The necessarity of between-sample normalization?

[SamCurt]

I'm currently perform some analyses involving cross-project expression data. Because it involves linear equations, we have decided to take log(TPM) as that algorithm's input.

While using older workflows involving rsubread or htseq-count would always require us to perform between-sample normalization, newer transcript quantification tools such as RSEM, Kalisto and Salmon gives out reads and (at a minimum) TPM as their raw output.

But even in that case, should I take out the reads, normalize it with DESeq2/edgeR, and calculate the TPMs instead? I'm not particularly comfortable with not doing between-sample normalizations, but I have a feeling that it's the norm these days.

[cmbetts]

Somebody more knowledgeable may correct me, but the statistical methods used by DESeq2 rely on having the raw read counts to calculate power and significance, and therefore can't use normalized values like TPM (There's a huge statistical difference in seeing one count in a million reads vs a thousand counts in a billion that's lost in normalization)

[SamCurt]

Somebody more knowledgeable may correct me, but the statistical methods used by DESeq2 rely on having the raw read counts to calculate power and significance, and therefore can't use normalized values like TPM (There's a huge statistical difference in seeing one count in a million reads vs a thousand counts in a billion that's lost in normalization)

Of course I know things like DESeq2 can't take TPM. The options I mentioned above are:

1. Take the TPM from the quantifier directly to downstream.
2. Take the expected read count from the quantifier, normalized by DESeq2 (etc), and then calculate TPM from this normalized number.