Hi everybody,
I am analysing RNA-Seq data of patient samples. One group of samples has a certain monoallelic deletion, the other group 2 copies of the wildtype gene.
Now, I want to compare the read counts between the groups to see if the RNA expression of certain target genes changes as a result of the deletion.
For example, I would expect a reduction of the read counts of the gene of interest to 50% in the patients with the monoallelic deletion in comparison to those with two wildtype copies.
But how should I approach this analysis? Can I just perform a student's t-test or do I need to use DeSeq2? Would really appreciate your comments.
Thanks a lot in advance!
Theoretically, the negative binomial model is more appropriate to analyze count variables than the normal model. However, for large read counts there is no practical difference (if log counts are used).
If the distributions of the data (group 1, group 2, group 1+ group 2) are normally distributed and do not violate the assumptions of all general linear models (t-test is one of them), then use t-test. The Kolmogorov-Smirnov (KS) test can be used to determine normality and uniformity of the distributions (or not). If any of the distributions are not normal (i.e. fail the KS tests for normality), then use the Mann-Whitney U test to compare the distributions of the two groups. The problem is that in order for the t-test results to be valid and publishable, the data must meet the assumptions for parametric models. The Mann-Whitney U on the other hand is a non-parametric test, and is considered to be equivalent to a t-test for data that is non-normally distributed and therefore disqualified from being analyzed in a statistically valid manner with a t-test. If your goal is to compare group 1 to group 2 and determine if there is any statistically significant difference (higher or lower) between the location (mean, median) of the distributions in the two groups, these tests (t-test, MWU) will achieve that goal. If your goal is to publish valid results at each step in the method, pay attention to the statistical assumptions of the models used, and use the appropriate statistical method, given the descriptive statistics of the data distribution being analyzed (centrality, variation, dispersion = normality or not).
I hope this helps. The appropriate test is always based upon the appropriate and valid method to be used given the nature of the data analyzed.
Following the two previous answers, i think would be easier to stick with DeSeq2 as it is aware of the statistical characteristics of RNA-seq data while also giving you options to check, visualize and normalize your data.
- Badges
- Science topic