Using LRT and Wald test to filter differentially expressed loci in RNA-seq dataset

Question

[a.rex]

I am using Kallisto and Sleuth for analysis of RNA-seq data.

When doing differentially analysis, I first run a LRT test in Sleuth (which compares the likelihood of the data assuming no differential expression (null model) against the likelihood of the data assuming differential expression (alternative model)). I then run a Wald test (which tests only one model and reports differential expression).

I then find a list of loci that have passed both the LRT and Wald tests (q value<0.05).

However, I fail to find any loci that pass the LRT test. Conversely, their are >2000 loci that pass the Wald's test.

For analysis of RNA-seq data, would it be okay to only look at Wald-passed LRT-failed data?

Conversely, for some experiments, I have a list of loci that pass both LRT and Wald? Are both Wald-passed LRT-passed, and Wald-passed LRT-passed data both equally valid for publication?

LRT is more stringent, I suppose, so would it be the case that if replicates are quite variable, then we will see fewer loci that pass the LRT test?

Answer 1

[Devon Ryan]

You're misunderstanding what you're doing.

The LRT is testing the likelihood of your full model vs. some reduced model. Individual genes with significant p-values according to that are declared differentially expressed.

A Wald test is fitting the full model and then using a coefficient (or more) and its (or their) standard error to perform a test for difference from 0. Genes with a coefficient (or more) significantly different from 0 are differentially expressed and their log2 fold-change is the coefficient.

Anyway, I recall that one assumption of an LRT is that the values should roughly follow a Chi-squared distribution under the null. I suspect that you'll find things to be 0-inflated, which may explain why the Wald test is giving you preferable results.

Either an LRT or Wald test are fine for publications and most people don't even bother to describe which they use.