Challenging specificity of chemicalcompounds targeting GPCRs with cellprofiling

(1)

Challenging specificity of chemical

compounds targeting GPCRs with cell profiling

Anton Davidsson

Degree project inbiology, Master ofscience (2years), 2020 Examensarbete ibiologi 60 hp tillmasterexamen, 2020

Biology Education Centre and Pharmaceutical Bioinformatics, Uppsala University

(2)

Abstract

Screening compounds with image-based analysis is an important part in the process of drug discovery. It is an efficient way to screen compounds as it gives more

information than for example HTS. High-content screening as it is also called, has really progressed in recent years, as the field of data science evolves, and with it so does the efficiency of how images can be processed into information. Another important part of the drug discovery field is the family of receptors GPCRs, a large family of over 800 different receptors in humans. The reason GPCRs are important in drug discovery is because of the large number of drugs targeting them. In this experiment we wanted to use image-based analysis to challenge drugs or

compounds that were said to be specific and see if they actually are that specific, or if we can see indications of the drug also working somewhere else. While the drugs we tested did not appear to cause any morphological perturbations large enough to distinguish them from the control, some drugs appear to cluster differently. This might suggest that they affect multiple targets, but it needs to be followed up upon in order to draw any substantial conclusions.

List of abbreviations

cAMP - cyclic adenosine monophosphate CBCS - Chemical Biology Consortium Sweden DMEM – Dulbecco’s modified eagle medium DMSO – dimethyl sulfoxide

DHM - digital holographic microscopy FBS – fetal bovine serum

GPCR – G-protein coupled receptor HCS - high content screening

HTS - high throughput screening MoA - mode of action

PBS – phosphate-buffered saline PCA - principal component analysis PC - principal component

PFA - paraformaldehyde PI - phosphatidylinositol PSF - point spread function QPI – quantitative phase imaging RPM – revolutions per minute WGA - wheat germ agglutinin

(4)

1. Introduction

1.1 Microscopy and screening

Over the past years, fluorescence microscopy has become an essential tool in cell biology, allowing visualization of cells and cell organelles using multiplexed dyes.

The advances in fluorescence microscopy, image analysis, and automated screening technologies have created opportunities to image cell states, extract quantitative information from images, and study changes in cell morphology induced by chemical compounds, drugs or environmental factors. However, studying live cells with fluorescent dyes still remains a challenge, mainly because the

fluorescence labels might be cytotoxic in the long run and thus influence the cell’s morphology, also fluorescent dyes might photobleach with time and limit the time for imaging. An alternative to fluorescence microscopy of live cells is Digital Holographic Microscopy (DHM) (Kühn et al. 2013). It’s a rather new technique, suitable for

experiments lasting up to several days with living cells that should allow quantitative data to be extracted from live cells (Rappaz et al. 2014).

1.1.2 High Throughput Screening

There are many different approaches to discovering new drugs in drug research, depending on the target of the drug, the disease or even the researcher. More often than not, the whole process combines different approaches. For example, there are ways of looking into possible interactions, or hits, between a compound and different targets. This approach usually involves different machine learning models, for

example using the chemical structures of the compound and protein sequence to predict these interactions (Lan et al. 2016). The follow-up could then be to screen these hits, thousands at a time, and look for leads among them in so-called High Throughput Screening (HTS) (Mayr & Bojanic 2009). Screening such a large volume of compounds is expensive and time consuming and therefore it is usually only done by companies, using automated labs, able to produce lots of data in a short period of time. The reason that compounds are run in such a large fashion is because in order for a hit to become a lead, it must succeed in multiple different assays. With the thousands and thousands of hits that are being screened, the success rate differs a lot depending on type of assay and target (Fox et al. 2006, Bender et al. 2008).

1.1.3 Cellular morphology, profiles and benchmarking

High content screening (HCS) is a method similar to HTS, but where the focus is on cell morphology instead. Cells are treated with compounds and images are

generated. An example of a method developed to help handle and analyse the large amounts of data produced in high content screening is by cell profiling. This can be done by labeling different organelles with fluorescent markers and capturing images

(5)

with different wavelengths, usually one wavelength per fluorescent marker. When analyzing these images, there will be similarities between the cells based on their morphological features, such as the size and shape of the cells and their organelles.

These similarities are what define a cell’s wild type profile (Bray et al. 2016). This is usually done using a software, to segment the different compartments of the cell, thus giving them numerical properties based on pre-set parameters (Carpenter et al.

