Kallisto/Sleuth DTE analysis - LRT test results in more DTE compared to Wald test.

Question

[MK]

Hello everyone, I am trying to get differential transcript/gene expression using sleuth on my dataset which consist of 3 replicates of well watered and not watered plants. I am using kallisto/sleuth pipeline. I realized I am getting more significant hits (qval <= 0.05) after running LRT test compared to Wald test. Isn't it supposed to be opposite? I read that LRT is a more stringent test. There is a difference of over 5000 transcripts between them!

so <- sleuth_prep(s2c, target_mapping = t2g, ~ condition, extra_bootstrap_summary = TRUE)
so <- sleuth_fit(so, ~condition, 'full')
so <- sleuth_fit(so, ~1, 'reduced')
so <- sleuth_lrt(so, 'reduced', 'full')
so <- sleuth_wt(so,'conditionwater','full')
results_table_lrt <- sleuth_results(so, 'reduced:full', 'lrt', show_all = FALSE)
results_lrt_significant <- results_table_lrt$target_id[which(results_table_lrt$qval <= 0.05)]
results_table_wald <- sleuth_results(so, 'conditionwater','full', test_type = 'wt')
results_wald_significant <- results_table_wald$target_id[which(results_table_wald$qval <= 0.05)]

table(results_table_lrt$qval<=0.05)
FALSE TRUE
53234 32237
table(results_table_wald$qval<=0.05)
FALSE TRUE
57549 27922

I would greatly appreciate any help.

[Devon Ryan]

How many levels does condition have?

[MK]

This is my metadata file.

sample  condition 
D4_1    drought
D4_2    drought
D4_3    drought
W4_1    water
W4_2    water
W4_3    water

[Devon Ryan]

I vaguely remember the original sleuth publication recommending LRT over Wald for some reason. If my recollection is correct then whatever the reason behind that was is likely the cause. I think sleuth has an email list, so you might ask on there.

[MK]

I think you are right. The reason why I want to run it together with Wald test is to be able to get beta value. I don't know how to get any estimation of fold change while running only LRT test on its own.

[Nicolas Rosewick]

[comment not directly related to the question] : ~30K transcript with qval <= 0.05 seems quiet a lot. Is this expected for watered/non-watered plants ?

[MK]

I might be wrong, but as wheat has a huge genome and drought stress will probably affect a lot of pathways this should be fine. Nevertheless, I am not experienced in working with wheat and this is my first RNA-seq analysis.