Implement cell type assignment methods

[nictru]

Description of feature

The following methods look interesting:

Tasks

Beta Give feedback

1. Implement celltypist cell type assignment #21
   enhancement +1
2. Implement Azimuth cell type assignment #22
   enhancement +1
3. Implement scPred cell type assignment #23
   enhancement +1
4. Implement SingleR cell type assignment #24
   enhancement +1

[grst]

What's your take here on cell-type assignment by atlas-mapping? In my experience, these approaches work best, but are of course limited by the availability of such an atlas and are conceptually different than the methods mentioned above.

The future will bring "foundational models" that are trained on all available data. scsimilarity goes in that direction -- it generates a low-dimensional embedding and cell-type labels in one go.

The Theislab is working on nicheformer.

I would keep that in mind during the pipeline design that you could jump to cell-type labels + low dimensional embedding in one step.

[nictru]

This sounds very interesting.
Implementation-wise the pipeline can produce labels+embeddings in one step already. An example of this is how the SCANVI process is implemented.

[vd4mmind]

If I may add, first iteration of cell type label assignment is better with some of the well benchmarked tool untill foundational models are fully operational and trustworthy. From cell type classification purposes, this first iteration label + embedding can still be reassigned and improved with human intervention if labels are a bit off, which will be case-specific (most likely depending on organ/tissue and disease type). So atlas level data will always serve as better ground truth training and for classification purposes with any of the transformer based or pure ML or Neural Network based models in future. So if such comes better, that can be implemented as a process.