2006). The wild type profile can be used in other assays, such as genotypic or

metabolic assays, by comparing it with cells that are phenotypically different from the wild type. Treating cells with compounds can be useful in making profiles based on perturbations that correspond to the compounds mode of action (MoA) (Caicedo et al. 2017). By profiling cells exposed to a compound with an unknown MoA and comparing it to profiles of cells exposed to a compound with a known MoA, it is possible to analyze phenotypic similarities to possibly deduce the MoA. Having access to a library of images of a specific line of cells treated with compounds with different MoA’s can therefore be used to identify the MoA of new compounds, without going through the trouble of screening all of them (Caicedo et al. 2016).

1.1.4 Specificity of drugs and their annotation

Where HTS is used for screening for hits among possible leads, HCS is more useful in testing the toxicity and specificity of the lead (Manganelli et al. 2014). Due to HCS being more labour-intensive than HTS, it is usually done in later stages of drug discovery if both methods are combined. Nonetheless, it is impossible to screen the entire proteome of humans and therefore off-target drugs are found both in later stages of drug development and early stages of clinical trials in the form of side effects (Playe & Stoven 2020). An additional issue with off-target effects is because cell lines that are used in in vitro studies are lacking many of the characteristics of cells in live humans. The environment is obviously very different, with cells missing external cell to cell communication and other influences, making them a less-than optimal model (Astashkina et al. 2012). It is not only the case of drugs in clinical and preclinical trials showing off-target effects, many drugs on the market are removed due to undiscovered side-effects (Li 2004). If drugs on the market have undiscovered off-target side-effects, it is also very plausible that drugs on the market have

off-target effects that show no or very little negative side-effects. It is also shown that drugs on the market with analogous structure to other compounds have different annotated targets (Hu et al. 2014). This all goes to show that drugs and compounds might have additional targets than annotated.

1.2 GPCRs

With just over 800 GPCRs in the human body, they are one of the larger families of receptors (Luttrell 2008). Although the far largest portion of them are sensory

(6)

receptors, with over half of them olfactory receptors, they vary greatly, from adhesion-type receptors to hormone receptors (Fredriksson et al. 2003). Being of such large variety and numbers, they are expressed in all different kinds of cells, all over the body with varying types of function (Thompson et al. 2008). However, what they generally have in common is that they have 7 transmembrane helices and they are coupled with a GTPase, the G-protein (Neer 1995). Depending on the receptor and the G-protein, they generally activate one of two pathways: the

cAMP-dependent pathway or the phosphatidylinositol (PI) signalling pathway (Neves et al. 2002). Because of its large variability and wide expression, defects in the GPCRs are prone to lead to disease (Roth & Kroeze 2015). Likewise, it is also the reason that there are a large number of drugs targeting GPCRs and as of 2017, 475 drugs are approved by the FDA, targeting 108 GPCRs (Hauser et al. 2017). Drugs targeting one GPCR often show off-target effects of other GPCRs in the same family (Schlyer & Horuk 2006) and they work either to activate, agonistically, or to inhibit, antagonistically (Wootten et al. 2013).

1.3 Aim of the study

Specificity of drugs is not as simple as one drug one target, especially for GPCRs.

By using a HCS assay called Cell Painting we hope to shed light on the precision of drug annotations and their specificity. To follow up, we wanted to gain additional data from a DHM instrument for label-free, live-cell imaging, called HoloMonitor, and test the reliability of the instrument. To analyse the data, we did a principal component analysis (PCA) to see if drugs targeting the same GPCR would cluster together if they are specific, or would cluster between different groups, suggesting that they show affinity towards other GPCRs as well. We also aimed to compare regular images with deconvolved images and see if the images would provide better data, but due to the Covid-19 situation, this experiment was cut.

2. Materials and methods

2.1 Cell culture conditions

U2OS cells were used for the experiment. The cells are stored at -150°C and thawed before the experiment. When thawing, the vial containing the U2OS cells is thawed in a 37°C water bath for approximately one minute until the cell suspension is almost completely liquid. The cell suspension is transferred to a T75 flask (cat. no.

156499) containing 10 mL DMEM growth medium (Thermo Fisher Scientific, cat. no.

11995) supplemented with 10% FBS (Thermo Fisher Scientific cat. no. 11550356) and 1% penicillin + streptomycin (Thermo Fisher Scientific, cat. no. 15-140-122), prewarmed to 37°C in a water bath, and incubated in a cell culture incubator at 37°C and 5% CO2. After 24 hours the growth medium in the flask is changed with new, prewarmed, supplemented DMEM medium. The U2OS cells are maintained regularly

(7)

in T75 flasks with supplemented DMEM growth medium. The cells are split at ~80%

confluence, by washing twice with 5 mL sterile PBS (1X) (Thermo Fisher Scientific, cat. no. 14190144) and further detached with 1 mL TrypLE Express Enzyme (1X) without phenol red (Thermo Fisher Scientific, cat. no. A12177) in 4 mL sterile PBS (1X) and incubated in a cell culture incubator at 37°C and 5% CO2 for 10 minutes.

