Home
Welcome to the Cell Painting wiki! This page is intended for the scientific community to add their tips and tricks about running the assay.
It is intended to be read alongside the current Nature Protocols paper and to provide material that supplements that. Please be careful not to add content that is redundant relative to the paper.
Create a github account to edit this page or email your tip to [email protected]
- Introductions to exploring Cell Painting profile data
- Updates to the Cell Painting protocol
- Cell Painting papers to get oriented
- Cell Painting datasets and publications
- Experimental design considerations
- Materials
- Procedure
Table of contents generated with markdown-toc
- Videos: Part 1 | Part 2
- Blog post: Help! Interpreting image-based profiles
- Tutorial: Interpreting Image-based Profiles using Similarity Clustering and Single-Cell Visualization
- The JUMP-Cell Painting Consortium performed extensive optimization in 2020-2021 to yield Cell Painting Protocol v3 (2021_03_19)
- The new Nature Protocols paper version (v3) differs from the previous Nature Protocols paper version (v2) in these ways:
- No media removal before addition of MitoTracker to minimize the loss of cells
- Combining permeabilization and staining steps to make the process more automation friendly
- Reduction of Phalloidin 4-fold, from 5µl/ml (33nM) to 1.25µl/ml (8.25nM)
- Reduction of Hoechst 5-fold, from 5µg/ml to 1µg/ml
- Overall reduction of staining volumes from 30µl/well to 20µl/well
- Our recommendation for MitoTracker staining concentration remains at 500nM, but previous versions of the protocol used instructions that unintentionally lead to a lower final concentration (375nM). The current protocol ensures a 500nM concentration after dilution.
- Increase Syto-14 2-fold, from 3µM to 6µM
- Reduction of Concanavalin A 20-fold, from 100µg/ml to 5µg/ml
- Note that changes 1-6 are also present in version 2.5; only changes 7 and 8 (changes in Syto and ConA concentrations) differ between v2.5 and v3.
- The JUMP-Cell Painting Consortium performed extensive optimization in 2020-2021 to yield Cell Painting Protocol v3. This version of the protocol contains most though not all the changes in v3; it was used to create the profiles in the CP-JUMP1 experiment (Chandrasekaran et al 2021) as well as the CP-JUMP-Scope experiment (Jamali et al 2022, in preparation). This version differs from the previous Nature Protocols paper version (v2) in these ways:
- No media removal before addition of MitoTracker to minimize the loss of cells
- Combining permeabilization and staining steps to make the process more automation friendly
- Reduction of Phalloidin 4-fold, from 5µl/ml (33nM) to 1.25µl/ml (8.25nM)
- Reduction of Hoechst 5-fold, from 5µg/ml to 1µg/ml
- Overall reduction of staining volumes from 30µl/well to 20µl/well
- Our recommendation for MitoTracker staining concentration remains at 500nM, but previous versions of the protocol used instructions that unintentionally lead to a lower final concentration (375nM). The current protocol ensures a 500nM concentration after dilution.
The Nature Protocols paper and BioRxiv pre-print version and % supplementary website for version (v2) differs from the previous (Gustafsdottir 2013) version (v1) in three ways:
- The Nature Protocols version specifies that staining with WGA should be done after fixation. In Gustafsdottir 2013, staining with WGA was done before fixation, during the live cell staining step, which was chosen because it resolved distorted WGA signals seen when permeabilization step was too long, as often necessary when processing a large number of plates). The Nature Protocols switch to using WGA post-fixation was a result of our experiments where we found that if we optimized WGA staining appropriately on fixed cells, the required concentration goes down significantly (40-fold, from 60 µg/mL to 1.5 μg/ml).
- Phalloidin concentration was diluted 5-fold, from 25 µL/mL to 5 μl/ml.
- Phalloidin also got bumped from an Alexa Fluor 594 conjugate to Alexa Fluor 568, and WGA from Alexa Fluor 594 to Alexa Fluor 555.
