The cell painting approach evokes new ideas and strategies for high-content screening (HCS) practitioners to identify new drugs and to study mechanisms of action by leveraging machine learning to measure subtle phenotypic changes at the cellular level. Dr. Alexander Schreiner, HCS Applications Leader, PerkinElmer and Dr. Martin Daffertshofer, High Content Screening Software Product Manager, PerkinElmer discuss how cell painting is being applied today in drug discovery and the current technical limitations and challenges.
How can we identify new therapeutics? This remains a fundamental question for drug discovery, as it becomes increasingly challenging to find new molecular entities (NMEs). It is widely accepted that there are two major modes of the drug discovery process: the target-based approach and phenotypic-based approach. For target-based campaigns, most often prior knowledge of a disease caused by an associated factor must be known. Screening approaches focus on a specific target and how it can be modulated, eg. by chemical compounds or genetic perturbation. For a phenotypic approach, typically, no prior knowledge of the factors involved is needed, and the screening focuses on identifying compounds that change the cellular phenotype.1
The common theme for both approaches is that the success will largely depend on how well the screening model recapitulates disease physiology, ie, how valid is the target and how well does the cellular model mimic the disease phenotype. A target-based approach will fall short if a target is not sufficiently well known, or if multiple targets are responsible for the disease state and each target contributes to the disease in a different way, most likely in an unknown order. Phenotypic assays, however, require model systems that ‘phenocopy’ the disease as closely as possible to be most successful. In this respect, advances in iPSC-derived model systems, CRISPR-Cas9 technology to validate involvement in genetic dispositions, and 3D spheroid/organoid cultures, which provide more physiologically relevant model systems, are beneficial for phenotypic assays.
Although both approaches have been used to generate new drugs, phenotypic approaches were mainly used in the early days of drug discovery. With advances in molecular biology and the completion of the human genome project, target-based screens increasingly became the central focus. However, in recent years there has been a gradual shift back to more phenotypic assays, maybe in part inspired by a publication from Swinney and Anthony (2011)² which looked at FDA-approved drugs and how phenotypic approaches contributed to the discovery of first-in-class NMEs.
Here, we focus on the phenotypic approach and particularly the ‘cell painting’ method as the most recent and comprehensive example of a powerful phenotypic assay being adopted by the academic screening centers and the pharmaceutical industry.
The term ‘phenotype’ can be defined as the sum of all observable features of a specimen that results from the genetic predisposition and environmental interactions. Therefore, the ideal phenotypic assay should be able to describe a phenotype regardless of the underlying cause, such as genetic mutations, epigenetic, physiological or metabolic alterations, or chemical perturbations, and just by describing what the system ‘looks like’.
As a cellular phenotype is the result of aggregating all observable features, microscopic imaging and, in particular, high content imaging is an ideal technology for phenotypic assays. The advantage of cell imaging is that an image conserves all phenotypic information, and this is available for as long as the image is preserved, well beyond the sample.
However, to comprehensively extract phenotypic data, every compartment of a cell should ideally be labelled. Of course, there will be a limit to the number of labels that can be applied at one time due to cost, time and spectral overlap in the visible spectrum.
In 2013, researchers from the Broad Institute suggested a set of six different dyes acquired in five fluorescent channels to label cells and extract as many features as possible from each channel.3 The six probes used in this study included: Hoechst 33342 to stain the nuclei, Concanavalin A to stain the endoplasmic reticulum, a nucleic acid to stain nucleoli, wheat germ agglutinate (WGA) to stain membrane compartments (prominently the Golgi), Phalloidin to stain the actin cytoskeleton, and a mitochondrial stain. Of these, WGA and actin were acquired in one channel. This combination ensures most of the cellular structures are labeled and features can be extracted.
In 2016, this approach was named ‘cell painting’ and a revised protocol was published by Bray, et al. (Anne Carpenter was one of the senior authors of the 2013 paper).4This protocol is now considered to be the ‘original’ cell painting protocol. The cell painting workflow consists of cell treatment and staining, image acquisition, image analysis/feature extraction, and computational analysis of extracted features. However, for each step in the workflow, some challenges remain.
