A versatile, automated and high-throughput drug screening platform for zebrafish embryos

Zebrafish provide a unique opportunity for drug screening in living animals, with the fast developing, transparent embryos allowing for relatively high-throughput, microscopy-based screens. However, the limited availability of rapid, flexible imaging and analysis platforms has limited the use of zebrafish in drug screens. We have developed an easy-to-use, customisable automated screening procedure suitable for high-throughput phenotype-based screens of live zebrafish. We utilised the WiScan® Hermes High Content Imaging System to rapidly acquire brightfield and fluorescent images of embryos, and the WiSoft® Athena Zebrafish Application for analysis, which harnesses an Artificial Intelligence-driven algorithm to automatically detect fish in brightfield images, identify anatomical structures, partition the animal into regions, and exclusively select the desired side-oriented fish. Our initial validation combined structural analysis with fluorescence images to enumerate GFP-tagged haematopoietic stem and progenitor cells in the tails of embryos, which correlated with manual counts. We further validated this system to assess the effects of genetic mutations and x-ray irradiation in high content using a wide range of assays. Further, we performed simultaneous analysis of multiple cell types using dual fluorophores in high throughput. In summary, we demonstrate a broadly applicable and rapidly customisable platform for high-content screening in zebrafish. The 96-well plates were imaged using the WiScan® Hermes High Content Imaging System (IDEA BioMedical, Rehovot, Israel). Images were taken in brightfield and green and red fluorescence channels at a 4x magnification with a well coverage of 150% and field density of 150%. This gave 4 overlapping images along each well. Brightfield was imaged with 35% light intensity, 40 ms exposure and 30% gain. Fluorescence channels were imaged with 90% light intensity, 200 ms exposure and 30% gain, except for Tg(kdrl:mCherry) which was imaged with 50ms exposure in the fluorescent channel. Z-stacks were taken in 5 planes with an inter-plane distance of 50.6 µm.