After detaching, the TrypLE is neutralized with 5 mL prewarmed supplemented DMEM growth medium and the cells are transferred to a Falcon tube. The cells are pelleted in a Sigma 2-7 centrifuge (Sigma Laboratory Centrifuges) at 3000 RPM for 3 minutes, the suspension aspirated and the pellet resuspended in 1 mL prewarmed, supplemented DMEM growth medium. The cells are transferred to a new T75 flask containing 12 mL prewarmed, supplemented DMEM growth medium and maintained at 37°C and 5% CO2.

2.2 Cell culture seeding and compound treatment

U2OS cells are seeded at a cell density of 5000 cells in 100 µL supplemented DMEM growth medium per well in a Corning Costar, flat bottom 96-well plate (Corning cat. no. 3603), not using the 36 outermost wells due to the edge effect, instead filling these wells with supplemented DMEM growth medium. The cell density is calculated with a Countess II FL (Thermo FIsher Scientific), using Tryptan Blue Stain (0.4%) (Thermo Fisher Scientific, cat. no. T10282). Before the addition of compounds, the media is aspirated and replaced with 100 µL new, prewarmed, supplemented DMEM growth medium per well. The compounds are spotted in 60 of the 96 wells in a 96-well plate and acquired from CBCS (Chemical Biology

Consortium Sweden), leaving the 36 outermost wells empty, and kept in a -20°C freezer. Before transferring the compounds, the plate is kept at room temperature for 30 minutes for thawing, after which 70 µL prewarmed, supplemented DMEM growth medium is added to all wells. The compounds are transferred onto the plate and put into the cell culture incubator until ready to run with the HoloMonitor. The compounds are run in three concentrations, 1 µM, 3 µM and 10 µM. Negative controls were used in form of cells treated with DMSO and cells not treated at all. Compounds were chosen from annotations of compounds targeting GPCRs in the DrugBank database and cross referenced with data from The Human Protein Atlas database about GPCRs available in U2OS cells (Table s1).

2.3 HoloMonitor/Quantitative phase imaging

A HoloMonitor M4 from Phase Holographic Imaging AB (PhiAB), located in a cell culture incubator at 37°C and 5% CO2, was used to capture time-lapse images of cells exposed to compounds. Using the Cell Morphology module of the AppSuite software (PhiAB) images were captured at 4 sites per well every 15 minutes for 24 hours, in order to be able to track individual cells. The AppSuite’s Kinetic

(8)

Proliferation Analysis pipeline was also used to obtain cell counts for each plate.

Immediately after compounds are added to cells, HoloLids (PhiAB) are put on the plate after soaking in ethanol for 15 minutes, and the plate is put inside the

HoloMonitor.

2.4 Cell Painting

Immediately before painting, the DMEM media is aspirated from the prepared cell plate and substituted with 100 µL Fluorobrite DMEM media (Thermo Fisher

Scientific, cat. no. A1896701), supplemented with MitoTracker Deep Red (Thermo Fisher Scientific, cat. no. M22426), for a working concentration of 900 nM under dark conditions, and incubated for 30 minutes in 37°C at 5% CO2 (Figure 3). After painting with MitoTracker, the cells are washed twice with PBS (1X) (Thermo Fisher

Scientific, cat. no. 18912014), the PBS is aspirated and replaced with 70 µL 4% PFA (Histolab Products AB, cat. no. 02176)for fixation and incubated 20 min in room temperature. The cells are again washed twice with PBS (1X), permeabilized with 70 µL 0.1% Triton (Thermo Fisher Scientific, cat. no. X100) and incubated for another 20 minutes at room temperature. The Triton is washed away with PBS (1x) twice and the cells are painted with 50 µL of a cocktail consisting of Hoechst (10 µg/mL)

(Thermo Fisher Scientific, cat. no. H3570), SYTO 14 (5 µM) (Thermo Fisher Scientific, cat. no. S7576), Concanavalin A, Alexa Fluor 488 (50 µg/mL) (Thermo Fisher Scientific, cat. no. C11252), Wheat Germ Agglutinin (WGA), Alexa Fluor 555 (3 µg/mL) (Thermo Fisher Scientific, cat. no. W32464) and Alexa Fluor 568

Phalloidin (10 µg/mL) (Thermo Fisher Scientific, cat. no. A12380) in PBS (1X) (Table 1), and incubated on a Fisherbrand Microplate Shaker (Thermo Fisher Scientific) in 200 RPM at room temperature for 20 minutes. Before storing, the plate was washed twice with PBS (1X), and lastly 100 µL PBS (1X) was added to all wells, and stored in 4°C until imaged, within one to two weeks (Table s2).

