The choice of size factor is confounded with disease stage in human liver transcriptomes: code, protocol, survey records and results

Code, protocol with fifteen dated amendments (a to o), literature survey records with a blind second coding, result files, figures and machine verifier for the manuscript of the same name. Version 2 adds the analyses written after two independent reviews (amendments n and o) and the protocol version 1.0 that version 1 of this record lacked. The protocol and its amendments are dated internal documents; they were not deposited in a public registry before the analysis.

In five public bulk liver RNA sequencing cohorts of fatty liver disease, 610 biopsies in the primary analyses, the log ratio of the median-of-ratios size factor to library size changed with disease stage in four cohorts: it rose by 0.025 to 0.058 per standard deviation of stage in three and fell by 0.064 in the fourth, where mitochondrial reads rose with stage. Of genes significant at a false discovery rate of 5 percent under either factor, 40 to 68 percent were significant under only one of library size and median of ratios, against 6 to 25 percent between two robust factors (TMM and upper quartile). The share of genes appearing to rise with stage under library-size scaling ranged from 22 to 79 percent, with a wide permutation null. Neither size factor is treated as ground truth; the ratio between them is offered as a diagnostic of when the choice matters. About half of 25 sampled reanalyses of these cohorts did not state their normalization.

Input data are public and are not redistributed; the accessions and files required are listed in README.md. The cell-composition arm failed its own validity checks and is deposited in full, because a failed control is a result.