Test PAC design on COVID datasets

[emdann]

Experiment design

• PAC design: reference trained on healthy PBMCs (excluding Stephenson), map Stephenson (COVID data as perturbation, healthy data as ctrl)
• PA design: as above but ignore the healthy controls
• PC design: just stephenson data

For all three: with and without including query in the reference model

Ground truth:

• IFN expressing cells
• Signature of T cell activation detected in challenge study by Rik
• Use CellTypist model trained by Lisa on Yoshida dataset, including labelling of IFN stimulated cells (/nfs/team205/lisa_dinithi/PBMC_yoshida_celltypist_model.pkl)

Expanding existing analysis Examine changes along continuum of activation trajectory for IFN stimulation and TNF alpha signatures: with nhood level analysis we can look at which stages of infection does the transition from quiet to activated happen.

[emdann]

Results so far:

• PAC design improves detection of IFN stimulated cells
• reference-free PAC design is even better than de novo integration, with and without the atlas

To do/expand:

• Comparison to other uncertainty metrics