Edit January 19, 2023: The 2021 Nature paper by Carlos P. Roca et al. introduced AutoSpill, which resolved the gating issue raised in this blogpost. In essence, AutoSpill uses iterations of linear regressions to get the best compensation or unmixing matrices, and removes the need for gating the positive and negative fractions of your controls. It saves you time, and generally provides better results. Learn more about AutoSpill here.
Edit October 21, 2021: It should be noted that the tips for gate placement described in this post can be applied to spectral unmixing as well, just replace the term “compensation” with “unmixing”. For more tips on spectral unmixing specifically, see this video.
When calculating compensation, automated tools are the gold standard. However, people often struggle to get good results from the automated compensation tools and will turn to manual compensation to fix any errors. Why is it difficult to get accurate results from automated tools? The only explanation that I’ve heard from the flow cytometry experts is that suboptimal results are caused by poor quality compensation controls. In this post I’d like to demonstrate that poor controls are not always the cause. The part that is often overlooked is that automated compensation tools are not fully automated – they require users to subjectively set gates on positive and negative cells within each control. Yes, some of the tools do have automated gate placement, but that feature often fails to generate appropriate gates. Gate placement is a critical component in determining the outcome of an automated compensation tool.
Setting Up the Automated Compensation Tool
Before I discuss gate placement I want to quickly go over how autocompensation tools work. Compensation determined to be correct when the median fluorescence intensity (MFI) of the positive population is in line with the MFI of the negative population (Figure 1). The flaw of automated compensation tools is that the user usually determines where to set the MFI for each population by placing gates on the positive and negative populations. Thus, the accuracy of the compensation is dependent on a subjective gate placement.
To demonstrate the effect of gate placement I’m using FlowJo’s automated compensation tool, though this should be applicable to any comparable software. I tested four different gate positions (Figure 2):
- The most positive (brightest) of the positive cells, which cuts through a high-expressing population and would never be used for analysis
- The whole bright/high-expressing population of the positive cells, assuming this control has a bright population and a dim population
- The entire positive population of cells, including dim and bright cells
- The whole dim/low-expressing population of cells, avoiding the bright/high-expressing cells
Results of Automated Compensation
To check the accuracy of the compensation determined by the automated compensation tool, I applied the compensation matrix to the same file and checked 5 plots. Again, the compensation is correct when the MFI of the positive and negative populations match. Here are my results:
The compensation using the first gate – on the most positive of the positive cells – gave the most accurate results. The second gate was nearly perfect, though two of the 5 plots are very slightly undercompensated. The compensation from the third and fourth gates produced suboptimal results with significant compensation errors. These data highlight that setting gates for autocompensation is very different from setting gates for analysis. The first gate would never be used to analyze data, yet it provides the best compensation results. The other gate options could be used for data analysis, but clearly do not work as well for calculating compensation.
Why does the first gate work so well to calculate compensation? If you remember the rules for compensation controls, one of the rules is that the control must be as bright or brighter than the sample. Compensation can be calculated for cells that are the same intensity or dimmer than the control, but calculating compensation for cells brighter than the control does not work well. Since compensation is calculated by MFI, that rule should be translated to state that the MFI of the control must be as bright or brighter than the sample. The first gate works the best for calculating compensation because the MFI (shown as the dashed vertical line in FlowJo in Figure 2) is the highest. In all the other gates, there are cells that are brighter than the MFI used to calculate the compensation, so those gates have essentially created a control that breaks the rules of a “good quality compensation control”.
In practice, my recommendation is to use an automated compensation tool, but know its limitations and always check the accuracy of the compensation prior to analyzing data. For best results, use high quality compensation controls and set the gates on the brightest cells. When setting the gates on the controls, remember that you are not analyzing the data, it is perfectly acceptable to cut a through a population with the goal of calculating a high MFI.