I have quantified the RNA-seq samples by Salmon. 2 groups are wild-type and 4 groups are Dnmt2 knocked-out. I've put all in one dataset for DE analysis. box plot of their normalized counts shows the median of knocked-out samples are zero, and maximum of 20 reads are assigned to each transcript. now to perform DE by DESeq2 I have 2 questions: 1- whether zero values should be deleted before? 2- to do DE what minimum of counts has to be selected?

thanks in advance

We have some minimal filtering code in the DESeq2 vignette you can take a look at. There's no point running a DE method when all the counts are 0 obviously, and you can additionally filter out genes which have very small counts for all samples, because these don't have enough precision for estimation of the LFC. A common rule is, for example a count of 10 or more in at least 3 or more samples. However, it does depend a bit on the dataset, for example UMI deduplicated data has counts < 10 which nevertheless give some precision to estimating the LFCs.

It is very suspicious that the maximum transcript count for a sample is 20. It is typically in the 100,000+ range for standard bulk RNA-seq of human or mouse. That there are many zeros is typical and expected, because cell types or tissues express only a subset of the transcripts in the genome.