Should zero count genes filtering by optional

[jsleight1]

There are several pre-filtering approaches that can be used in RNAseq count data analysis:

• Filter genes with zero counts across all samples
• Filter genes that have sum(reads) < threshold across all samples
• Filter genes that have CPM < threshold across given number of samples
• More complex approaches such as those outlined by the function filterByExpr in the edgeR R package (https://f1000research.com/articles/5-1438)

I have noticed that the between sample normalisation methods in rnanorm automatically remove genes with zero counts across all samples. However would it be possible for this to be optionally, allowing the user to perform custom count filtering prior to passing to rnanorm's TMM normalisation for example.

[ftucos]

The filtering of genes with zero counts across all samples already happens inside edgeR even when you do not call filterByExpr() and rnanorm simply replicates that behaviour.

edgeR source: calcNormFactors.R

---

I don't see a context where disabling this step would help.

In a typical RNA-seq experiment you start with ~60 000 annotated genes, roughly two-thirds of them being pseudogenes, inferred genes, or non-coding genes. Depending on library preparation method and sequencing depth, more than half of the genes often ends up having zero counts in every sample. If you keep these all-zero rows:

• Upper Quartile normalisation would turn into something closer to median normalisation.

• TMM normalization: only genes that are non-zero in both the reference and target samples are used to compute log-ratios, so all-zero rows are ignored anyway. The TMMwsp variant can recycle only genes that are non-zero in at least one of the two samples.

• CPM normalization: zeros don't contribute to library size sums.

---

filterByExpr() main purpose is not improving normalisation, it removes low-expression genes that can inflate false-positive results in downstream differential expression analysis. That's also why that filter depends on experimental group structure.
Also consider that rnanorm is designed primarily for count normalisation to feed into machine-learning workflows, not for DE analysis.
For this kind of workflow you could just write manually a filter for low expressed genes (before or after normalization) or look into scikit learn feature selection module.
If you need to perform DE analysis in python you should give have a look at diffxpy, edgePy, or PyDESeq2 packages.