Introduction
screens (Chiu et al., 2008, Vlasits et al., 2012, Esterberg et al., 2013, Pei et al., 2018, however all still rely on manual counting or scoring of individual fish, creating a significant bottleneck in analysis throughput. Zebrafish have also emerged as a valuable patient-derived xenograft model for drug screening for novel chemotherapeutic compounds or to identify patient-specific responses, having several advantages over mouse models including far more rapid time scales, cost effectiveness and the ability to use large numbers (Xiao et al., 2020, Fazio and Zon, 2017, Fior et al., 2017, Jung et al., 2012, Tonon et al., 2016. Zebrafish xenografts have been used successfully to identify or validate chemotherapeutic compounds, for example to validate BPIQ against lung cancer cells (Chiu et al., 2015) and the identification of regorafenib against adenoid cystic carcinoma (Chen et al., 2017).
Although there have been a number of examples of how zebrafish could provide a useful highthroughput screening platform for xenograft models (Haney et al., 2020, Zhang et al., 2014, Lin et al., 2019, Fu et al., 2018, Somasagara et al., 2021, the use of these on large scale has so far been limited. A platform with greater flexibility and ease of customisation would greatly aid the use of xenograft models for drug screening, due to the variation and complexity between different models, and different read-outs for results for example tumour size (Zhang et al., 2014) or cell count (Somasagara et al., 2021).
The optimal zebrafish high-throughput screening platform would permit automation of both image acquisition and analysis across a range of multiplexed assays and phenotypes with minimal human intervention. One of the primary challenges in this process is getting the embryo in the desired orientation for imaging without manual manipulation of the embryos. One approach has been to use small glass capillaries for imaging embryos, as in the VAST Bioimager which allows for automated imaging of zebrafish embryos in a chosen orientation (Pulak, 2016, Pardo-Martin et al., 2010, Pardo-Martin et al., 2013, Chang et al., 2012, Haney et al., 2020. Zebrafish embryos can also be imaged in a semi-automated format by standard microscopes in 96-well plates (Romano and Gorelick, 2014), although there is limited control over the orientation of the fish without manual manipulation or inspection of the images.
Existing automated image analysis solutions are much more limited, with most platforms developed as bespoke solutions. One popular technique is to design algorithms to automatically analyse fluorescent images of transgenic zebrafish, or fish fluorescently labelled with a dye. A number of such screens have adapted the ImageXpress High Content Screening System by Molecular Devices for automated image acquisition. This has included assays to quantify the number of angiogenic blood vessels (Tran et al., 2007), to analyse axon length (Kanungo et al., 2011) and to measure with four overlapping images along each well to permit accurate image stitching and full-fish visualisation (Fig. S1). Image acquisition was carried out in brightfield and fluorescent channels, taking around 15 minutes to obtain the 3,840 raw images per full plate, allowing for the relatively high throughput. Image pre-processing steps are carried out automatically by the accompanying batch image processing software package (IDEA Bio-Medical, see Methods) in around 20 minutes.
We performed best z-slice selection (most in-focus z-plane) for the brightfield channel and maximum z-axis intensity projection for fluorescence channels prior to image stitching for full-fish analysis (Fig. S1), with images processed in batches by accompanying software. Fluorescence max-intensity was chosen over best-slice selection to capture all fluorescently labelled cells throughout the fish volume. In this fashion we were able to visualise individual fluorescently labelled cells without the need for additional processing or image deconvolution while also having both the tail and head of the fish in proper focus for analysis of the brightfield channel.
IDEA Bio-Medical developed a novel Zebrafish image analysis application for their WiSoft® Athena software package that identifies the fish and internal anatomy with no required user input. This analysis application uses a unique deep-learning AI algorithm to analyse the brightfield images of full zebrafish (see Methods), trained utilizing hundreds of zebrafish images as input, in conjunction with multiplexed fluorescence channel analysis. 13 separate datasets were used to train the algorithm, with a further 5 datasets used for testing, to ensure the quality of the trained AI for detecting features of naive data. The software was trained to use brightfield images to automatically detect the outline of zebrafish embryos, while concomitantly identify internal anatomical structures and regions (Fig. 1B). Currently, the internal anatomy detected includes: eye, heart, tail fin, yolk sac, spine, bladder and otic vesicle (Fig. 1B). The embryo body is also subdivided into the head, trunk and tail regions. Each segmented region can be analysed for features such as morphology (i.e. eye size) or count (i.e. count 1 vs. 2 eyes to determine orientation), and is combined with fluorescence quantification including intensity metrics and identification of labelled structures, such as fluorescent cells or granules (Fig. 1C).
This analysis approach also permits the automatic selection of wells containing, exclusively, the desired side-oriented fish (illustrated in Fig. 1D), without manual image inspection. Although the use of the alignment plate improves the number of fish in the correct orientation for analysis (Fig. 1A,B), a small number of fish in each plate are not properly aligned in the well (lying on their back, tail out of viewing window etc., (Fig. S2)). This is usually <5% per plate. The side-orientated fish were selected using anatomical attributes identified by the Athena software, specifically those having both an eye count and a tail count equal to one, thereby excluding empty wells and fish in undesirable orientations. Defining a minimum fish and tail area in the software also excludes wells where the fish is only partially in the well (see Methods). Automatic exclusion of these wells without any manual inspection of the images is a critical feature for high-throughput applications and avoids skewing statistics through incorrect or biased sample selection. While these improperly oriented fish could be correctly oriented via manual intervention, the required hands-on time far exceeds that needed to prepare, scan, process and analyse an additional 96 well plate and results in a lower total amount of useful data.
The rapid image acquisition combined with batch image processing and automatic detection of the fish and its anatomy allows for high-throughput analysis with minimal user input. The variety of features detected and the rapid, easy-to-use customisability of the application makes this platform a highly versatile tool for a number of screening applications, as we detail below.

