularly high rate of false positives in ELISA measurements, but not in IF (Table  265 1). We conclude that the threshold values determined achieve our goal of yielding highly specific 266 IF results (at the cost of sub-maximal sensitivity). 267 Roughly 10.6% (IgA) or 3% (IgG) of the samples were classified as positive or 268 potentially positive by ELISA ( Table 1 ). The notably lower specificity of the IgA determination in 269 a seronegative cohort observed here is in accordance with findings in other studies [29, 30] accordance with other reports [30] [31] [32] . Consistent with other reports [32] , SARS-CoV-2 specific IgM 293 was not detected notably earlier than the two other antibody classes in our measurements. At 294 the earlier time points (up to day 14), a similar or higher proportion of positive samples was 295 detected by IF compared to ELISA for IgG. Although the sample size used here is too small to 296 allow a firm conclusion, these results suggest that the sensitivity of IgG detection by the semi-297 quantitative IF approach is higher than that of an approved semi-quantitative ELISA assay 298 routinely used in diagnostic labs. In the case of IgA detection at earlier time points (< day 11) 299 ELISA performed slightly better (11/17 samples scored positive) compared to IF (9/17 scored 300 positive) however that came with the price of a very low specificity of ELISA IgA assay (10.6% 301 false negative detection) compared to IF (0.5%). 302 3. Discussion 303 304

Here, we describe the development of a semi-quantitative IF based assay for detection 305 of SARS-CoV-2 specific antibodies in human samples that complements available ELISA-based 306 testing systems [33, 34] . Alternatives to ELISA-based commercial test kits are important in 307 situations where those kits are not available either because they are not yet developed in early 308 days of the pandemic or due to high global demands for tests and required reagents. The 309 microscopy-based assay described here has been developed during the early phase of the 310 COVID-19 pandemic to support the serological testing needs of the University Hospital 311

Heidelberg, Germany and is employed as a confirmatory assay in clinical studies [35] and ongoing 312 studies]. The assay displayed comparable or slightly better sensitivity and specificity than a 313 commercially available semi-quantitative SARS-CoV-2 ELISA approved for diagnostic use at the 314 time. More importantly, combining two technically different serological assays, IF and ELISA, 315 and classifying as "positive hits" only those that scored positive in both assays was instrumental 316 to minimize false positive results while maintaining high sensitivity, and thus serves as a principle 317 for serological studies or diagnostics where specificity of detection is of critical importance. 318

Specificity of detection is essential in settings of relatively low SARS-CoV-2 antibody prevalence 319 [36] [37] [38] in conjunction with high prevalence of potentially cross-reactive anti-ccCoV antibodies in a 320 global population [39] . 321

One advantage of the IF based assay presented here is that the specimens used for 322 detection present the entire viral proteome, while ELISA or chemiluminescent approaches use 323 a single recombinantly expressed antigen. Both the N and S protein of coronaviruses are highly 324 immunogenic, and antibodies binding to the receptor binding domain on the S1 subunit are 325 considered most relevant for neutralization. However, the relative importance of antibodies 326 directed against the N protein for potential protective immunity against SARS-CoV-2 and the 327 possible relevance of the overall breadth of the antibody response is currently unclear. Other 328 SARS-CoV-2 structural and non-structural proteins might play a role in immune response as it 329 was shown for proteins 3a and 9b of the closely related SARS-CoV [40] . In addition, expression 330 of the viral proteome in permissive cells ensures correct protein folding and post-translational 331 modification patterns. Alterations in post-translational modifications are likely to influence the 332 ability of serum antibodies to bind to different viral epitopes as it was shown for other viruses 333 such as HIV-1 [41] . It has to be noted that the detection of viral RNA requires fixation and 334 permeabilization of cells, which has the potential to affect epitope preservation. However, based 335 on the high sensitivity of antibody detection and the good correlation to ELISA measurements 336 observed we conclude that this was no major concern in this case. 337

Two major disadvantages of typical IF-based serological assays as applied in the past 338 are manual microscopy acquisition steps and evaluation of samples based on a visual 339 inspection. This procedure is incompatible with high throughput approaches and results are 340 subjective, not quantitative and difficult to standardize. We have addressed these disadvantages 341 by implementing automated microscopy acquisition and developing a robust software platform 342 that is able to identify individual cells, classify infected and non-infected cells and take into 343 account specific and non-specific background in order to generate semi-quantitative results. high-throughput application [42, 43] . Combining such cell lines with spectral unmixing microscopy 360 [44] would not only enable simultaneous determination of levels of all three major classes of 361 antibodies (IgM, IgG and IgA), but also identification of the viral antigens recognized, in a single 362 multiplexed approach. The high information content of the IF data (differential staining patterns) 363 together with a machine learning-based approach [45] and the implementation of stable cell lines 364 expressing selected viral antigens in the IF assay will provide additional parameters for 365 classification of patient sera and further improve sensitivity and specificity of the presented IF 366 assay. 367

The described analysis pipeline can be readily applied for serological analysis of other 368 virus infections, provided that an infectable cell line and a staining procedure that allows 369 differentiating between infected and non-infected cells are available. The assay described here 370 thus offers potential as an immediate response to any future virus pandemic, as it can be rapidly 371 deployed from the moment the first isolate of the pathogen has been obtained without requiring 372 information on the expression of immunogenicity of viral proteins. 

