As far as I know, there isn't a universally agreed-upon threshold or an approach to clean the data. I want to remove the genes that don't contribute, or in other words, the noise genes, BEFORE I normalize the data, using CPM or TPM or any other approach.
I've picked the threshold randomly, I tried not to set it too high so that I dont delete important genes that might have infomative value. This is my code:
thresh = data > 0.5
keep = rowSums(thresh) >= 1.5
data = data[keep,]
What do you think? thanks!
•
written
by
AlexStar •
3 answers
For the sake of simplicity and because it well stood the test of time I always use use edgeR::filterByExpr() which has reasonable defaults. In general I would try to respect sample size. If you have 100 samples then a filter like "1 CPM in at least 3 samples" makes little sense. Rather "in at least 10 or 20 samples" would be fine.
Honestly, I am a fan of visualizing to select a threshold since these can depend on the specific experiment. I would even just run standard DEG analysis with removing only 0-count genes and visualizing data to see which FPKM threshold removes the noisy genes. I am far from an expert on this though. Otherwise, I'd echo others, use something standard or reasonable and document.
Log in to answer this question.
-
Filtering genes after TPM normalization
written by AlexStar •
-
Is it adviced to remove genes with very low expression in TPM data?
written by AlexStar •
-
EdgeR: apply minimum cpm threshold only to certain samples
written by 826afca9 •
-
Comparing gene expression between two groups, using boxplots
written by AlexStar •
-
A check over my code using Limma DE, starting from normalized data
written by AlexStar •
-
limma - voom with TPM data
written by AlexStar •
-
How to get gene lengths ?
written by AlexStar •
-
Filtering lowly expressed genes
written by Barista •
-
Problem with low count removal in edgeR ?
written by Björn •
-
count based or cpm based filtering for miRNA normalization when library size is small.
written by Björn •