Accurate counts HSPCs in the tail of 2-4dpf zebrafish embryos
The primary goal of our screen is to accurately enumerate HSPCs, using the Tg(itga2b:GFP) reporter line ( Fig. 1C). At 3dpf the HSPCs reside in the caudal haematopoietic tissue (CHT, analogous to fetal liver) in the base of the tail and in close proximity to the caudal-most tip of the yolk sac (Fig. 1D). We aimed to develop a screening platform that would provide an automatic count of the number of cells exclusively in this region to allow us to analyse the effects of myelodysplastic syndrome (MDS)-related genetic mutations that have been identified in the HSPCs of patients. Ultimately such a tool will enable a large-scale small-molecule library screen to target mutant stem cells, to identify drugs that exclusively deplete the mutant HSPC population. One of the challenges in automating our assay is that zebrafish at 3dpf have substantial auto-fluorescence, particularly around the head and yolk (Fig. 1C). The HSPCs of interest are the GFP lo cells, whilst the GFP hi cells are thrombocytes, some of which are in circulation (Lin et al., 2005) . This meant that simply thresholding out the autofluorescence in the whole image erroneously eliminated a large number of cells of interest, or the inclusion of fluorescent spots not relevant to our question. The Athena zebrafish application can provide an automatic count of the fluorescent granules in a chosen region of the fish, in our case the tail (Fig. 1C, outlined in red), without manual inspection or segmentation of the images. Analysis of a 96-well plate takes around 10 minutes. To verify the counts measured, we analysed a 96-well plate of wild-type transgenic zebrafish at 3dpf with the Athena software and compared results to a manual cell count in the same images. The two counting methods were strongly correlated, r=0.844 ( Fig. 1E), indicating that automated detection provides equivalent results to manual counting.
Stem cells first migrate to the CHT around 2dpf where they then start to proliferate (Chen and Zon, 2009). We tested our platform on embryos of different ages to see the accumulation of HSPCs in the CHT over time ( Fig. 2A). Using our automated HSPC counting assay, we confirmed that small numbers of HSPC are present in the CHT at 2 dpf and this steadily increases over the next 48 hours ( Fig. 2A-B). In this experiment different batches of embryos were used for each age. However embryos are imaged alive and can easily be removed from the plate with a pipette, permitting longitudinal studies on the same embryo if desired. We also attempted to image embryos at 5dpf but due to the inflation of the swim bladder, very few fish aligned in the desired side-orientation and our analysis could no longer be carried out with the desired throughput. The system could be used to image younger embryos, although this would require the use of a different plate, such as a round bottomed plate, due to the shape of the embryo and the fragility of the yolk sack to centrifugation.
Currently the software is trained to identify fish 2-6dpf, although with additional training datasets it could be trained to identify younger fish.

Detection of phenotypic differences in HSPC and myeloid cell count
We have developed several models for MDS in Zebrafish, including an Rps14 mutant line (Peña et al., 2020). In this model, heterozygous embryos are phenotypically indistinct from wildtype animals unless the animals are subjected to stress. Phenylhydrazine (phz) is used as a haemolytic stress in zebrafish, causing oxidation of haemoglobin and anaemia (Lenard et al., 2016, Shafizadeh et al., 2004, Ferri-Lagneau et al., 2012. If phz is applied for 24 hours at 24-48hpf, only the wildtype embryos recover from the induced anaemia (Schneider et al., 2016, Peña et al., 2020. Using our HSPC counting assay, we showed that at 3dpf, phz leads to an increase in HSPCs at 3dpf in the wild types in response to anaemia, a response not observed in the Rps14 mutants (P < 0.01, Fig To further define the sensitivity of the platform we assessed the effects a range of non-lethal doses of X-ray irradiation on HSPC numbers in the embryo (Traver et al., 2004, McAleer et al., 2005. Embryos were irradiated at 2dpf before imaging at 3dpf. This led to a significant reduction in stem cells with 40 Gy and a further reduction with 100 Gy (Fig. 3A). The platform allows for swift analysis of these images in large numbers, and the phenotypic differences are easily measured with high statistical significance (P < 0.001, Fig. 3B).
Beyond analysing solely HSPCs in our MDS models, we aimed to develop a versatile screening platform that we could easily extend to other assays including other zebrafish transgenics to label different cell types. We performed multiplexed fluorescence image acquisition using a double colour transgenic line with Tg(itga2b:GFP) and Tg(lyzC:mCherry), which has mCherry-tagged myeloid cells (Buchan et al., 2019), for imaging in both the red and green channels (Fig. 3C). We used this fish line in the X-ray irradiation assay and identified both the GFP-tagged HSPCs and mCherry-tagged myeloid cells in the same fish (