Treatment
A notable strength of cell painting is that a compound-induced phenotype can be described completely independently from the mode of action (MoA) or target. Therefore, reference phenotypes are needed as landmarks. Hence, the questions you can answer with cell painting will depend on your reference phenotypes. In a recent publication for example, the cell painting approach was used to screen for bio-activity of environmental chemicals.5
Staining
For staining, all the fluorescent dyes need to be purchased and stock solutions prepared. To ease the latter task, an optimised cell painting kit was recently brought to market containing all the dyes in the required amount to perform the ‘original’ cell painting application, making it easy to start the workflow.6 Figure 2 shows example images of cells stained using this kit and treated with compounds known to affect the targets of the respective dyes.
Image acquisition
We have previously shown that for small chemical spaces image acquisition can be varied without significantly influencing the outcome of phenotypic clustering.6 However, for larger chemical spaces producing more diverse phenotypes, the acquisition parameters become more restrictive. Therefore, it is important to have an imaging device that can scale the image quality to the required level. For example, specimen with more background noise will benefit from confocal imaging whereas monolayer cultures can be sufficiently imaged using non-confocal acquisition. Furthermore, a spinning disc allows for acquiring z-stacks which can then be analysed using a maximum intensity projection. This is especially useful if the signal of interest is distributed within a volume exceeding the depth of focus of the objective. Water immersion objectives will always be an advantage because of the higher numerical aperture allowing more light into the instrument’s optical pathway.7For high content screening approaches, and particularly for the cell painting assay, the specimen is ‘mounted’ or seeded in an aqueous solution. Therefore, using a water immersion objective will provide the best possible image quality because the water used in the immersion matches the refractive index of the aqueous solution.6As a result, the acquisition speed will be increased using water immersion objectives because the exposure time to capture a fluorescent signal is reduced.
Image analysis/feature extraction
Image segmentation and extraction of hundreds to thousands of parameters can be difficult and time consuming to set up. Ideally, analysis workflows would require minimal set up time and be conducted without the need for any image analysis script development knowledge.
For larger screening campaigns, leveraging the scalability of computing clusters or cloud-based computing will be required.4 However, for small screens, time savings from increased computational power in a cloud environment may not justify the time spent uploading and transferring datasets. In such scenarios, on premise analysis solutions that do not require data transfer are more efficient. The ideal workflow uses the same analysis pipeline for both on premise and cloud-based solutions and is seamlessly connected to or integrated with the image acquisition system. This allows a swift transition from assay development to screening.
Particularly for cell painting, the secondary analysis of extracted features is considered the most challenging and critical step in the workflow after choosing the biological system. Dealing with data tables comprising several thousand columns and rows requires state of the art visualisation and analysis tools. This becomes even more challenging if single cell data is to be analysed. For example, the analysis of a 384-well microplate where three fields per well were acquired with a 20x water immersion objective and 3107 features per cell were extracted, yielding a single cell results table of 12.3Gb. For such a data table containing 3107 columns, each analysed object is defined by 3107 features; in other words, a 3107-dimensional space. Therefore, one of the first tasks should be reducing the number of features or dimensionality of the data. One approach is to take all features and use a feature reduction and visualisation method like principal component analysis (PCA), uniform manifold approximation (UMAP)8, or t-distributed stochastic neighbor embedding (tSNE).In the end, the data is pared down to a set of a few hundred features with sufficient variance to separate the phenotypes.
So why extract thousands of features only to reduce them down to a few hundred? The simple answer is that at the start of the process, the features needed to separate the phenotypes are unknown and only become apparent once they have been extracted and analysed.
A vibrant future
Cell painting is an effective tool for analysing the effect of most forms of perturbagens tested to date. It can be used for the ‘classical’ screening approach whereby a cell culture model is treated with hundreds of thousands of compounds. One advantage of a cell painting-driven screen is that the phenotypic clustering will be independent of a known target, enabling identification of novel compounds with different MoAs. These results could also prove valuable in the future to identify novel compounds. For example, if a compound is found to be effective against a disease, other compounds causing similar phenotypes in the ‘original’ screen could be tested and may result in novel hits and leads without the need to set up a new screen.9 The success of such an approach will depend on the biological model system used and how physiologically relevant it is for a given disease.