- Cimini, et al. Nat Protocols 2023: Details the full protocol, with troubleshooting
- Caicedo et al. Nature Methods 2017 best practices: A collaboration between ~20 labs, this outlines a typical workflow for data analysis and discusses choices at each stage
- Chandrasekaran, et al. Nat Reviews Drug Discovery 2020: A review article describing successes and opportunities in image-based profiling
- Poulsen lab paper: https://www.sciencedirect.com/science/article/pii/S096808961930416X : A nice explainer that adds to the Bray Nature Protocols article, and also uncovers the mechanism of action of a drug (9-methylstreptimidone as a protein synthesis inhibitor)
- Pahl A, Sievers S. The cell painting assay as a screening tool for the discovery of bioactivities in new chemical matter. In: Ziegler S, Waldmann H, eds. Systems chemical biology. New York, NY: Humana Press; 2019:115–126.
- *Pahl A., Sievers S. (2019) The Cell Painting Assay as a Screening Tool for the Discovery of Bioactivities in New Chemical Matter. In: Ziegler S., Waldmann H. (eds) Systems Chemical Biology. Methods in Molecular Biology, vol 1888. Humana Press, New York, NY. Pahl-Sievers2019_Protocol_TheCellPaintingAssayAsAScreeni.pdf
Please see the Cell Painting Gallery (we no longer maintain a list of papers here): https://broad.io/CellPaintingGallery See also the site maintained by the CytoData Society: https://github.com/cytodata/awesome-cytodata
Neurons: Some tips when working with neurons from Matthew Tegtmeyer in the Ralda Nehme lab: “Geltrex has been very reliable for us, across several cell types. It’s from Thermo Fisher cat# A1413201. It’s used at 1:100 in any protein-rich basal media (usually DMEM/F12). (Another lab reported a strong green background when using Matrigel). In terms of the plates, we’ve used a few different ones with similar results. For experiments in 384-well or 96-well plates we use Perkin Elmer Cell Carrier Ultra plates (now know as PhenoPlate). https://www.perkinelmer.com/product/cellcarrier-384-ultra-lid-50x1b-6057300
Organoids: A preprint describes an adaptation of the protocol for 3D organoids/tumoroids: "Evaluating Drug Response in 3D Triple Negative Breast Cancer Tumoroids with High Content Imaging and Analysis" Sirenko et al. https://doi.org/10.21203/rs.3.rs-1859525/v1
Suspension cells: As mentioned by Spring Discovery at SLAS 2021: "Luckily most of these cells can be coaxed into being somewhat adherent under the right conditions. The process involves plating the cells in a serum-free media for 30-60 minutes before layering a high-serum concentration trigger media along with gentle centrifugation. That said, we acknowledge that this process is going to result in some bias in the final populations." Separately, we’ve been told BlueWasher works well for suspension cells for Cell Painting.
How many replicates? Using the TA-ORF pilot genetic perturbation experiment (Rohban et al.) data, Shantanu Singh found that performance drops marginally if we reduce from n=5 to n=4, but drops much more as we drop to n=3. Details:
- Percentage of hits is calculated as n.set.sig.pos / n.set.sig.tot.
- For n = 5, 4, 3 replicates, the percentage of hits is 131/264, 129.5/264, 111/264 respectively.
- The drop in percentage of hits going from n = 5 to n = 4 is 1 - 129.5/131 = ~1%, and from n = 5 to n = 3 is 1 - 111/131 = ~15%.
- We therefore recommend 5 replicates for experiments where every sample is critical; 4 replicates for most large-scale experiments.
An early list of recommended control compounds is available here.
