Scale for predicted vs observed plot in SCVI-SCANVI mapping?

Help scvi-tools
1

I’m curious to know:

  1. is the plot generated in scanvi mapping where yellow and green I assume are signifying similarities—publishable? If so, would adding a scale bar which could tell for eg yellow being the most similar and green almost similar help the readers?
  2. I was able to map my in vitro dataset to an in vivo dataset. Among these, can I annotate my unlabelled clusters with the ones that have yellow/green level similarity with the in vivo dataset? Overall it had an accuracy of 62%.
    IMG_6676
    IMG_66761179×1158 163 KB
2

These signify confusions. Generally high confusion means similar cell-types. However, it doesn’t need to be that all cells are similar but can also mean that some cells within a celltype are more similar and some are less or that the frequency of this celltype was low in the reference data. I would recommend using a heatmap of marker genes to distinguish these cases

1 Like
3

A follow up question: if the markers genes are not easily distinguishable (because my dataset has a healthy and 2 patients), would you recommend going ahead with the labels that get predicted even if the accuracy is 62%?