Organ specific genes: signatures or cell states?

[emdann]

• Analysis à là JP on aggregated profiles per cell-type (organ and samples), but using a model like MOFA or MEFISTO
• If we find an organ specific signature, how do we figure out if it's indicative of higher heterogeneity within that cell type? E.g. let's say we find specific expression of a Tfh specific marker in spleen T cells. That could be because ALL the spleen T cells are Tfh or because a subset of spleen T cells are Tfh, with very preposterous marker expression. To tease this out we could do factorization within a celltype in one organ, or on the residuals between the 'average' cell type profile: this would allow us to distinguish if a specific subset of cells deviates from the avg profile, or all the cells.

[emdann]

Suggestions from John:

• checking residuals of single-cells on linear model: we expect that if a signature is common to all cells of one type in one organs, the residuals should be balanced, not so if a few cells express that signature gene
• Use GlmGamPoi for reliability and speed https://github.com/const-ae/glmGamPoi

[emdann]

Control genes: in mature B and B1 we expect AICD, HGHI and AICDA to be thymus specific

[emdann]

Problems with JPs approach:

• collinearity between celltypes and organs
• incorporating all the technical covariates (donor, method, n_genes) in DE testing

Solution: using a (generalised) linear mixed model to avoid (1) running out the degrees of freedom by introducing too many fixed parameters (i.e. the many covariates), and (2) collinearity between covariates including organ and celltype. We use the implementation by Natsuhiko introduced here. Full model description from the supplementary notes is here (provided by Natsuhiko).

[emdann]

Check gamma-delta signature in NKT cluster