The current ability to precisely generate a disease state and revert it, for example, by using iPSC-derived models and/or CRISPR-Cas9 to specifically edit endogenous genes, will be a boost for phenotypic screening approaches such as cell painting. Especially with CRISPR-Cas9, it is possible to more specifically perturb a cell than using a chemical compound that will always have off-target effects by affecting unrelated endogenous components. If CRISPR-Cas9 is combined with a clearly-defined genetic background, such as haploid cells guaranteeing full allele targeting, unrelated off-target effects will be minimised. For genome-wide imaging screens, arrayed CRISPR-Cas9 libraries are available.
Another more complex approach is to apply cell painting to 3D model systems. There are complexities along every step of these assays, starting with production then staining, imaging and analysis. Imaging and analysis of 3D models is particularly challenging due to light scattering reducing imaging depth. One method to improve this is optical clearing.10,11 As this process can remove proteins and lipids from the specimen, it may affect the staining pattern and most optical clearing approaches also alter the volume of the whole 3D specimen.
Regarding staining, it will be interesting to see which dye combinations will be considered going forward. Currently, the combination from Bray et al. publication is favored. In the future, perhaps we will see more ‘hybrid’ approaches, combining cell painting dyes with antibody staining against a target or a combination of more disease-relevant dyes, ie those that would stain cellular compartments involved in a disease phenotype. These approaches may prove advantageous for connecting compounds to MoA.12
Conclusion
Cell painting as the most consistent implementation of phenotypic screening in drug discovery allows new strategies for high content screening (HCS) scientists to identify new drug candidates and study mechanisms of action by taking advantage of both cellular imaging and machine learning.
It was introduced to pharmaceutical applications a couple of years ago and efforts have been taken to adapt the approach to high content instrumentation with high resolution. The first cell painting reagent kit has been launched and software has been released to support the entire workflow. However, various challenges remain that require further optimisation.
The volume of data is proving challenging for information technology departments in the pharmaceutical industry and academic institutes. Leveraging cloud-based implementation as well as high performance computing technologies will support time and cost savings, particularly in downstream analysis.
In addition, as with other high-content screening assays, there is inherent bias built into the data from end-user input, and the use of supervised and unsupervised machine learning, artificial intelligence (AI), deep learning, convolutional neural networks (CNNs), and other computational processing requires careful examination to reduce bias as much as possible.
Cell painting has invoked the use of several analytical tools to effectively analyse large amounts of extracted feature data generated from HCS experiments. These tools, which may include AI or machine learning guidance, enable researchers to better understand, fingerprint, and interpret the data by classifying phenotypic profiles to determine MoA and identify new discoveries from unknown perturbagens.
Figure 1 – Schematic drawing of a cell and the organelles/structures stained in the Cell Painting application
Figure 2 –Representative images of Cell Painting staining patterns. Images were acquired on an Opera Phenix Plus High Content Screening System using a 40x water immersion objective. (A) Cropped images of HeLa cells labeled individually with PhenoVue Hoechst 33342 Nuclear Stain, PhenoVue Fluor 488 – Concanavalin A, PhenoVue 512 Nucleic Acid Stain, PhenoVue Fluor 568 – Phalloidin, PhenoVue Fluor 555 – WGA and PhenoVue 641 Mitochondrial Stain either untreated (upper panel) or after treatment (lower panel). Fenbendazole treatment leads to giant multinucleated cells, Tetrandrine to abundant endoplasmic reticulum, Etoposide to large nucleoli, Ca-074Me affects abundance of Golgi, Cytochalasin D leads to disruption of the actin cytoskeleton and Berberine chloride to redistribution of mitochondria. (B) Composite images of HeLa cells stained with all Cell Painting kit dyes. Left image shows staining pattern of untreated cells. The images to the right show staining patterns after indicated treatment.
