I have raw gene read counts and would like to perform an analysis across multiple samples. I've found conflicting info online on how this should be done. One commonality however is that FPKM/RPKM should not be used. On the other hand, there seems (to me at least) to be conflict about whether TPM is suitable. For example, this StatQuest video seems to suggest that TPM is suitable for analysis across samples. One person on this Biostar post (among other resources I've looked at) suggested that none of the units (FPKM, RPKM, TPM) are suitable for cross sample comparisons. I am quite confused. Can someone clear up the confusion?
Neither of RPKM, FPKM or TPM is a good choice. All these methods are similar, they only perform the operations in a slightly different order. Eventually they correct for gene length and sequencing depth. Correction for sequencing depth is necessary but not sufficient in many cases. Correction for length is typically not done for a differential analysis as it reduces the counts of many genes and therefore throws away statistical power. I will not describe this in detail since the StatQuest video series you link includes videos about DESeq2 and edgeR normalization procedures which extensively cover the normalization procedures and why these approaches are superior to RPKM/FPKM/TPM. In short: You have to correct for differences in library composition to compare between samples. Please watch these videos and read the manuals of the established RNA-seq analysis packages such as but not exclusively limma, edgeR and DESeq2.