What would you like to see added to AutoSpectral?

[DrCytometer]

Are there tools that would be useful to have for assessing the unmixing, spectra, matrices, etc.? Do you want to see other unmixing methods, such as TRU-OLS? Do you need NxN plots?

[SamGG]

Hi,
I am not experienced in unmixing, so I have basic questions and remarks. I read your article, but I confess, I partially understood it as there is such huge knowledge, experience and work.

A) I downloaded the Sony ID7000 dataset you provided, and I noticed that in every raw/mixed channel the minimum is negative. Is it attended? I replaced negative intensities with zero and redo OLS unmixing on a few cells but didn't notice obvious change.

B) from the OLS or WLS current implementation, I think that the auto-fluorescence spectrum is not removed beforehand. Is this really what vendor programs are performing? I thought their program is integrating auto-fluorescence in some way. Is there a way to remove auto-fluoresence before applying OLS unmixing?

C) "Our computational pipeline uses the per-cell residual signal following application of the unmixing matrix to calculate the optimal autofluorescence subtraction, being the profile which minimises residual". There is something I don't catch here. What I think about is that the pipeline looks like (1) applying unmixing matrix to the observed fluo, (2) considering the residual as the auto fluo, but it is too basic. I wish there is some pseudo for helping me. I think there are two parts that should be very well separated : determining the spectra of each marker and unmixing taking into account auto-fluo.

Best.

[DrCytometer]

So, I probably don't have everything right either. I've mostly been learning this on my own, and I do not have a background in linear algebra or computational science.

A) Yes. In almost all cytometers, negative values are allowed in the raw data. These are usually capped at some values. I suspect this has to do with the internal pre-processing of the signals coming from the PMTs/APDs, but that is beyond me. As you saw, this will have little impact on the unmixing in most cases because this is but a small part of the electronic noise. The A8 and S8 may be exceptions here--they have a tendency to kick out much greater negative values in the raw files in samples with high levels of positive signal.

B) I believe I am do the OLS and WLS autofluorescence unmixing as it is done on the vendor platforms, but I could be wrong. For a single autofluorescence parameter, you pass a single autofluorescence spectrum as a row in the spectral mixing matrix (spectra). This gets unmixed ("extracted") as a separate parameter in the unmixed data. It is not removed beforehand as that would effectively be background subtraction (background subtraction is fairly useless). AutoSpectral automatically generates a single median-like AF signature that you can pass to do AF extraction in OLS and WLS unmixing. In support of this being what happens on vendor software, you can see Autofluorescence parameters in both the SPILL keyword from BD files and in the .expt files from Cytek. If you want to do multiple autofluorescence extraction, this just means adding additional rows to the spectral matrix.

C) I'm not sure I'm following entirely. AutoSpectral can be thought of as at least two pipelines: 1) optimizing the extraction of spectral signatures from single-stained controls and 2) application of those signatures in the unmixing calculation(s). In regard to the unmixing workflow, a basic OLS or WLS unmixing in AutoSpectral follows the clear separation you describe: first we identify the spectral signatures, then we unmix with them. In OLS and WLS, even with multiple AF extraction, there is no accounting for variability in autofluorescence at the cellular level.

We do not do this: "considering the residual as the auto fluo"
The residuals are the part of the data explicitly outside the model with least squares. Including the residuals as an AF parameter, by definition, does nothing to the unmixing in the other channels. I know this because I tried it before I realized why it was that way. The residuals actually don't have much information left in them in high parameter datasets because almost all of the raw detector information gets converted into unmixed data. The problem is that the conversion is not really correct in many cases. An unstained sample unmixed with a large panel of fluorophores will have minimal residual--all the data goes into the model and manifests as AF.

What the residuals are good for is determining the quality of fit of the model. The lower the residuals, the better the prediction of the model for the raw data. In OLS and WLS, we fit a single spectral matrix to the entire cell dataset, which is a good overall fit. At the level of the cell, though, there is variability, particularly in the autofluorescence. So, we can work to minimize the residual for each cell.

In the unmixing pipeline, we first get an initial unmixing of the data using only the fluorophore spectra, no AF extraction at all. Then, we cycle through each of the identified AF signatures, tracking whether each one improves the fit per cell (or minimizes the absolute signal in the fluorophore space--the dist.0 metric, which works better). We do this by adding a single AF signature to the spectral mixing matrix in each iteration, so there is effectively one degree of freedom in the unmixing, with that freedom being the AF signature. Once all AF signatures have been checked, each cell has its best AF signature (from the choices provided). This goes on the understanding that a single cell can only have a single AF background.

