Verify logistic regression assumptions

Question

[AKypR]

Hello,

I have single-cell RNAseq samples of cancer cell populations from either a relapse state or just at the diagnostic stage. So there are two states and I would like to characterize the relapse state according to the expression of certain genes. I have previously identified different genes that might be characteristic of a relapse state.

To try to verify this, I plan to apply a logistic regression with the target variable being the state (relapse or diagnostic) and the predictor variables being the expression of the selected genes (about 200 genes). However, I have difficulties in verifying the different assumptions that allow the application of a logistic model, especially concerning the absence of multicollinearity and the presence of a linear relationship between the logit function of the target variable and the predictors.

Are there any particular packages/functions that allow to verify these assumptions?

Answer 1

[Ming Tommy Tang]

Logistic regression is a sensible way to try. Many genes are correlated as they may be in the same pathway. You can remove the highly correlated genes, or use a method such as PLSR which is fine with multicollinearity.

Read my blog post https://divingintogeneticsandgenomics.rbind.io/post/partial-least-square-regression-for-marker-gene-identification-in-scrnaseq-data/

[AKypR]

Hey Ming Tang,

Thanks for your answer.

I'm sorry for the late reply but I wanted to read the article you sent me before replying.

So I read it and I also read about PLS then I have two questions that come to me that you might have an answer to:

• Since you are working on a scRNAseq sample, i assume that there is a lot of sparsity. So why did not you use a (s)PLS-DA instead of a classical PLS-DA?
• I saw that when we work with PLS method, we must verify the linearity of the data. I am not sure that you verify this part in your article, have you verified it on your side and not mentioned it into the article.

PS: Great blog by the way!