Hi everyone, I'm relatively new to this type of RNA-seq analysis, so apologies if these are basic questions. I've been reading the RUVSeq vignette and related discussions, but I'm still unsure about some aspects of the recommended workflow when combining RUVg with DESeq2.

I'm using RUVSeq's RUVg() with in silico empirical control genes (following Section 2.4 of the RUVSeq vignette), where the control genes are obtained from a first-pass DESeq2 analysis. I have two related questions.

First-pass design formula

Besides my variable of interest (group), I also have known technical/biological covariates (e.g. batch, RIN).

For the first-pass DESeq2 analysis used to identify empirical control genes (for example, genes with p > 0.3), should I fit:

(a) ~ group only, allowing RUVg() to capture all unwanted variation (including variation associated with the known covariates), or

(b) ~ covariates + group, so that the empirical control genes are those that remain non-significant for group after adjusting for the known covariates?

My intuition is that option (a) could misclassify truly group-associated genes as controls if their signal is masked by unmodeled technical variation. On the other hand, I wonder whether option (b) introduces some degree of circularity, especially if the measured covariates are themselves noisy or imperfect proxies for the unwanted variation.

Collinearity between estimated W factors and known covariates

Suppose I run RUVg() and find that one or more estimated W factors are strongly correlated with a known covariate such as Batch.

In the final DESeq2 model, is it generally recommended to include both, for example:

~ Batch + W_1 + ... + W_k + group

or could this lead to multicollinearity, inflated standard errors, unstable Wald statistics, or even rank-deficient design matrices?

Would the recommended approach be to:

• examine the correlation (or R²) between each W factor and the known covariates before fitting the final model;
• keep only the measured covariate if it is well-characterized and trusted;
• keep only the W factors if they appear to capture the unwanted variation more effectively; or
• include both unless there is evidence of severe collinearity?

Thank you in advance, and apologies if these are beginner questions—I'm still getting familiar with RUV-based normalization and want to make sure I'm applying the workflow correctly.