Extended applications using fluorescent outputapoptosis, hair cell detection and angiogenesis
We wished to further assess the utility of our platform beyond haematological-related screens, to other assays that are commonly used for drug screening in zebrafish embryos, showing the broad applicability to a variety of screening applications. Apoptosis assays in zebrafish can be utilised for drug screens, including analysis of regulated and induced apoptosis and for assessing toxicity (McGrath andSeng, 2013, Parng et al., 2004). Acridine orange is a fluorescent apoptosis marker that can easily be used to visualise cell death in zebrafish by incubation of live embryos in the stain followed by fluorescence imaging (Abrams et al., 1993, Furutani-Seiki et al., 1996, Tucker and Lardelli, 2007. Rapid automated quantification of this fluorescence could be a valuable tool for drug screening and toxicity assessment of chemical screens. To test if our screening platform could be adapted to assess cell death, X-ray irradiation at 2dpf was used to induce apoptosis in zebrafish embryos before analysis at 3dpf. Irradiation in zebrafish embryos leads to apoptosis, with particularly sensitivity towards cells within the spinal cord (Geiger et al., 2006). We labelled dying cells post irradiation with the supravital stain acridine orange and imaged on the Hermes in high throughput (Fig. 3F). The Athena analysis pipeline was optimised by adjusting the threshold, smoothing and area parameters of the granule detection, to count the small stained granules which define the dying cell population in the whole fish. The increased cell death with irradiation was easily observed (P < 0.0001, Fig 3G).
Assays analysing hair cells in the lateral line in zebrafish have also been utilised in drug discovery for compounds that mediate deafness and ototoxicity due to the similarities between these cells and the inner ear hair cells in humans (Whitfield, 2002, Nicolson, 2005, Ton and Parng, 2005. Hair cells in the lateral line are arranged into neuromasts along the head and body of the zebrafish and can be stained using fluorescent marker YO-PRO-1. This has been utilised in several chemical and genetic screens (Chiu et al., 2008, Vlasits et al., 2012, Esterberg et al., 2013, Pei et al., 2018, Owens et al., 2008. To date these screens have required manual evaluation of the neuromasts either by counting or scoring for degradation. We therefore sought to replicate the results of one of these screens using  , 2007). At 48hpf the embryos were imaged on the Hermes, with both compounds leading to reduced angiogenesis (Fig. 4D). Athena was then utilised to identify areas of fluorescence, and the total area of the fluorescence within the whole fish used as a surrogate of the degree of vessel formation. As expected, both AG1478 and SU4312 led to a significant reduction (Fig. 4E-F, P < 0.0001), with the latter causing a more significant reduction at lower concentrations, consistent with the lower IC 50 reported previously (Tran et al., 2007). A fibre identification algorithm could be introduced to Athena, to provide a more direct read out of angiogenesis.

Brightfield phenotypic analysiseye size
In addition to fluorescence assays, we assessed if our platform could be adapted for morphological screens in brightfield. Mab21l2 is involved in healthy eye development, and zebrafish carrying homozygous mab21l2 mutations display microphthalmia (smaller eyes) (Gath and Gross, 2019, Deml et al., 2015, Hartsock et al., 2014. Using an incross of mab21l2 u517 heterozygous embryos (Wycliffe et al., 2020), imaged in brightfield with the Hermes at 4dpf, we used Athena to identify they eye in brightfield and measure the size. Consistent with previous studies, mab21l2 u517 homozygous are smaller in size (Fig. 4G), as measured using Athena (P < 0.05, Fig. 4H).

