Question 8 of 20 Fundamentals Concept
Why should DESeq2 or edgeR be given raw counts rather than TPM or FPKM values?
Short answer
These tools model read counts with a negative binomial distribution, and how precise a measurement is depends on how many reads were actually counted. TPM and FPKM have already been rescaled, which hides that information, and the tools do their own library-size normalization.
Why
A count of 10 against 20 reads and a count of 1,000 against 2,000 reads are both a doubling, but the first is far noisier. A count model knows this because it sees the raw numbers. Once values are scaled to TPM or FPKM, a gene with 10 reads and a gene with 10,000 reads can carry the same value, and the model can no longer judge precision.
The DESeq2 documentation says plainly that counts scaled by library size should not be used as input, because only the count values let the model assess measurement precision correctly. TPM has a second problem. It is a within-sample proportion, so a few very highly expressed genes changing in one condition shifts the TPM of every other gene, even if those genes did not change.
That makes TPM a poor basis for comparing samples. When reads are quantified with Salmon or kallisto, the right route is tximport, which passes estimated counts and an average transcript length offset to the statistical tool. TPM still has a use: ranking genes within one sample and making plots that readers find intuitive.
A strong answer covers
- Count models use the raw number of reads to estimate precision
- DESeq2 and edgeR normalize for library size internally
- TPM is a within-sample proportion and is sensitive to composition
- Transcript quantifier output goes in through tximport as counts plus length offsets
Common mistakes
- Rounding TPM values and feeding them to DESeq2 as if they were counts
- Normalizing counts first and then normalizing again inside the tool
- Treating TPM as comparable across samples sequenced with different protocols
The interviewer might ask next
What does tximport pass to DESeq2 that a plain count matrix does not contain?