There is a second part to the unmixing pipeline, the per-cell fluorophore optimization. This is where it gets a bit complicated. This second part addresses variability in fluorophore emissions, in much the same way we deal with variability in AF. There are a lot of reasons why fluorophore emission can vary at the level of the individual cell, but basically, it's tandem dyes, including Brilliant tandems, that are the major source of variability. We can measure the variation in output from the fluorophores using the single-stained controls. We are used to seeing this variability as spillover spread, although the variability I'm addressing is not what is usually discussed as the theoretical basis for spillover spread. When we measure the spectral signature of a fluorophore, we are effectively taking the median/mean of the distribution of emissions in the raw space. We can, instead, sample multiple points in that distribution, all of which are completely valid if they come from the single-colour controls. Allowing each cell to vary between any of those points has the effect of reducing some of the variability. If you haven't read through the help page on this topic, maybe have a look at that. It's a little hard to me to explain, but it can be thought of as analogous to weighting. With weighting, you emphasize or de-emphasize certain detectors based on how reliable we consider their measurements to be. With spectral variation, we allow the spectral signature for any given fluorophore to vary within the restricted, measured range, selecting the optimal variant for any given cell based on which variation best "fits", as determined by minimization of the cell's residual.

[SamGG]

Thanks for your feedback and patience. I really appreciate.

A) OK.

B) I forgot that there is an AF spectrum in the spectral matrix and that it will compete with the spectrum of each marker. My bad.

C) Thanks for your clear explanations. It helps me a lot. I will keep the variability of fluorpohore emissions for later.

Concerning the extraction of per-cell AF from an unstained sample, I have some questions. If it is easier for you to point me to Web page or a paragraph of your article, don't hesitate. I feel you provide some much information that I don't know where to look. Actually, I prefer to read and understand the code rather than sentences as I interpret the latter with my own prism without really undertanding what happens under the hood.

D) in get.af.spectra(), the FCS file is not cleaned up nor filtered (as the processing of single-stained controls, I mean the negative/unstained cells). Should it be cleaned up beforehand?

E) in get.af.spectra(), what is the rationale concerning the data used to feed the SOM algorithm in order to create clusters of AF? from your experience, what the OLS unmixed expressions bring to the clustering? from a naive point of view, why the FSC and SSC channels are not included for clustering?

F) in unmix.autospectral(), I think I do understand the first part (af.only). The selection of one AF spectrum among the AF spectra determined by the previous clustering is made from an optimization point of view, ie minimization of residuals between the raw spectrum and the adjusted contribution of the markers and the tested AF. It seems to be the only to assign an AF spectrum to a cell. To evaluate the gain of this approach vs the us use of the global AF, we have to read bi-parametric plot. Right? Did you try to store the residual spectrum as a FCS file in order to analyse/view it with a commercial software?

G) From a naive point of view, I would cluster the unstained sample using the FSC and SSC channels to extract AF spectra. Then I would map stained cells using FSC and SSC channels. As you certainly thought of this approach, what are the pros and cons? Do you feel/determine the cells within gates defined in the FSC/SSC bi-plot are not homogeneous when looking at their AF spectrum? Do you think FSC/SSC could be used as a matching criterion when assigning an AF to a cell rather than relying only on the residual minimization?

[DrCytometer]

C) Single cell AF article
D) No, I chose not to do any clean up via gating or anything else. This means 1) the workflow will work no matter what type of sample is supplied, and 2) every cell and particle will have autofluorescence extracted as best as possible so any downstream analysis will be improved, no matter what.
E) FSC and SSC are useful, but provide another failure point. Scatter parameters are the most commonly affected by fluidics errors, so any input on those is problematic. Combining unmixed and raw data as input helps, empirically. Theoretically, it helps because we are translating the raw data into the unmixed space, where distortions of the negative through unmixing-dependent spread will be seen and can contribute to clustering. Those distortions are some of the most impactful on data quality.
F) Please read the pre-print and help articles.
G) Doesn't work well. Not enough info there. As you said, not homogeneous. And, you're using 2-6 channels to model 60+ channels of information--that's essentially impossible to do well. You can try incorporating it into the model. I see it as highly failure-prone based on the types of data I get from people.

[SamGG]

Thanks for the feedback. I will wait for the trials of my colleagues.

[claude-chew]

Would like a way to submit additions to the fluorophore database. I have been manually adding fluors that I come across that aren't in that .csv file but not sure if there's a way to compile them so that they appear in future releases of Autospectral?

[DrCytometer]

This has been added in the last update. See README for info. Links to relevant Google sheets below:
Fluorophore database
Marker database

[claude-chew]

Thanks, I'm going to update this as I come across fluors that aren't currently in the database