Today’s post is about how to get the best results with spectral unmixing. Poor quality unmixing is the most common problem I help people with, and those unfamiliar with large panels often don’t realize there’s any issue. I’ll lay out my guidelines for preparing controls, using those controls, and understanding whether you have unmixing issues. While this post focuses on spectral unmixing, the principles are the same for compensation on conventional machines.
The bottom line: For your single-color controls, the best thing you can do is prepare cells by staining them exactly the same way as in your fully stained samples.
First, a bit of background. I’m not a mathematician, and my suggestions here are based on empirical testing rather than theory. I started out in flow cytometry doing three color experiments on a FACScan, performing compensation manually. As new machines came out and the number of markers in my experiments increased, I tried to use automated compensation tools, but found the results to be subpar. By the time I got up to running 28-color panels on the FACSymphony, I was pretty frustrated with doing manual compensation. As you might imagine, editing a 28 x 28 matrix until all 28 single color controls are correct takes some time, and I usually had to repeat the process a couple of times per experiment to get it close. Fortunately, we had a mathematician in the lab, Carlos Roca, who thought what I was doing was ridiculous, and programed the computer to do it instead. The result of that was AutoSpill, which is now integrated into FlowJo. Compensation with that approach reduces the spillover error by around 100,000-fold. If you’re interested in learning more, you can check out the paper.
Step 1: Recognize you have a problem.
How can you recognize the spillover errors that result from incorrectly compensated or incorrectly unmixed data? Here are my tips:
In your single-color control cells the positives are above or below the negative with respect to another channel. You don’t have single-stained cells? We’ll get to that next.
There are asymmetrical hypernegative events. Spreading errors are symmetrical, so you will have some events below zero, particularly in poorly designed panels. Populations that are distinctly negative or leaning negative as they increase are likely spillover errors.
There are positive correlations in the data. Very few markers we work with are linearly correlated, an example of a pair that is being TCR and CD3.
In less extreme but more common versions of 1 & 2, the data curve up or down as the expression increases. Again, suspicious. In a best-case scenario, you may have an FMO control for the marker in question, and you’ll see that the “positive” doesn’t disappear in the absence of staining.
There are biologically impossible populations. Once with a mixed bone marrow chimera (a mix of cells expressing CD45.1 and CD45.2), I was looking at CD4 vs CD8 and had four distinct, well separated populations. It wasn’t the thymus.
The staining pattern doesn’t look like it did the last time. There are many things that can cause this, but incorrect unmixing or compensation is one.
You have massive spreading in the negative; for example, you can’t tell where the positive begins.
Step 2: Prepare better controls.
My recommendations for controls in order of preference:
Use cells treated identically as in the full panel.
This means the same antibody at the same dilution factor.
This means the same number of cells in the same volume.
This means using the same buffer(s), including fixatives and permeabilization agents.
Same staining time.
At least as high of expression as the highest expected in your panel.
In the case of tandems, it’s also important to use precisely the same source of cells as much as possible. Tandem breakdown is catalyzed by cellular enzymes.
Cells stored in reference library.
Run your cells on another day when you can really focus on preparing good controls.
Beads or CD4-type staining.
Both of these are slightly inaccurate compromises. Both are particularly bad for labile tandems.
• A suggestion: prepare multiple controls and determine empirically what works best for you. If you’re struggling to get good controls for a marker, prepare 1) the single-stained cell controls with that antibody, 2) cells with another antibody with the same fluorophore against a bright common antigen like CD4, and 3) beads with the first antibody. Try the unmixing with all three and validate the results against the cells stained with the first antibody only.
• If you need compensation and unmixing, something has gone wrong.
Step 3: Get the best out of those controls.
Now, let’s look an example. For unmixing examples, I’ll be using this 30+ color panel.
Everybody loves to use comp beads because they’re easy. When you unmix or compensate large panels with comp beads, however, the unmixing is not correct and we get these spillover effects.
Why I don't recommend compensation beads:
By looking at the single-stained cell controls, we can see that the spectra identified by using the comp beads produces unmixing that is inaccurate for cells. This is not due to tandem breakdown or brilliant violet interactions. The larger your panel, the more inaccurate comp beads will be. Also, the more comp bead controls you use, the less accurate it will be. If you only need to use comp bead controls for a couple of markers in your 40-color panel, you’ll probably be fine.
Single-stained cells unmixed with comp beads:
So, we need to use cells, but cells are hard because there’s autofluorescence and the signals can be dim. How do we fix this? In this example, I’m using ICOS in BV605. ICOS is expressed on activated T cells, which aren’t super common in healthy mice, and BV605 gets interference from macrophage autofluorescence. If I set a wide scatter gate, I see uncertainty in the signature due to the combination of BV605 and autofluorescence being detected.