(9)

Figure 3: Schematic pathway of the cell painting method.

Table 1: The stock and working concentrations of the dyes used in the Cell Painting.

Dye Hoechst Syto WGA Phalloidin Concanvalin

Stock Conc. 10 mg/mL 5mM 1 mg/mL 1,5 mg/mL 2,5 mg/mL

Working Conc. 10 ug/mL 5 uM 3 ug/mL 10 ug/mL 50 ug/mL

2.5 Confocal Imaging

Confocal images were captured with an ImageXpress (Molecular Devices), at 9 sites at 5 channels with different emission wavelengths at 447, 536, 593, 624 and 692 nM.

2.6 Cell Profiler

Images from the confocal microscope were processed with CellProfiler 3.1.9

(Carpenter et al. 2006) and the computations were performed on resources provided by SNIC through Uppsala Multidisciplinary Center for Advanced Computational Science (UPPMAX) under Project SNIC 2019-8-149. The cells were identified with the nucleus as a primary object and cells were segmented to distinguish it from

(10)

neighbouring cells using watershed transformation to extract numerical values based on the cell’s morphology, and exported to a CSV file.

2.7 PCA

The PCA (Principal Component Analysis) was done in R 3.6.0 (R Core Team, 2019) with the numerical data from CellProfiler. A PCA is done because it lets you handle large amounts of data with many variables. It is a method to visualize which principal component (PC) has the largest proportion of variance on the data. The mean

average per well was used for the PCA and the different plates were normalized based on the controls.

2.8 Deconvolution

Deconvolution of the images was done in the Huygens Essential 19.10

Deconvolution Wizard (Scientific Volume Imaging). Only one Z layer was used and over 30 iterations. The Huygens Deconvolution Wizard calculates a theoretical PSF (Point Spread Function) (Jin et al. 2017), a constant calculated from the settings from the camera, as a way to minimize the noise from higher intensity in images.

3. Results

3.1 HoloMonitor

The data from the HoloMonitor was not used for any morphological analysis due to the high noise in many of the images (Figure 4). Despite this it was still used to compare cell count with the CellProfiler, where the HoloMonitor’s mean cell count is systematically higher than CellProfiler’s mean cell count (Figure 5).

Figure 4: U2OS cells imaged with HoloMonitor that shows the granular and ripple-like noise created from too thin cells.

(11)

(12)

Figure 5: Cell count with data from CellProfiler (A, C, E, G, I, K and M) and HoloMonitor after 24 hours (B, D, F, H, J, L and N) with the average number of cells per well in that plate.

3.2 Cell Painting

Visually, it appears that some drugs have an effect on the cells on some images (Figure 6), however from the PCA data it is not as apparent. When looking at the PCA from a single well, none of the GPCRs separate enough to form a profile, they all seem to more or less group around the controls (Figure 7). The same can be said when looking at each drug by itself, no drug really separates itself from the control enough for its own profile (Figure 8). However, when looking at the data for each

(13)

target individually, the drugs do not seem to cluster together in one group as clearly (Figure 9).

Figure 6: U2OS cells imaged in ImageXpress, nucleus stained blue, ER and golgi stained red and actin filaments stained green. A: Treated with Dipivefrin (1 µM). B: Treated with Desipramine (10 µM).

C: Treated with Esmolol (3 µM).

(14)

Figure 7: PCA plot on the values from each well for all plates, with each drug going against single targets coloured, while drugs going against multiple targets are merged under one group, Multi.

Figure 8: PCA plot with a single drug against the controls for a subset of four of the 131 different drugs.

(15)

Figure 9: PCA plot with every drug going against a single target, with the 5 targets with most drugs in this experiment (table s1). Marked in blue are two drugs that seem to cluster together, marked with a

*, Methoxamine (B) with Midodrine (D) and Nadolol (C) with Oxprenolol (E).

3.3 Deconvolution

I was unable to analyze the deconvoluted images completely due to the situation with Covid-19, and therefore have no results regarding whether or not it had a significant effect. However, after deconvolving a few images, there is an observable difference, which would likely give different data (Figure 10).

(16)

Figure 10: U2OS cells imaged with ImageXpress after cell painting, composed from three channels.

A: Image deconvoluted with Huygens Essential, where each image was deconvolved separately and later composed. B: Composed image from ImageXpress.