The JUMP-Cell Painting Consortium developed standards and optimized processes to facilitate the worldwide community using Cell Painting. Settling on these standards allows scientists across institutions to align and compare their data. They have created the following compound sets (Side note: the company Specs has assembled these compounds for purchase; for info contact [email protected] and see the Specs website):
- JUMP-Target: Lists and 384-well plate maps of 306 compounds and corresponding genetic perturbations, designed to assess connectivity in profiling assays. JUMP-Target is described here: https://github.com/jump-cellpainting/JUMP-Target
- JUMP-MOA: List and a 384-well plate map of 90 compounds in quadruplicate (corresponding to 47 mechanism-of-action classes), designed to assess designed to assess connectivity in profiling assays. JUMP-MOA is described here: https://github.com/jump-cellpainting/JUMP-MOA
We typically recommend a minimum of 16 negative controls per plate for negcon normalization (which in most cases we recommend - see Rebecca Senft's blog post about how to normalize Cell Painting data.).
The Carpenter lab tested 24 vs 48 hours incubation with compound in a very informal experiment years ago; evaluation was challenging due to the lack of ground truth for the compounds tested. We saw more compounds distinguishable from neg controls at the longer timepoint. I don't believe we ever tested even longer timepoints; our anxiety has been that at very long timepoints one would begin to see major nonspecific toxicity that would begin to make distinct compounds look similar, but we have no experimental evidence of this. I think it is worth testing longer timepoints.
The dyes chosen for Cell Painting were carefully selected to be inexpensive and compatible with each other in terms of wash steps and fluorescence response. All the stains that we ended up using worked beautifully in other wavelengths too, so if you are adapting the staining procedure switching the fluorophore should be pretty safe. Here are some notes on dyes we ended up avoiding in our creation of the original Cell Painting assay:
- Celltrace - bound to too many structures in the cell
- Lysotracker - staining was not compatible with the fixation step in our protocol.
- For the Syto stains, of the 6 or 7 options, we chose the one with the highest affinity for RNA, for the best visualization of nucleoli.
This table contains our notes during the development of the assay that might be helpful if you are trying to adapt it.
Someone asked whether there is a reason we use the more expensive ConA-alexa fluor reagent instead of the similar-spectra ConA-fluorescein. The Carpenter lab does not know the answer, but it could be that fluorescein must be used in higher concentration. The company might be able to confirm whether this guess is true.
Another thing to consider about ConA is that Johanna Nyffeler of the EPA reports that ConA gave them some trouble: "We always dissolved it in buffer at pH 8.3-8.4 until we prepared new buffer that had a different pH. The stain looked very different, much brighter (i.e. we couldn’t use the same acquisition protocol because it was so bright). I continued to analyse the data and saw much more ER endpoints being affected compared to other experiments. We used reference chemicals from Gustafsdottir 2013 and could usually not reproduce the ER phenotypes. My hypothesis would be that in your lab your buffer has a different pH and therefore our results might be different. Another issue with ConA that we still have is that it forms precipitates after staining as fast as overnight. The precipitates are all over the plate and disturb the acquisition. We currently wash each plate before imaging to get rid of those aggregates. The precipitate only forms in the wells. I rarely have precipitate in the stock, although I centrifuge them thoroughly. It happens both if I use frozen or refrigerated ConA."
Yet, Kate Hartland of the Broad Institute reports: "We do not test for pH of the buffer. We rehydrate the ConA to 1mg/ml in 0.1M sodium bicarb, which I make up from powder, 840 mg/100 ml. We use HBSS (at 10x concentration, 14065-056, Invitrogen) diluted with Milli-Q water. I can't recall ever being told of precipitates from ConA interfering with the reading, and we routinely read plates days after they are stained."
If you consider swapping out one of the regular Cell Painting stains:
- This fluorophore spectrum viewer could be helpful, from Thermo.
- Keep in mind that when two stains share the same output channel on your imager, it doesn't help to remove just one of them!
- We ran some rather messy analyses to assess the "value" of each existing Cell Painting stain in order to help guide decisions on which stain to leave out.