Here’s that macrophage AF profile for comparison so you can see how it’s mixed in there.
By unmixing with this gate and the resulting uncertainty and inaccuracy in the spectrum, both the single-color and fully stained samples have spillover unmixing issues.
So, we need to better identify the ICOS-expressing cells and exclude the non-specific signals. In this case, we want T cells, so I can tighten the scatter gate and focus on the brightest events. This gives us a clean BV605 signature.
And the unmixing improves:
In this case, I set a small lymphocyte gate because I had a T cell marker. If I’d had a neutrophil marker, I would have needed to set the scatter gate to include high SSC events because that’s where my real positives are. How do I know where my positives are? If you don’t already know from experience or reading the literature, you can find out by back-gating or scatter gating. To back-gate, gate on the positive events in the channel with the expected peak emission (e.g., V1 for BV421), then plot these events on FSC/SSC. Where are they? Set your FSC/SSC gate to include those. Alternatively, just move the FSC/SSC gate around until you see where you have signal.
Step 4: Check against a reference spectrum.
How do we know if we’ve isolated the correct signature from our cells? There are a couple of tools for this. First, you can look up the spectra for most fluorophores on websites these days. Cytek, Sony, BioLegend, Thermo and BioRad all have spectrum viewer tools that will show you the expected emission profiles. Second, you can run an easy control such as compensation beads or a bright clean marker to see what the normalized spectrum should look like in the absence of noise.
Some spectrum viewer tools:
Step 5: Troubleshooting.
Problems and suggested solutions:
My marker isn’t expressed enough to use cells.
If you plan to measure it by flow in your panel, it is probably expressed adequately to use as a single-color control. If not, why is it in your panel? You should be acquiring at least 100,000 cells, more if your marker is really rare. Ideally you should have a couple hundred real positive events in your file. If you know the frequency, you can figure out how many you need to acquire.
My marker isn’t expressed in my control cells, only in X, Y, Z (tumor infiltrate from patient, cytokine in stimulated cells, transduced cells).
Usually we actually have control over this, but the issue is that it’s more work. For instance, if the marker is only present after transduction or in a transgenic animal, prepare some transduced cells and freeze them for use as controls on demand. If the marker is only expressed in a specific tissue, you can isolate additional cells from that source for use as controls.
On the new spectral machines, you can store your controls in the reference library. These can then be used in place of fresh controls if needed. These library controls are updated with the machine QC, but will eventually depart from reality. Using a few of these is okay, but the larger your panel or the older the controls, the less accurate it will be.
If you really can’t get ahold of cells that express your marker, design your panel so that you’re using a stable, simple fluorophore (e.g., AF488, PE, BV421) and use another antibody conjugate to the same fluorophore from the same manufacturer against an easy, high expression target (e.g., CD8).
I don’t have enough cells.
If you don’t have enough the same day but can get more soon, do that.
If you can get similar cells (e.g., buffy coat instead of PBMCs), that will work.
My marker is expressed on highly autofluorescent cells.
This approach works with autofluorescent cells provided you extract the autofluorescence of those cells as a separate channel. I’ll cover that in another post.
In this situation, the autofluorescence can be combined with the fluorescence of the fluorophore in that channel. For instance, the channel for your eosinophil marker can now include the eosinophil autofluoresence plus Siglec F or CCR3.
If you’re doing really high-depth phenotyping on autofluorescent cells, you may wish to use a reference negative that is the unstained autofluorescent cell population.
Use a different antibody (same fluorophore, same manufacturer) against a non-autofluorescent cell type.
Use beads. Using a few bead controls in a large panel can work okay, particularly if the marker isn’t highly expressed so that any error is also small. The more bead controls you use, the more error will creep in.
Some of my samples are correctly unmixed but others aren’t.
Most likely, all of them are incorrect but it is easier to see in some.
Alternatively, you did something different to some of your samples. Some may have been left out longer, left longer before acquiring, or you may have different tissue sources with unaccounted for autofluorescence.
Preparing so many single stained cell controls is hard.
Not as hard as preparing a properly stained full panel.
I like using beads.
Good for you.
I’ll finish by saying that I think there’s quite a bit of work to still be done to improve unmixing, both with the controls and, mainly, the software. There are new approaches to synthetic controls being developed (like Slingshot particles), and perhaps we’ll reach a point where these are as good as cells. If anyone thinks they have compensation beads that work as well as cells, I’d be happy to test them. If you want to play around with some alternative unmixing approaches, I suggest you check out Mirek Kratochvil and Abhishek Koladiya’s PanelBuilder tool. This is an RShiny app, so it’s easy to use.
