search

AzSaied / Az_MAGMA_Benchmarking

This is a repo for my MAGMA benchmarking project with Imperial
1 stars 0 forks source link

Choose an outcome measure to convert truth matrix into a single number #4

Open AzSaied opened 8 months ago

AzSaied commented 8 months ago

The most crude level - we can choose GWAS phenotypes which logically ought to be associated with only one cell type.

AzSaied commented 8 months ago

In reality - all genesets associated with a given phenotype, are likely to have some correlation with other cell types. A certain amount of 'signal' ought to be expected with other cell types. This would not be a measurement error - but would be biologically 'true'.

AzSaied commented 8 months ago

At the most crude level, we could try to optimise the MAGMA_celltyping to output zero for all celltypes other than the one we logically expect...

But it would be so much better if we could predict where we might expect positive signal, and then optimise the MAGMA celltyping to produce a signal that matches our expectation.

AzSaied commented 8 months ago

This might be a little naive, but could we look at each GWAS that we include in our matrix, look at all of the SNPs that are associated with positive but non statistically significant p values (no MAGMA) but then map these SNPs to genes and then use EWCE to see if this 'second tier' geneset is associated with any particular celltypes?