The rlog transformation is calculated as follows. The experimental design matrix X is substituted with a design matrix with an indicator variable for every sample in addition to an intercept column. A model as described in Equations (1) and (2) is fit with a zero-centered normal prior on the non-intercept terms and using the fitted dispersion values \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha _{\text {tr}}(\bar \mu)$ \end{document}αtr(μ¯), which capture the overall variance-mean dependence of the dataset. The true experimental design matrix X is then only used in estimating the variance-mean trend over all genes. For unsupervised analyses, for instance sample quality assessment, it is desirable that the experimental design has no influence on the transformation, and hence DESeq2 by default ignores the design matrix and re-estimates the dispersions treating all samples as replicates, i.e., it uses blind dispersion estimation. The rlog-transformed values are the fitted values, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $$ \text{rlog}\left(K_{ij}\right) \equiv \log_{2} q_{ij} = \beta_{i0} + \beta_{ij}, $$ \end{document}rlogKij≡log2qij=βi0+βij, where βij is the shrunken LFC on the