4 Discussion

4.1 HoloMonitor

This was the first time that the HoloMonitor was used in the lab, and therefore there was a lot of trial and error. We expected clearer images, but they came out with a lot of noise, occasionally covering the entire image. The reason why the data quality produced by the HoloMonitor was lower than expected was because the thickness of U2OS cells was too thin. A more suitable cell line would have given less noise, but on the other hand, U2OS cells are well established as good cells for cell painting experiments. Additionally the plates that were used are not recommended for the HoloMonitor, but they were required to be used for the cell painting. However, when testing these plates with A549 cells, that are thicker, in another experiment, the noise was reduced. As can be seen in Figure 5, there is a large difference in cell count, and the mean cell count is higher for the HoloMonitor, in one case even over three times higher. The difference in cell count between the two instruments is most likely due to this noise being segmented as cells, however, it is also plausible that this is because the cells are being put through cell painting after HoloMonitor. While the Appsuite has an option to increase or decrease the sensitivity for which cells segment, it was not perfect either. For example, the lowest available sensitivity still segmented some noise on images with high noise while omitting most of the cells from images with low noise, while on higher sensitivity much of the noise was

segmented as cells, which will give a higher cell count. All in all, just this experiment is not enough to fully evaluate the HoloMonitor as a tool for live-cell imaging in

image-based analysis but shows the limitation for working with cells that are not thick enough. These results were communicated to the industry partner Phase

Holographic Imaging (PHIAB).

(17)

4.2 Cell Painting

This was the first attempt to profile GPCR-binding drugs using Cell Painting, and evaluate their groupings using clustering approaches based on morphological profiles. While the PCA on multiple targets (Figure 7) doesn’t show such clear separation as we had hypothesized that it would, it is hard to say why without any further studies. It seems to be the same with single drugs against the controls, they do not seem different enough from the controls to separate (Figure 8). It might be that despite U2OS cells having expression of GPCRs, the receptor is inactive in some component, such as the G protein being inactive, or another downstream agent being inhibited. It could also be due to methodical reasons, such as 24 hours not being enough for the drug to work, or 10 µM being too low of a concentration.

This, however, would unlikely be the case since just over 130 drugs are being tested and while some of them might require higher doses or longer time, it is most likely not the case for all of them and some would show different morphologies. While these results are unclear whether this means that U2OS cells are unfit for

image-based analysis with GPCRs as drug target, it would need further testing with different cell lines, provided they express the same GPCRs. This does not mean that none of the data worked along our hypothesis, in Figure 9 you can clearly see some tendency of clustering. While the targets with less than 10 drugs show less of a pattern, the targets with more than 10 show a clearer separation from each other.

This trend could mean that some of these drugs work in another manner to the rest, and might act on another target as well, but it needs to be confirmed with additional experiments.

It seems that at least two drugs cluster together in ADR1B and ADRB2 respectively, despite them binding the same target: Methoxamine with midodrine and nadolol with oxprenolol. Both methoxamine and midodrine are agonists and that might be a reason why they cluster together, however, so is metaraminol and it clusters very differently (Aronson 2016a, Aronson 2016b). It is also possible that they are clustering due to what in my opinion is a somewhat similar structure, but as I have limited experience in judging 2D images of compounds it is not a safe assumption.

What seems to be a likelier case, however, is that both of them would work on another target. According to ChEMBL’s target prediction both are theorized to work against the 5-HT3A receptor (Bosc et al. 2019). This receptor is not present in U2OS cells, but it could be a hint towards them working against another receptor, serotonin or other. Oxprenolol and nadolol on the other hand do not share many qualities, as they aren’t very similar in structure (McDevitt 1987). The only thing they have in common is that they both work antagonistically, but that is the most common form of activity within the list of compounds. Therefore, it seems plausible that there is an

(18)

unknown common factor, such as having another unannotated target that they both work against.

4.3 Deconvolution

Despite the comparison of the data before and after deconvolution fell through due to the events of Covid-19, I think the method has promise. Although it is entirely based on my opinion and not actual data, the images in Figure 10 show that there is an increase of focus and a slight decrease of noise. Having sharper focus on an image will of course show more details, but in image analysis this is important, as it can more significantly measure subcellular compartments, or even compartments within organelles. Although, it is important to remember that it is not a rendition of the actual image, but more of a calculated “what-if”. It can therefore be important to keep in mind that if done incorrectly, it can provide data inaccurately.

4.4 Conclusions

What we learned was that the HoloMonitor definitely has advantages over many instruments used for cell imaging, such as it being label-free. On the other hand it will produce blurry and noisy images if one uses cells that are too thin, like we used.

More work is needed to evaluate the performance of the HoloMonitor for live-cell imaging and profiling. It seemed difficult for us to separate targets from each other based on the cell’s morphologic features from the cell painting profile of the drug’s effect on GPCRs. However, we were able to identify clusters of drugs for some targets that show different patterns from other, similar drugs. This indicates that cell painting can most likely be used for studies in drug annotation specificity, but more studies are needed to confirm this.