- In an analysis from 2011-03-24, we counted the number of wells (samples) in a small-molecule experiment that were "active", i.e., distinguishable from negative control wells. This is a reasonable metric for "how information-rich are the profiles?" The answer was ~350 wells if we included all data from all channels of the Cell Painting assay; if we removed any of the channels of data (other than DNA; one would never want to eliminate DNA staining), the metric is negatively impacted. As you can see, removing the "phgolgi" channel (now called AGP for actin, golgi, plasma membrane) has the worst impact. By contrast, removing ER, SYTO (now called RNA), or Mito have less impact on the "power" of the extracted morphological profiles. These three channels are roughly equivalent in terms of their value.
- In an analysis from 2015-10-12, we came at this question from another direction: For each set of features coming from a channel, how well-correlated are the features' values across replicates of the same samples? Description from Shantanu Singh here: "For one of our Cell Painting experiments, I partitioned the data into two replicate groups (across 2260 pairs in total, 565 unique treatments) and computed the Pearson correlation between these two replicate groups in each of the 1474 features we analyzed. I categorized the features into the channel, compartment (Cell, Nucleus, or Cytoplasm), and very roughly into feature type (Shape, Texture, etc.). The plot shows the distribution of Pearson correlations for each features group. Please note that these features correspond to per-well medians, not single cell."
- The most comprehensive analysis of the impact of dropping individual channels is described in the supplement of Cimini et al., Nature Protocols 2023 (Extended Data Figures 7-9).
- In an analysis from 2011-03-24, we counted the number of wells (samples) in a small-molecule experiment that were "active", i.e., distinguishable from negative control wells. This is a reasonable metric for "how information-rich are the profiles?" The answer was ~350 wells if we included all data from all channels of the Cell Painting assay; if we removed any of the channels of data (other than DNA; one would never want to eliminate DNA staining), the metric is negatively impacted. As you can see, removing the "phgolgi" channel (now called AGP for actin, golgi, plasma membrane) has the worst impact. By contrast, removing ER, SYTO (now called RNA), or Mito have less impact on the "power" of the extracted morphological profiles. These three channels are roughly equivalent in terms of their value.
This is the overall schedule for a batch of plates, giving plate washer protocol names at the Broad Institute: Paint Schedule
Per Kate Hartland at the Broad: "The key automated components are the pin tool (for transferring the test compounds into assay plates - we use a CyBio pin tool), the liquid dispenser for dispensing stains and PFA (we use a Thermo combi, but a Biomek would work too), and a plate washer (rather ancient Biotek but it does the job) for removing liquid and washing the cells (a Biomek may work but is probably a lot slower). We have only run Cell Painting in walk-up mode (myself as robotic arm). This is because I can be faster than a robot, if less precise in terms of timing. We used an ImageXpress for the early Cell Painting efforts (this is what Sigrun Gustafsdottir worked with) before upgrading to the PE Phenix. The files are huge though and can create issues around storage and transfer. We have fully automated robotics systems which can do all of the functions and move plates around at exactly specified time points. However, this process requires that plates be spaced enough apart at the inception of the process that everything can get done in the correct time frame without one plate waiting on another. And robots are slower than people. So overall the throughput declines significantly."
How much does it cost to store and process Cell Painting data? At the Broad in 2018, our AWS costs for storage were $11.2 /plate-month including overhead (so the actual AWS cost was around $6.50/plate-month). Cost of processing was $74/plate including overhead (actual AWS cost $44/plate). This means storing a 275-plate screen (5 replicates x 20,000 samples plus controls) for 1 year on S3 was $36,960 and compute $20,350 (for an experiment with 6 channels (including brightfield), no binning, 9 sites / well, 384 well). Storage is MUCH cheaper for glacier options, if you do not need to routinely access the data.
