Version 1.0.0: Faster unmixing

[DrCytometer]

• Version 1.0.0 brings a revamp to how AutoSpectral identifies the best spectra on a per-cell basis. The theory behind it remains the same--we are still trying to identify the variation in the autofluorescence and fluorophores that best reduces the residual on a per-cell basis. Now, however, we do not need to do that using brute force. Instead, we can search only through variants (or autofluorescences) that align with a given cell's residual. Thus we can pre-screen the variants to a select few and then test just those. This means we can figure out the solution in way less time. It also means that a native R implementation of the algorithm is possible in R in a somewhat reasonable time frame. So, that may help for anyone struggling to use the fast C++ version in AutoSpectralRcpp.
• Since we can now quickly identify which variants are useful for a given cell, we can test more variants, allowing a finer-grained view of the variation, which may improve unmixing quality.
• Autofluorescence extraction and fluorophore variation extraction are now modified to allow searching for more variation, focusing on "problematic" cells that remain far from where they should be when the first batch of variation is applied. This is most helpful for extracting autofluorescence in complex tissue samples, where AutoSpectral previously struggled to deal with the last few messy cells. Use "refine = TRUE" for this.
• Speed in unmixing should be the biggest change, particularly if you run using AutoSpectralRcpp.
• When extracting autofluorescence using get.af.spectra() with "refine = TRUE", you will now get a set of plots showing you the unmixed data for the channels most affected by the autofluorescence ("worst channels"). The same channels will be plotted after a single round of autofluorescence extraction per cell (as in AutoSpectral v0.9.2 and earlier) as well as after the second round, using data from more difficult
  cells. To see this, run with refine = TRUE.
• Autofluorescence is now assigned to each cell using a shortcut to "project" where the AF will impact on fluorophore or residual space. This is especially fast for residual-based assignment.
• Perhaps most importantly, discontinuities that sometimes appeared in the data after unmixing using per-cell-fluorophore optimization, particularly with the "fast" approximation, should now be gone or at least greatly diminished.

---

This discussion was created from the release Version 1.0.0: Faster unmixing.

[claude-chew]

Installed dev version 1.1.0 after not using AutoSpectral for a while, all steps are noticeably faster, thanks for maintaining this @DrCytometer!