Novel analytical pipeline reduces spectral flow cytometry errors up to 9000-fold
Tuesday, October 28, 2025 at 10:36AM Pre-print alert! And this one really is a must read for anyone that does spectral flow cytometry. It is a complete, fully-automated spectral unmixing pipeline that reduces error up to 9000-fold, created by our cytometry guru Oliver Burton, of Colibri Cytometry fame.

We've all seen the problems - spreading, skewing, autofluorescence intrusion. Unmixing errors are so ubiquitous in high parameter panels they are often thought of as unavoidable, intrinsic to the way the hardware works. Surprisingly, they are largely artefacts of the unmixing software being used.

The problem is that spectral unmixing is complex. The basis is a linear regression of positive versus negative signals, a highly error-prone process. This issue is largely solved by the use of robust linear regression with iterative rounds of improvement (which we pioneered with AutoSpill). However there are three additional problems, which become bigger the more fluorophores are used:
1)This unmixing solution still requires ideal positive-negative matching to find the right linear regression. This isn’t trivial, as the cells positive for one marker might have completely different autofluoroscence profiles to the cells positive for another marker. Using the same negative population gives you spillover calculation errors.
2) Cells have variation in background fluorescence. An unmixing matrix that doesn't account for autofluorescence will force all signal into one of the flurophore channels, giving misassigned signal. Past approaches only use a single autofuorescence index, which means heterogenous mixtures have cells with misassigned signal.
3) Fluorophores actually stuck on cells have variation in emissions, and using only a single profile will lead to misassigned signal on some cells.
Some of these problems can be tackled (partially) by a highly skilled flow cytometrist, willing to spend days on each unmixing matrix, manually selecting populations for positive and negative cells and running multiple sets of calculations depending on which markers they want to assess. AutoSpectral does it all in a completely automated pipeline, using a robust statistical model that is highly reproducible and visibly reduces the error.
For positive-negative calculations, intrusive events are purged and scatter-matching is used to identify the suitable negative population for each positive population. We then use robust linear regression with iterative improvement to find the ideal unmixing matrix.

We can also deal with heterogeneity in the cells by identifying all autofluorescence patterns in the unstained sample, then applying each pattern to each individual cell in the real sample. We select the autofluorescence index that leaves the least residual, subtract that signal and unmix the rest.
The same is true for fluorophore variation - we can test the different fits on a per cell basis, and use the fit that leaves the least residual. It means more signal is attributed to the correct fluorophore.
The cumulative effect of these improvements is enormous. For tough samples, like lung, incorrectly assigned signals are reduced by up to 9000-fold, and a 10- to 3000-fold improvement is common. We demonstrate the improvement in synthetic experiments with known ground truth, and multiple real-world complex panels, where we can use known biology to see the improvements. For example, look at this experiment, where the wildtype has no GFP signal and the GFP transgenic should have GFP in CD4 and CD8 T cells. Since this sample is from the lung, the autofluorescence of macrophages gives a huge GFP signal in the wildtype mouse, which completely confounds the genuine GFP signal in T cells. Switching over to AutoSpectral, and exactly the same samples, with exactly the same cells, behave just as you would expect, little signal in the wildtype and a CD4+ and CD4- population in the GFP transgenic.

The whole pipeline is available right now on GitHub. Don't be intimidated by R, it includes comprehensive notes on every step from installation to utilisation, and only takes a couple of minutes to run per experiment. Hopefully soon (like with AutoSpill) it becomes standard on commercial platforms too.
Full article is available here.
flow cytometry 