At Broad, Xiaodong Lu suggested the following seeding densities based on his visual examination under a microscope.
| 6h | 24h | 48h | |
|---|---|---|---|
| A549 | 2,900 | 1,900 | 900 |
| MCF7 | 1,900 | 1,200 | 600 |
| U2OS | 2,900 | 1,900 | 1,000 |
Several laboratories report using assay-ready plates (ARPs) successfully for Cell Painting, where the compound is plated in the correct amount within the well of a dry plate, stored frozen until ready to use, and then medium with cells is added to the plate. Although we are not aware of a side by side comparison of data quality, no downsides have been reported. In particular, no local effects of the compound (e.g. in the original compound-spot location within the well) have been observed. In general, we have used 48 hour compound exposure with ARPs (as compared to the standard protocol with dispensed compounds: seed cells, wait 24 hours, then add compounds to wells and incubate 48 hours).
In step 15, the cells are fixated using PFA. There is no additional wash step and the MitoTracker remains on the cells. The MitoTracker is washed out with the PFA after fixing.
According to Sigrun Gustafsdottir, the fixation needs to be at least 15 minutes, 20 minutes was even better, to avoid cells lifting off the plate and to avoid compromising the signal/background ratio.
We suspect that after fixation, plates can be stored for a period prior to imaging (in HBSS at 4degC, wrapped in foil to protect from light, and be careful about bugs growing, especially if the plates are at room temp for a significant period during imaging). But we are not certain whether there is any information loss by waiting prior to imaging. Collaborators report images looking good after plates were stored for a month, but did not quantitatively test whether or not they're as good as when freshly made: we estimate a roughly 2-fold loss of intensity. We welcome input on this issue, whether qualitative experience or quantitative testing.
(For v2 of the protocol) These files are used at the Broad for preparing the stock solutions and calculating amounts to order for a given number of plates. Supply Prep Calculations, Stock Solution Prep
In one experiment, we saw illumination gradient effects occur in the w3 (SYTO) channel (contact us for details - internal reference: 2008_12_04_Imaging_CDRP_for_MLPCN_%28Imaging_Platform%29:Part_III%28March-December_2010%29#2010_05_07_Email:Illumination_correction_in_SYTO_channel) and w5 (Mito) channel (contact us for details - internal reference: 2008_12_04_Imaging_CDRP_for_MLPCN%28Imaging_Platform%29:Part_III%28March-December_2010%29#2010_11_04_Email). In the SYTO case, this was noticed by Sigrun Gustafsdottir while training for a phenotype in which requests for positive cells always returned (i) cells from the same edge of the image, (ii) that appeared to be normal cells. (contact us for details - internal reference: 2008_12_04_Imaging_CDRP_for_MLPCN_%28Imaging_Platform%29:Part_III%28March-December_2010%29#2010_05_04_Email). In both cases, the artifact is characterized by an uneven illumination pattern in which one image edge is excessively dark. Usually this is apparent when checking the illumination correction function. All wells' images were affected, but for one channel only.
Solution: Filter fix by the vendor. The channel still has a slight gradient, but much reduced and correctable for by a posteriori illumination correction (contact us for details - internal reference: 2008_12_04_Imaging_CDRP_for_MLPCN_%28Imaging_Platform%29:Part_III%28March-December_2010%29#2010_05_20_Update_on_SYTO_staining)
Be consistent in the order images are acquired; it's possible that some channels bleach a bit when imaging other channels. We advise imaging from red to blue, so that the emission of one fluorophore doesn't excite a fluorophore with overlapping emission/excitation wavelengths before the image for that channel is captured.
- See this page for an explanation of Cell Painting features, as derived from CellProfiler.
- Current Cell Painting pipelines updated for recent versions of CellProfiler can be found here: https://github.com/broadinstitute/imaging-platform-pipelines/ (note the subfolder for JUMP_production pipelines, related to the JUMP Cell Painting Consortium).
- Rebecca Senft of the Cimini lab at Broad Institute wrote a blog post about how to normalize Cell Painting data.