Two of our processing steps require manually annotated data: in order to train the convolutional 458 neural network used for boundary and foreground prediction, we needed label masks for the 459 individual cells. To determine suitable parameters for the infected cell classification, we needed 460 a set of cells classified as being infected or non-infected. We have produced these annotations 461 for 10 images with the following steps. First, we created an initial segmentation following the 462 approach outlined in the Segmentation subsection, using boundary and foreground predictions 463 from the ilastik [46] pixel classification workflow, which can be obtained from a few sparse 464 annotations. We then corrected this segmentation using the annotation tool BigCat 465 (https://github.com/saalfeldlab/bigcat). After correction, we manually annotated these cells as 466 infected or non-infected. Note that this mode of annotations can introduce two types of bias: the 467 segmentation labels are derived from an initial segmentation. Small systematic errors in the 468 initial segmentation that were not found during correction, could influence the boundary 469 Cell segmentation forms the basis of our analysis method. In order to obtain an accurate 488 segmentation, we make use of both the DAPI and the serum channel. First, we segment the 489 nuclei on the DAPI channel using the StarDist method [19] trained on data from Caicedo et al. 490 2019 [48] . Note that this method yields an instance segmentation: each nucleus in the image is 491 assigned a unique ID. In addition, we predict per pixel probabilities for the boundaries between 492 cells and for the foreground (i.e. whether a given pixel is part of a cell) using a 2D U-Net [20] 493 based on the implementation of Wolny et al. 2020 [21] . This method was trained using the 9 494 annotated images, see above. The cells are then segmented by the seeded watershed algorithm 495 [18] . We use the nucleus segmentation, dilated by 3 pixels, as seeds and the boundary predictions 496 as the height map. In addition, we threshold the foreground predictions, erode the resulting 497 binary image by 20 pixels and intersect it with the binarised seeds. The result is used as a 498 foreground mask for the watershed. The dilation / erosion is performed to alleviate issues with 499 very small nucleus segments / imprecise foreground predictions. In order to evaluate this 500 segmentation method, we train 9 different networks using leave-one-out cross-validation, 501

training each network on 8 of the manually annotated images and evaluating it on the remaining 502 one. We measure the segmentation quality using average precision [22] at an intersection over 503 union (IoU) threshold of 0.5 as described in https://www.kaggle.com/c/data-science-bowl-504 2018/overview/evaluation. We measure a value of 0.77 +-0.08 with the optimum value being 505 1.0. 506 507 Quantitation and Scoring 508

To distinguish infected cells from control cells we use the dsRNA virus marker channel: infected 510 cells show a signal in this channel while the non-infected control cells should ideally be invisible 511 (see Fig. 3 ). We classified each cell in the cell segmentation (see above) individually, using the 512 following procedure. First, we denoised the marker channel using a white tophat filter with a 513 radius of 20 pixels. To account for inaccuracies in the cell segmentation (the exact position of 514 cell borders is not always clear), we then eroded all cell masks with a radius of 5 pixels and 515 thereby discard pixels close to segment boundaries. This step does not lead to information loss, 516 since the virus marker is mostly concentrated around the nuclei. On the remaining pixels of each 517 cell, we compute the 0.95 quantile ( ) of the intensity in the marker channel. For the pixels that 518 the neural network predicts to belong to the background ( ), we compute the median intensity 519 of the virus marker channel across all images in the current plate. Finally, we classify the cell as 520 infected if the 0.95 quantile of its intensity exceeds the median background by more than a given 521 threshold: 522

For additional robustness against intensity variations we adapt the threshold based on the 523 variation in the background in the plate. Hence, we define it as a multiple of the mean absolute 524 deviation of all background pixels of that plate with N=4.8: 525

To determine the optimal values of the parameters used in our procedure, we used the cells 526 manually annotated as infected / non-infected (see above). We performed grid search over the 527 following parameter ranges: 528 In order to determine the presence of SARS-CoV-2 specific antibodies in patient sera, it was 560 necessary to define a decision threshold r*. If a measured intensity ratio r is above a decision 561 threshold r* than the serum would be characterized as positive for SARS-CoV-2 antibodies. For 562 this an ROC analysis was performed [28] . Each possible choice of r* for a test corresponds to a 563 particular sensitivity/specificity pair. By continuously varying the decision threshold, we 564 measured all possible sensitivity/specificity pairs, known as ROC curves (Fig. S7 ). To determine 565 the appropriate r* we considered two factors [26] : where is the prevalence or prior probability of disease. 573