Discussion
Zebrafish embryos provide an excellent opportunity for in vivo screens, but often such screens rely on time-consuming manual assays or the development of bespoke automated solutions that require time and expertise to develop. Utilising the Hermes and Athena HCS platform from IDEA Bio-Medical offers an automated solution for zebrafish screens, which is both versatile and easy to use.
Image acquisition is straight-forward to set-up, with simple plate loading and imaging. The plates are loaded in the lab with a standard pipette with no special treatment of the embryos required beyond the use of the alignment plate. Imaging is easy to carry out -the plate is placed in the microscope and the parameters chosen using user-friendly software, where changes can be visualised interactively before imaging. Imaging parameters can be saved and loaded to provide consistency between images acquired at any time. Unlike the VAST system which offers 360° imaging of embryos in a capillary (Pulak, 2016, Pardo-Martin et al., 2010, Pardo-Martin et al., 2013, Chang et al., 2012, the Hermes microscope images the fish from beneath the plate. This allows for side-on multiplexed images of fish at 2-4dpf ( Fig. 2A) in multiple fluorescence colours (Fig. 3C). For many applications this is sufficient, and although not all embryos are perfectly aligned in the wells with our method, this is limited to around 5% per plate and the software parameters permit automatic exclusion of these wells without manual inspection of the images.
The Athena Zebrafish application allows for easy, automated analysis of the acquired images. We have demonstrated one of the principle advantages of this system is the ease with which customisation to different assays can be undertaken. Following set up with the HSPC assay, each of the subsequent optimisations shown in this manuscript were completed in a single experiment. We It is worth noting that a potential significant barrier for the use of this system is the cost, with both the microscope and the analysis packages being potentially out of reach of some laboratories. The flexibility across a large range of assays means the system can be both multi-purpose and multi-user allowing it to be shared between groups. The Hermes microscope and Athena packages both have applications beyond zebrafish, with many applications in cell biology, which does gives potential for shared use across multiple laboratories and departments, and the opportunity to share the cost.
Zebrafish provide a unique opportunity for drug screening in whole live animals since they produce large numbers of fast developing, transparent embryos. The development of an automated screening platform that is fast, customisable and easy to use without specialist knowledge or training will allow for much better utilisation of this unique system. As shown in this manuscript, the platform can be applied to a large number of different assays across a large range of research topics, with new assays being straightforward to develop using the user-friendly environment. The speed of acquisition and analysis allows access to larger compound libraries, the consistency between experiments allows for comparison of results across multiple experiments, and the ease of use makes the platform accessible to all. The training set used to identify fish and anatomy are centralized and continuously updated/maintained which also provides reproducibility and consistency between research groups. This platform will allow for high-content screening in zebrafish, as well as being useful for more targeted experiments by allowing for the use of large numbers of embryos to give statistical power.
The ease of use, speed of imaging and analysis, and consistency between images acquired at different times makes this platform useful for both small and large-scale experiments. We now plan to use this platform for a small-molecule drug screening, using the automated HSPC count presented here to conduct a synthetic lethal screen in an MDS-related mutant background.

Zebrafish Husbandry and Experimental Conditions
Zebrafish (Danio rerio) stocks were maintained according to standard procedures in UK Home Office approved aquaria (Westerfield, 2007). Embryos were obtained from wild-type AB or AB/TL,

Preparation of Embryos for Imaging
At 24hpf, embryos were dechorionated using pronase, and raised in 0.003% phenylthioureasupplemented E 3 medium to prevent pigment formation. At 2-4 dpf embryos were anaesthetised with tricaine, and then loaded into a 96-well zebrafish alignment plate (Hashimoto ZF plate, Japan, 96-well) in 75 µL of the E 3 medium using a wide orifice tip. The plates were then gently centrifuged at 200 rcf for 20 seconds before imaging.