Personal reflection

Not everything I did and learned fit in a report, and I will list some of what learned. A large part of the analysis was done in R, a language I was only vaguely familiar with, and jumping from basics to relatively complex analysis meant I learnt a lot, not only about statistics and analysis, but about the R language as well. On top of that I learnt some basic python programming as well. This was in order to work with a pipetting robot, Opentron’s OT-2, which has its API in python. This was meant to be part of the group’s plans of lab automation, and learning programming as well as the logistics that go into such plans was very valuable. The plan was that the pipetting robot was also going to perform some of the cell painting experiments, alleviating some of the workload, but in the end, it was never fully set up.

References

Aronson JK (ed). 2016a. Methoxamine. Meyler’s Side Effects of Drugs (Sixteenth Edition), p.

(19)

912. Elsevier, Oxford.

Aronson JK (ed). 2016b. Midodrine. Meyler’s Side Effects of Drugs (Sixteenth Edition), pp.

1030–1031. Elsevier, Oxford.

Astashkina A, Mann B, Grainger DW. 2012. A critical evaluation of in vitro cell culture models for high-throughput drug screening and toxicity. Pharmacology &

Therapeutics 134: 82–106.

Bender A, Bojanic D, Davies JW, Crisman TJ, Mikhailov D, Scheiber J, Jenkins JL, Deng Z, Hill WAG, Popov M, Jacoby E, Glick M. 2008. Which aspects of HTS are empirically correlated with downstream success? Current Opinion in Drug Discovery &

Development 11: 327–337.

Bosc N, Atkinson F, Felix E, Gaulton A, Hersey A, Leach AR. 2019. Large scale comparison of QSAR and conformal prediction methods and their applications in drug discovery.

Journal of Cheminformatics 11: 4.

Bray M-A, Singh S, Davis CT, Borgeson B, Hartland C, Kost-Alimova M, Gustafsdottir SM, Gibson CC, Carpenter AE. 2016. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes | Nature Protocols. Nature Protocols 11: 1757–1774.

Caicedo JC, Cooper S, Heigwer F, Warchal S, Qiu P, Molnar C, Vasilevich AS, Barry JD, Bansal HS, Kraus O, Wawer M, Paavolainen L, Herrmann MD, Rohban M, Hung J, Hennig H, Concannon J, Smith I, Clemons PA, Singh S, Rees P, Horvath P, Linington RG, Carpenter AE. 2017. Data-analysis strategies for image-based cell profiling. Nature Methods 14: 849–863.

Caicedo JC, Singh S, Carpenter AE. 2016. Applications in image-based profiling of perturbations. Current Opinion in Biotechnology 39: 134–142.

Carpenter AE, Jones TR, Lamprecht MR, Clarke C, Kang IH, Friman O, Guertin DA, Chang JH, Lindquist RA, Moffat J, Golland P, Sabatini DM. 2006. CellProfiler: image

analysis software for identifying and quantifying cell phenotypes. Genome biology 7:

R100–R100.

Fox S, Farr-Jones S, Sopchak L, Boggs A, Nicely HW, Khoury R, Biros M. 2006.

High-Throughput Screening: Update on Practices and Success. Journal of Biomolecular Screening 11: 864–869.

Fredriksson R, Lagerström MC, Lundin L-G, Schiöth HB. 2003. The G-Protein-Coupled Receptors in the Human Genome Form Five Main Families. Phylogenetic Analysis, Paralogon Groups, and Fingerprints. Molecular Pharmacology 63: 1256–1272.

Hauser AS, Attwood MM, Rask-Andersen M, Schiöth HB, Gloriam DE. 2017. Trends in GPCR drug discovery: new agents, targets and indications. Nature Reviews Drug Discovery 16: 829–842.

Hu Y, Lounkine E, Bajorath J. 2014. Many Approved Drugs Have Bioactive Analogs With Different Target Annotations. The AAPS Journal 16: 847–859.

Jin X, Liu L, Chen Y, Dai Q. 2017. Point spread function and depth-invariant focal sweep point spread function for plenoptic camera 2.0. Optics Express 25: 9947–9962.

Kühn J, Shaffer E, Mena J, Breton B, Parent J, Rappaz B, Chambon M, Emery Y, Magistretti P, Depeursinge C, Marquet P, Turcatti G. 2013. Label-Free Cytotoxicity Screening Assay by Digital Holographic Microscopy. Assay and Drug Development

Technologies 11: 101–107.

Lan W, Wang J, Li M, Liu J, Li Y, Wu F-X, Pan Y. 2016. Predicting drug–target interaction using positive-unlabeled learning. Neurocomputing 206: 50–57.