The optimal decision threshold r*, given the false-positive/false-negative cost ratio and 575 prevalence, is the point on the ROC curve where a line with slope m touches the curve. As 576 discussed in the main text, a major concern regarding serological assays for SARS-CoV-2 577 antibody detection is the occurrence of false-positive results. Therefore, we choose m to be 578 larger than one in our analysis. In particular, we determine r* for the choice of m=10 (see Fig.  579 S7). 580

We performed quality control of the images and analysis results at the level of wells, images and 582 cells. The entities that did not pass quality control are not taken into account when computing 583 the score during final analysis. We exclude wells that contain less than 100 non-infected cells, 584 that have a median serum intensity of infected cells smaller than 3 times the noise level 585 (measured by the median absolute deviation), or that have negative intensity ratios, which can 586 happen due to the background subtraction. Out of 1.736 wells, 94 did not pass the quality control, 587 corresponding to 5.4 % of wells. At the image level, we visually inspect all images and mark 588 those that contain imaging artifacts using a viewer based on napari [49] . We distinguish the 589 following types of artifacts during the visual inspection: empty, unstained or over-saturated 590 images, as well as images covered by a large bright object. In addition, we automatically exclude 591 images that contain less than 10 or more than 1000 cells. These thresholds are motivated by 592 the observation that too few or too many cells often result from a problem in the assay. Thus, 593 296 of the total 15.624 images were excluded from further analysis, corresponding to 1.9 % of 594 images. Out of these, 295 were manually marked as outliers and only a single one did not pass 595 the subsequent automatic quality control. Finally, we automatically exclude segmented cells with 596 a size smaller than 250 pixels or larger than 12.500 pixels that most likely correspond to 597 segmentation errors. These limits were derived by the histogram of cell sizes investigated for 598 several plates. Two percent of the approximate 5.5 million segmented cells did not pass this 599 quality control. In addition, we have also inspected all samples scored as positives. For the IgA 600 channel, we have found a dotty staining pattern in ten cases that produced positive hits based 601 on intensity ratio in negative control cohorts, but does not appear to indicate a specific antibody 602 response. We have also excluded these samples from further analysis. 603

In order to scale the analysis workflow to the large number of images produced by the assay, 605

we implemented an open-source python library to run the individual analysis steps. This library 606 allows rerunning experiments for a given plate for newly added data on demand and caches 607 intermediate results in order to rerun the analysis from checkpoints in case of errors in one of 608 the processing steps. To this end, we use a file layout based on hdf5 [50] to store multi-resolution 609 image data and tabular data. The processing steps are parallelized over the images of a plate if 610 possible. We use efficient implementations for the U-Net [21] , StarDist [19] and the watershed 611 algorithm (http://ukoethe.github.io/vigra/) as well as other image processing algorithms [51] . We 612 use pytorch (https://pytorch.org/) to implement GPU-accelerated cell feature extraction. The 613 total processing time for a plate (containing around 800 images) is about two hours and thirty 614 minutes using a single GPU and 8 CPU cores. In addition, the results of the analysis as well as 615 meta-data associated with individual plates are automatically saved in a centralized MongoDB 616 database (https://www.mongodb.com) at the end of the workflow execution. Apart from keeping 617 track of the analysis outcome and meta-data, a user can save additional information about a 618

given plate/well/image in the database conveniently using the PlateViewer (see below images that correspond to three different ratio scores (ratio score is indicated above the image) 865 determined from samples stained with three different human sera, followed by staining with an 866 anit-IgG secondary antibody coupled to AlexaFluore488. Images represent overlays of three 867 channels -nuclei (blue), IgG (green) and dsRNA (red). White boxes mark the zoomed area. 868

Cells in the insets are highlighted with yellow or cyan boundaries, indicating infected and non-869 infected cells, respectively. Scale bar = 10 m. 870 

thank EMBL, especially the EMBL IT Services Department for providing computational 641 infrastructure and support, as well as Wolfgang Huber for discussions on computing image 642 based scores and statistical tests. We thank the patients who participated in this study

Berlin and the European Virus Archive (EVAg) for the 644 provision of the SARS-CoV-2 strain BavPat1. Individual images used in the Fig. 1A courtesy of 645 medical illustrations database

SB is supported by the Heisenberg program (project 650 number 415089553) and MLS is supported by the DFG (project number 41607209). The funders 651 had no role in study design, data collection

Bartenschlager 661 Microscopy development: S

Merle 666 Data interpretation

Laketa 668 Study design

All authors have read and approved the final version of the manuscript

MICCAI 2018. MICCAI

Van den Bogaard

Prevalence of COVID-19 in children in Baden-Württemberg Preliminary study report

Handb. Open Source Tools

We would like to thank Martin Weigert and Uwe Schmidt for their help with setting up prediction 638 for StarDist. We would like to acknowledge Infectious Disease Imaging Platform (IDIP) at Center 639 for Integrative Infectious Diseases Research (CIID) for microscopy support. We would like to