Can spectral unmixing extract an unknown channel?

Hey there!

I haven’t used spectral unmixing even for simple cases so I apologise if my question is naive or too simple.

Say I have a (technically) two-channel regular widefield image with overlapping spectra, where:
Ch1: Known spectrum of a protein
Ch2: Inconsistent spectrum (varies from region to region of the cell in a range of about 50-60 nm)

Is there any plug-in or an existing algorithm to extract the spectrum of Ch2 for each pixel within the cell from the widefield image - the idea is to later use these spectra for further post-processing.

Thanks in advance! :slight_smile:

Some clarifying questions:
You say each channel is a spectrum; generally I think of that meaning you have measurements from many wavelengths, collected with the same excitation. Are you sure you don’t mean a known fluorophore that you don’t believe has any bleedthrough? While channel 2 probably has some degree of bleedthrough?

If that is the case, you might be able to use a sample with no second channel signal (whatever you are trying to unmix from), which would allow you to get a good read on the bleedthrough, which should be linear. If the ratio of channel 2/channel 1 pixel values is ~X, you can then subtract channel1 *X from your second channel in future images (essentially anything that does linear unmixing). If you don’t have a sample like that, you would need to find a best estimate of an area with no “real” channel 2 signal, though that might be difficult to justify down the line.

If you do mean different channels having spectral information, maybe you could give more detail on what type of images you are analyzing? I am not familiar with that!

Thanks a lot for the quick reply. I don’t fully understand your question (due to my lack of experience in this area, nothing to do with the actual question!) but I hope the following info helps:

My Ch2 is a solvatochromic dye (it changes its emission depending on its environment), that’s why I can’t give a reference spectrum for it. In my case the bleedthrough would come when I excite the protein channel (Ch1), because this will also excite to some extent the solvatochromic dye as well (and no I haven’t found a reliable fluorescent protein that manages to evade this problem). When I excite only Ch2, there shouldn’t be any bleedthrough in the protein channel if I choose the construct wisely.

The images will be acquired on a widefield fluorescence microscope, NOT spectral imaging with a confocal microscope, but I was told spectral unmixing is still possible nonetheless even with widefield images (?).

Another issue which I guess might occur is the signal in the protein channel will be from a transfected fluorescent protein, so some cells will have a lot brighter signal compared to others (the staining won’t be even). My understanding is that with spectral unmixing the intensity of the emissions should be known and be constant among all cells, so that your suggestion with the subtraction will work?

Thanks again!

With linear unmixing, in a perfect world (hah), and assuming you had a sample with no dye, your unmixed (math above) calculation would result in an almost blank channel 2. Then, any further channel 2 signal would be only from the dye.

It sounds like, though, the problem is additional signal in channel 1… which is more problematic. You now have three variables and only two equations, which doesn’t seem solvable. If you could find a channel where you knew your protein signal was pure (even if not optimal), you might be able to use that. Otherwise… I dunno :slight_smile: Sounds very math unlikely.

Pixel in channel 1 = protein+dye1
pixel in channel 2 = dye 2
If the dye1 to dye2 ratio (their relative emission strength in each channel) isn’t fixed, you cannot create an equation to solve for your variables, I think.
If you can get a protein only channel, then you can solve for at least dye1, with:
channel 3 = protein.
channel 1 = protein +( dye1=0) [in a sample with no dye]
Once you have that ratio, you can add the dye and extract the intensity of dye1… though at that point you may as well only take channel 2 and channel 3!

That said, I am basing this on the assumption that the emission change for your dye is in relative channel intensity, not just emission strength. IE the emission peak is shifting. If it just gets brighter or dimmer based on environment (but the ratio of a pixel in channel1 to channel 2 will always be the same), you should be able to eliminate the signal from the dye in Ch1, and ignore what I said above. Sorry for the somewhat meandering post.

1 Like


Thank you so much for the explanation, I really appreciate it!
I can easily make a sample without any dye in order to get a reference spectrum of the protein tag, this isn’t a problem.
However, whatever way I think about it, it seems I will have some dye bleedthrough in the protein channel anyway, which as far as I understand is the bigger problem, since it is impossible to establish a common spectrum shape or fraction relative to the protein emission.
I will also have protein bleedthrough in the dye channel(s)
You are right in that the emission peak is shifting along the spectrum.

Yes exactly, the intensity ratio changes. It is technically only 1 dye, but I image it in 2 channels to detect the emission peak shift.

I was hoping that knowing only the protein emission spectrum will be enough to unmix it from any other bands, if the spectrum shape and intensity are sufficiently different, but I see it is more complicated than that.

Thanks anyway!

1 Like