Li AP. 2004. Accurate prediction of human drug toxicity: a major challenge in drug development. Chemico-Biological Interactions 150: 3–7.

Luttrell LM. 2008. Reviews in Molecular Biology and Biotechnology: Transmembrane Signaling by G Protein-Coupled Receptors. Molecular Biotechnology 39: 239–264.

(20)

Manganelli G, Masullo U, Filosa S. 2014. HTS/HCS to Screen Molecules Able to Maintain Embryonic Stem Cell Self-Renewal or to Induce Differentiation: Overview of Protocols. Stem Cell Reviews and Reports 10: 802–819.

Mayr LM, Bojanic D. 2009. Novel trends in high-throughput screening. Current Opinion in Pharmacology 9: 580–588.

McDevitt DG. 1987. Comparison of pharmacokinetic properties of beta-adrenoceptor blocking drugs. European Heart Journal 8: 9–14.

Neer EJ. 1995. Heterotrimeric G proteins: organizers of transmembrane signals. Cell 80:

249–257.

Neves SR, Ram PT, Iyengar R. 2002. G protein pathways. Science; Washington 296:

1636–9.

Playe B, Stoven V. 2020. Evaluation of deep and shallow learning methods in

chemogenomics for the prediction of drugs specificity. Journal of Cheminformatics 12: 11.

Rappaz B, Breton B, Shaffer E, Turcatti G. 2014. Digital Holographic Microscopy: A Quantitative Label-Free Microscopy Technique for Phenotypic Screening.

Combinatorial Chemistry & High Throughput Screening 14: 80–88.

Roth BL, Kroeze WK. 2015. Integrated Approaches for Genome-wide Interrogation of the Druggable Non-olfactory G Protein-coupled Receptor Superfamily. Journal of Biological Chemistry 290: 19471–19477.

Schlyer S, Horuk R. 2006. I want a new drug: G-protein-coupled receptors in drug development. Drug Discovery Today 11: 481–493.

Thompson MD, Cole DEC, Jose PA. 2008. Pharmacogenomics of G Protein-Coupled Receptor. In: Yan Q (ed.). Pharmacogenomics in Drug Discovery and Development:

From Bench to Bedside, pp. 77–107. Humana Press, Totowa, NJ.

Wootten D, Christopoulos A, Sexton PM. 2013. Emerging paradigms in GPCR allostery:

implications for drug discovery. Nature Reviews Drug Discovery 12: 630–644.

Supplementary Data

Table s1: The different drugs used in the experiment, what library they were obtained from and the GPCR they target.

Library Drug Name GPCR(s) Library Drug Name GPCR(s) Prestwick adenosine AA1R,

AA2AR, AA2BR

Prestwick mephentermine ADA1B, ADA2C, ADRB2 Prestwick alfuzosin ADA1B Prestwick mesoridazine 5HT2A Prestwick alprenolol ADRB2 Prestwick metaraminol ADA1B Prestwick alprostadil PE2R2 Prestwick methoxamine ADA1B Prestwick amitriptyline 5HT2A Prestwick methyldopa ADA2C

Prestwick amoxapine DRD1 Prestwick metixene ACM4

Prestwick antazoline HRH1 Prestwick mianserin HRH1,

(21)

5HT2A Prestwick aripiprazole 5HT2A Prestwick midodrine ADA1B Prestwick asenapine ADA2C,

DRD1, HRH1, 5HT2A

Prestwick minaprine DRD1, 5HT2A

Prestwick astemizole HRH1 Prestwick mirtazapine ADA2C, 5HT2A

Prestwick atropine ACM4 Prestwick misoprostol PE2R2

Prestwick azelastine HRH1 Prestwick molindone 5HT2A

Prestwick baclofen GABR1 Prestwick nadolol ADRB2

Prestwick betahistine HRH1 Prestwick naphazoline ADA1B, ADA2C Prestwick brompheniramine HRH1 Prestwick nefazodone ADA1B,

5HT2A

Prestwick caffeine AA1R,

AA2AR, AA2BR

Prestwick olanzapine ADA1B, ADA2C, ACM4, DRD1, HRH1, 5HT1D, 5HT2A Prestwick carbinoxamine HRH1 Prestwick olopatadine HRH1 Prestwick carteolol ADRB2 Prestwick orphenadrine HRH1 Prestwick carvedilol ADA1B,

