I have the following two genes with expression taken from single-cell
RNAseq:
I am conducting pairwise comparison between sample (S1, S2...) on those data,
using t.test. The data has been normalized using Monocle3 normalization function.
My question is whether this is a correct approach? If not what are the better statistical test to use?
And also, if it's also acceptable to perform statistical test for raw count?
And the need of multiple comparison test like Bonferroni, FDR?
Yes, the t-test has been shown to perform well, probably due to the large number of data points when treating each single cell as a "replicate", see Soneson & Robinson (2018) Nat Methods. This paper also covers alternative testing regimes and discusses/benchmarks up/downsides. The big advantage of t- or wilcox tests is speed and simplicity. This is (iirc) for single-sample comparisons e.g. comparing sample1 vs sample2 in the absence of true biological replicates. With true replicates I would often prefer pseudobulk-level DE analysis as summing cells per cluster into a pseudobulk will compensate for zeros/dropouts, so (for me personally) the estimated fold changes are more intuitive to interpret. Others might disagree here.
And also, if it's also acceptable to perform statistical test for raw count?
No, raw counts are confounded by sequencing depth.
And the need of multiple comparison test like Bonferroni, FDR?
Yes, FDR correction is needed when conducting multiple tests, there is no magic in single-cell analysis, it is has to obey the basic rules of statistics.
