Score from correlations per genes in DGE results

Question

[Lopiniatre]

Hi, I am looking for a scoring method. I have on one side a correlation score (-1,1 range) per gene (>22,000) for different fates (A, B, C) giving me a link between each gene and each fate. On the other side I have the result of several differential expressions (so log2FC values (-3 to +3 range) and their p-value) per gene.

Which statistics/formula could I use to obtain a score per fate for each differential expression result. The goal is to understand if the differential gene expression leads to an increase or a decrease of the fate according to the correlation (integrating both positive and negative impact into one value over all genes).

I was thinking of multiplying logFC x corr for each genes, summing over all genes to get an overall score. But I don't know what would be the validity of such score?

(sum (logFC * corr) )/ n_genes

Any suggestion to a statistical method/ publication applicable in this case? Thanks!

Answer 1

[rpolicastro]

If you are going to use Log2FCs for a downstream purpose like this, I would use the shrunken Log2FCs from DESeq2, so that less confident FC values are shrunken closer to 0.

I think a good place to start would be a simple parsimonious linear regression using the formula Log2FC ~ corr_score + fate. Do the proper linear regression quality control, and if things look good the slopes should give you an idea whether this is some significant relationship between Log2FC and corr_score, and whether there are some differences based on fate.

The direction of the analysis after this will somewhat depend on how the QC and your model coefficients look.

[Lopiniatre]

Thank you, I will try lm approach.

One question though: Since each gene has 1 logFC value but multiple corr_score (one for each fate A, B, C), shouldn't I run a lm for each fate Log2FC ~ corr_score ? this way I would use the slope and p-value as a "fate alteration score" ? (maybe my description of the data was unclear)

[rpolicastro]

If you only have one Log2FC per multiple fates, you should start off with separate models per fate. If you find linear regression working well you could build a slightly more complex regression such as Log2FC ~ FateA + FateB + .... If you do format your model like this you would probably want to use lasso linear regression to reduce the less important coefficients closer to 0 to get a better idea of which fates are most correlated with Log2FC.