Perhaps my understanding is wildly misguided, but I'm seriously confused about why people are still interested using FPKM values for cross sample gene expression analysis. My understanding is, that for some gene ${ i }$, its $FPKM$ is calculated as follows:
$$ FPKM_i \propto \frac{ \mathrm{fragments\ mapping \ to\ gene \ }i } { \sum_j \mathrm{ fragments\ mapping\ to \ gene\ } j } \bigg/ \mathrm{effective\_length}(i) $$
For comparing gene expression within a single sample, this doesn't seem too harmful to me. In fact, this seems like a good idea.
However, the use of $FPKM$ or even the famous $log_2(FPKM+1)$ in a similarity analysis seems to me plain wrong. True differences in RNA abundances of a couple of genes would significantly impact the FPKM values of the other genes leaving perhaps confused about the true differences between the samples.
Am I missing something here? Is it ever meaningful to use FPKM values for analyzing across samples?