Proof of principle on mouse gastrulation atlas and chimera data

emdann commented 2 years ago

Which data

Missing blood lineages - DA logFC < 0 and significant
Tal1-/- specific endothelium: more abundant (at E8.5) and showing differential expression compared to atlas cells (higher uncertainty), upregulating Nkx2-5, Mef2c, and Tnnt2 (cardiomyocytes genes) and Pcolce, Tdo2, and Plagl1, but still expressing endothelial genes (Esam and Sox17)

[x] Save objects in h5ad format + metadata yaml
[x] Train scVI model on atlas
[x] Train scVI model on control
[x] Map WT and KO on atlas
[x] Store uncertainty for each design object: scArches uncertainty, CellTypist uncertainty (only for PA and PAC, doesn't make a lot of sense for PC), ...
[x] Make neighbourhoods: for PAC embedding of P+C
[x] Store mean nhood confidence
[x] Compute DA logFC per neighbourhood: PAC = P vs C / PA = P vs A / PC = P vs C
[x] Plot DA logFC vs mean confidence

Does the design sensitively detect the missing blood lineages? annotate neighbourhoods using the control annotations (C in PAC), compute TPR for clusters annotated as Blood progenitors or Erythroid (where positive == logFC < 0 and SpatialFDR < 0.1). Consider FP all the other predicted depletions.
Does the design detect the Tal1-/- specific endothelium? sort neighbourhoods that contain at least 1 KO cell by uncertainty score, in which rank are the endothelium neighbourhoods?

emdann commented 2 years ago

This dataset works for no. 1, but not really for no.2: low effect size and only 2 replicates per KO condition Analysis in https://github.com/emdann/diff2atlas/blob/master/notebooks/mouse_chimeras/mouse_chimeras_design_comparison.ipynb

Possible Solutions:

emdann commented 2 years ago

Address no 2 with Augur implementation combined with DA fold-change