https://liorpachter.wordpress.com/2014/04/30/estimating-number-of-transcripts-from-rna-seq-measurements-and-why-i-believe-in-paywall/

^ See this blog post which references some literature about FPKM cutoffs.

For some things in biological sciences, cut-offs actually make sense, for example, there is statistical backing via Power analysis to having an expression study of at least 3 samples - such a study is at the absolute fringes of statistical power, i.e., the ultimate limit at which one can reasonably detect a statistically significantly differentially expressed gene.

For NGS, the read depth cut-off of 30 also has backing via simulation and validation studies. A read depth of 30 via NGS can confer 99.999% sensitivity for NGS with Sanger seq as the gold standard. One doesn't need DeepVariant or any other fancy nonsense to do this - SAMtools / BCFtools mpileup suffices, for single nucleotide variants at least.

For FPKM, a useful start may be to read through the zFPKM package docs. My colleague John Thompson --together with Ron Ammar-- designed a cut-off for gene-level FPKM data via conversion to Z scores. Take a look: https://www.bioconductor.org/packages/release/bioc/html/zFPKM.html

Thank you and kind regards,

Kevin