WiScan® Hermes Image Acquisition
The 96-well plates were imaged using the WiScan® Hermes High Content Imaging System (IDEA Bio-Medical, Rehovot, Israel). Images were taken in brightfield and green and red fluorescence channels at a 4x magnification with a well coverage of 150% and field density of 150%. This gave 4 overlapping images along each well. Brightfield was imaged with 35% light intensity, 40 ms exposure and 30% gain. Fluorescence channels were imaged with 90% light intensity, 200 ms exposure and 30% gain, except for Tg(kdrl:mCherry) which was imaged with 50ms exposure in the fluorescent channel. Zstacks were taken in 5 planes with an inter-plane distance of 50.6 µm.
An accompanying image pre-processing software package (Advanced Data Processing software, IDEA Bio-Medical, Rehovot, Israel) was used to process the raw images of multiple data sets in batch prior to quantitative analysis. This software loads one or more image data sets obtained on the WiScan® Hermes microscope, then performs one or more sequential image processing operations in batch.
Image processing operations can include intensity projection (through time or Z-slices), fluorescence deconvolution, sharpest (most in-focus) Z-plane selection, and image montage (stitching). Here, the sharpest z-slice for each raw field of view in the brightfield channel was selected, and maximum intensity projection for each raw field of view in the fluorescence channel was performed.
Subsequently, all fields of view (sharpest brightfield and maximum fluorescence) in a well were stitched together to provide a single, 2-colour channel image per well.

WiSoft® Athena Image Analysis
Image quantification was performed with the WiSoft® Athena software Zebrafish Application (IDEA Bio-Medical, Rehovot, Israel). The software application performs automated, multiplexed image analysis by processing brightfield and fluorescence channels simultaneously, with different algorithms. Brightfield analysis of the whole zebrafish embryo utilized a novel deep learning based artificial intelligence algorithm to identify the fish and internal structures. The method is based on a supervised convolutional neural network and was trained using hundreds of images of 2-4 dpf zebrafish embryos as input, each paired with manually segmented ground-truth outputs. An option for custom, manual segmentation is also present in the Zebrafish application. The resulting algorithm automatically identifies the fish outline, internal anatomical structures and three fish regions (head, trunk, and tail). Users input only minimum and maximum size constraints to select the appropriate range for anatomical objects/regions that are to be identified and analysed. Fluorescence channels are analysed using image analysis techniques with optimized algorithms (smoothing, background Biology Open • Accepted manuscript subtraction, thresholding and area constraints, Table 2) to identify high signal-to-background objects, such as fluorescent granules.
The morphological features (length, area, shape, etc.) of the fish, organs and regions are quantified based on the structures identified in the brightfield channel. Objects (i.e. fluorescent granules) identified in the fluorescence channels that are contained within the fish outline are included in quantification, while spots outside the fish are disregarded. Fluorescent spots are quantified regarding their count and intensity within the fish as a whole, as well as within anatomical or regional structures (e.g. granule count within the tail, spot intensity within the spine, etc.).
Parameters were set to define the permitted size of detected fish and internal features in 2-4dpf embryos (Table 1). No additional parameter input was required to identify these objects, since the AI algorithm was trained to identify them automatically. A population of on-side orientated fish for analysis was defined as those with an eye count and a tail count equal to one.

Statistical Analysis
For each experiment the desired data was exported from Athena in csv format, for example granule in the tail count for HSPC count experiments, fluorescence area or eye size, for individual embryos.
Statistical analysis was performed using GraphPad Prism v9.0.2. The probability level for statistical significance was P < 0.05 and unpaired two-tailed t-tests were used for analysis. Data are presented as mean ± standard deviation.

Genotyping Mutants
For Rps14 and mab21l mutants, genotyping was required post imaging. Embryos were removed in 50 µL of nuclease-free water using a multi-channel pipette and wide orifice tips and added directly to 1 µL 50x HotSHOT base solution (KOH 1.25 M, EDTA 10 mM). After incubation at 95°C for 30 minutes the solution was neutralised with 1x HotSHOT neutralising solution (40 mM Tris-HCl). To identify the genotype of Rps14 and mab21l mutants we used the Kompetitive Allele Specific PCR genotyping assay (KASP, LGC Genomics).     Fin 15000 300000

Figures
Bold text shows parameters that have been adjusted from the default For fluorescence analysis, granules were identified within the Athena Zebrafish Application by defining the fluorescence threshold, area of the granules, background subtraction and smoothing.
Parameters were customised for each assay (Table 2).