Figure 3 – Principle component analysis to reduce dimensionality and visualise the data. Wells containing cells with similar phenotypes will cluster together. For each compound three wells were treated, and principal component analysis was performed using PerkinElmer Signals Screening platform. From this PCA, it can be seen that the phenotypic clustering for this small chemical space was not affected by the different acquisition parameters. The ER dye and the RNA/Nucleoli dye are prone to spectral crosstalk and it is often questioned if this influences the phenotypic clustering. Therefore, we show here with two different filter settings that even when acquiring both signals within the same channel (500-550) the phenotypic clustering is not significantly different from an optimal spectral separation by acquiring the ER dye within the range of 500-530nm and the RNA/Nucleoli stain within 570-630nm. Images were acquired on an Opera Phenix Plus High Content Screening System using a 20x water immersion objective and image analysis was performed using a cell painting specific analysis option of Harmony high-content analysis software.Volume 22, Issue 2 – Spring 2021
About the authors
As Team Leader of the Biological Applications Team at PerkinElmer in Hamburg Germany, Alexander Schreiner is responsible for developing applications and providing application related support for PerkinElmer’s high content screening systems. He holds a doctorate in Biology from the Goethe University in Frankfurt.
Martin Daffertshofer is Platform Leader for High Content Screening Software at Perkin Elmer since 2007. From 2002 – 2007 he has been working as Vice President Drug Discovery Technologies at Evotec Technologies. Martin is a PhD in Natural Science from the University Stuttgart, Germany.
References
- Roberts JP and Tesdorpf J: White Paper (2015): Phenotypic Drug Discovery with High Content Screening, PerkinElmer.
- Swinney DC and Anthony J (2011): How were new medicines discovered?, Nature Reviews Drug Discovery, 10, 507-519
- Gustafsdottir SM, Ljosa V, Sokolnicki KL, Anthony Wilson J, Walpita D, Kemp MM, et al. (2013) Multiplex Cytological Profiling Assay to Measure Diverse Cellular States. PLoS ONE 8(12): e80999.
- Bray MA, Singh S, Han H, 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. Nat Protoc. 2016 Sep;11(9):1757-74. doi: 10.1038/nprot.2016.105. Epub 2016 Aug 25. PMID: 27560178; PMCID: PMC5223290.
- Nyffeler J, Willis C, Lougee R, Richard A, Paul-Friedman K, Harrill JA. Bioactivity screening of environmental chemicals using imaging-based high-throughput phenotypic profiling. Toxicol Appl Pharmacol. 2020 Jan 15;389:114876. doi: 10.1016/j.taap.2019.114876. Epub 2019 Dec 30. PMID: 31899216
- Letzsch S and Schreiner A: Application Note (2021): Cell Painting for Phenotypic Screening, PerkinElmer.
- Boettcher K and Schreiner A: Technical Note (2017): The Benefits of Water Immersion Lenses for High-Content Screening, PerkinElmer
- Gregory P. Way, Maria Kost-Alimova, Tsukasa Shibue, William F. Harrington, Stanley Gill, Federica Piccioni, Tim Becker, Hamdah Shafqat-Abbasi, William C. Hahn, Anne E. Carpenter, Francisca Vazquez, Shantanu Singh (2020). Predicting cell health phenotypes using image-based morphology profiling. bioRxiv 2020.07.08.193938; doi: 10.1101/2020.07.08.19393
- Arany, A., Hochreiter, S., Moreau, Y., Ceulemans, H., Arany, A., & Steijaert, M. (2018). Repurposing High-Throughput Image Assays Enables Biological Activity Prediction for Drug Resource Repurposing High-Throughput Image Assays Enables Biological Activity Prediction for Drug Discovery. 1–8. https://doi.org/10.1016/j.chembiol.2018.01.015
- Letzsch S, Boettcher K and Schreiner A: Technical Note (2018): Clearing Strategies for 3D Spheroids, PerkinElmer
- Foitzik A, Boettcher K, Preckel H and Schreiner A: Technical Note (2018): 3D Volumetric and Zonal Analysis of Solid Spheroids, PerkinElmer
- Trask J: White Paper (2020): Cell Painting Phenotypic Drug Discovery Re-Imagined, PerkinElmer.