ADRB2

Prestwick oxprenolol ADRB2

Prestwick celiprolol ADRB2 Prestwick oxymetazoline ADA1B, ADA2C

Prestwick cetirizine HRH1 Prestwick trazodone 5HT2A

Prestwick chlorcyclizine HRH1 Prestwick trifluoperazine 5HT2A Prestwick chloropyramine HRH1 Prestwick triflupromazine DRD1 Prestwick chlorpheniramine HRH1 Prestwick trimeprazine HRH1 Prestwick chlorpromazine 5HT2A Prestwick trimipramine ADA1B,

5HT2A

(22)

Prestwick chlorprothixene DRD1, HRH1, 5HT2A

Prestwick tripelennamine HRH1

Prestwick cinnarizine HRH1 Prestwick triprolidine HRH1 Prestwick clenbuterol ADRB2 Prestwick tropicamide ACM4 Prestwick clonidine ADA2C Prestwick vigabatrin GABR1 Prestwick clozapine DRD1,

HRH1, 5HT2A

Prestwick yohimbine ADA2C

Prestwick cromolyn sodium GPR35 Prestwick ziprasidone 5HT2A Prestwick cyclizine HRH1 Prestwick zuclopenthixol DRD1 Prestwick cyclobenzaprine 5HT2A Selleck aclidinium ACM4 Prestwick cyproheptadine HRH1,

5HT2A

Selleck almotriptan 5HT1D

Prestwick desipramine ADRB2, HRH1

Selleck bepotastine HRH1

Prestwick desloratadine HRH1 Selleck bimatoprost PF2R Prestwick dihydroergotamine 5HT1D Selleck dexmedetomidin

e

ADA2C

Prestwick dimenhydrinate HRH1 Selleck droxidopa ADA1B, ADA2C, ADRB2 Prestwick diphenhydramine HRH1 Selleck epinephrine,

adrenaline

ADA1B, ADA2C, ADRB2 Prestwick diphenylpyraline HRH1 Selleck fingolimod S1PR1, S1PR5 Prestwick dipivefrin ADA1B,

ADA2C, ADRB2

Selleck iloperidone DRD1, HRH1, 5HT2A Prestwick dobutamine ADRB2 Selleck indacaterol ADRB2 Prestwick doxazosin ADA1B Selleck istradefylline AA2AR

Prestwick doxepin ACM4,

HRH1, 5HT2A

Selleck lurasidone 5HT2A

(23)

Prestwick doxylamine HRH1 Selleck LY310762 5HT1D

Prestwick emedastine HRH1 Selleck nebivolol ADRB2

Prestwick esmolol ADRB2 Selleck paliperidone 5HT2A

Prestwick fexofenadine HRH1 Selleck phenylephrine ADA1B

Prestwick flunarizine HRH1 Selleck plerixafor CXCR4

Prestwick fluphenazine DRD1 Selleck rizatriptan 5HT1D

Prestwick formoterol ADRB2 Selleck rupatadine HRH1

Prestwick guanabenz ADA2C Selleck solifenacin ACM4

Prestwick guanfacine ADA2C Selleck tolterodine ACM4

Prestwick haloperidol 5HT2A Selleck trospium ACM4

Prestwick homatropine methylbromide

ACM4 Selleck vismodegib SMO

Prestwick hydroxyzine HRH1 Selleck WZ 811 CXCR4

Prestwick isoetharine ADRB2 Selleck zolmitriptan 5HT1D

Prestwick itraconazole SMO Tocris clemastine HRH1

Prestwick ketotifen HRH1 Tocris eletriptan 5HT1D

Prestwick labetalol ADA1B, ADA2C, ADRB2

Tocris histamine HRH1

Prestwick levocabastine HRH1 Tocris isoproterenol ADRB2

Prestwick levodopa DRD1 Tocris methysergide 5HT2A

Prestwick lofexidine ADA2C Tocris procaterol ADRB2

Prestwick loratadine HRH1 Tocris propranolol ADRB2

Prestwick loxapine 5HT2A

(24)

Table s2: Plate order and experimental dates. Experiments marked with a * were disrupted midway, due to HoloMonitor malfunctions.

Plate ID Seed date HoloMonitor date Cell Painting date Imaging date

P010719 11/12/19 N/A 13/12/19 18/12/19

P010720 13/1/20 14/1/20 15/1/20 20/1/20

P010721 14/1/20 15/1/20* 16/1/20 20/1/20

P010722 15/1/20 16/1/20* 17/1/20 20/1/20

P010723 20/1/20 21/1/20 22/1/20 27/1/20

P010724 21/1/20 22/1/20 N/A N/A

P010726 27/1/20 28/1/20 29/1/20 11/02/20

P010727 28/1/20 29/1/20 30/1/20 11/02/20

P010728 29/1/20 30/1/20 31/1/20 11/02/20

Challenging specificity of chemicalcompounds targeting GPCRs with cellprofiling