Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression

Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression Christoph Hafemeister Rahul Satija

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression 2019 Genome biology 10.1186/s13059-019-1874-1 hafemeister2019-normalization-and-variance-stabili hafemeister2019-normalization-and-variance-stabili.txt, the project's staged copy be76b92c9173be8657a8ccdb05ba1402d24e0d06 cc-by biology

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them normalisation of single-cell RNA-seq counts - how to remove the effect of sequencing depth without removing the biology along with it Quotation withheld because this record is not cleared for redistribution. whether any single per-cell scaling factor can normalise genes of different abundance, and if not, what error model should replace it Quotation withheld because this record is not cleared for redistribution. regularized negative binomial regression - fit a GLM per gene with depth as covariate, pool parameters across genes of similar abundance, and take the Pearson residuals as normalised values Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. a preprocessing choice that decides what counts as biological signal - the same data yielding heterogeneity or not depending on how the error model is constrained Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. counts of low- and high-abundance genes respond differently to sequencing depth, so one factor per cell cannot correct both; dividing by depth leaves the high-abundance tiers still confounded and the variance contribution of the other tiers dampened abundance-dependent depth effect defeating a single scaling factor Quotation withheld because this record is not cleared for redistribution. an unconstrained per-gene model has enough freedom to explain real cell-type differences as technical effects of depth, so the residuals it produces have had the biology subtracted along with the noise - even cell-type markers come out with low residual variance overfitting absorbing biological signal Quotation withheld because this record is not cleared for redistribution. estimating each gene's parameters independently makes them a function of that gene's sampling noise; smoothing them against mean abundance replaces the noise with the population trend, and the estimates become reproducible across bootstraps pooling across genes as a constraint Quotation withheld because this record is not cleared for redistribution. megakaryocytes carry little RNA, so they appear at low sequencing depth for reasons of biology rather than sampling; an unconstrained fit reads that as a negative slope between depth and the marker's count, which cannot be true of any single cell type low RNA content masquerading as a depth relation Quotation withheld because this record is not cleared for redistribution. the regularisation assumes the majority of genes do not vary biologically, which the authors say may be overly simplistic on a highly heterogeneous sample Quotation withheld because this record is not cleared for redistribution. the kernel bandwidth is a user-adjustable parameter; robust over an order of magnitude and producing over- or under-smoothing at the extremes Quotation withheld because this record is not cleared for redistribution. the batch extension requires all cell types to share a similar batch effect, which is an assumption about the experiment rather than a property of the model Quotation withheld because this record is not cleared for redistribution. Pearson residuals are an intermediate rather than a model - an alternative would fit the count distribution directly, as GLM-PCA does, instead of passing residuals to methods that assume normal errors Quotation withheld because this record is not cleared for redistribution. parameter regularisation by pooling across similar features, carried from bulk RNA-seq differential-expression error models into single-cell normalisation Quotation withheld because this record is not cleared for redistribution. generalized linear models and Pearson residuals, carried from regression theory to serve as the normalised data values themselves rather than as a diagnostic Quotation withheld because this record is not cleared for redistribution. a dataset with NO biological variation by construction - droplets that are technical replicates of one bulk RNA sample - run through the unregularised model to see whether the parameter heterogeneity persists when biology cannot explain it eliminates: true biological variation as the explanation of the parameter heterogeneity across genes - in a sample where every droplet is a technical replicate of one bulk RNA extract there is no biology left that could produce it · outcome: failed Quotation withheld because this record is not cleared for redistribution. the same control dataset run through the FINISHED method, where a correct model must find no high-variance genes at all eliminates: residual technical variation as the source of the high-variance genes the finished method selects - run on a sample containing no biology, a correct model must find none, and does · outcome: held Quotation withheld because this record is not cleared for redistribution. bootstrapping the fit over randomised subsets of cells, to separate real between-gene differences from sampling variation in the estimates eliminates: real between-gene differences as the explanation of the parameter spread - estimates that move when the cells are resampled are measuring the sample and not the gene · outcome: failed Quotation withheld because this record is not cleared for redistribution. substituting a Poisson and an unconstrained NB error model into the same pipeline, so the regularisation step is tested rather than assumed eliminates: the generalised linear model framework itself as the source of the result - swapping only the error model shows the work is done by the regularisation, not by using a GLM · outcome: held Quotation withheld because this record is not cleared for redistribution. checking what the high-residual-variance genes actually are, rather than reporting that some exist eliminates: residual technical noise as the explanation of the surviving high-variance genes - noise does not produce a list that is exclusively markers of known immune cell subsets · outcome: held Quotation withheld because this record is not cleared for redistribution. repeating the whole analysis on five UMI datasets from plate- and droplet-based protocols rather than on the one dataset the method was developed against eliminates: the single development dataset, and the droplet protocol it came from, as the source of the result · outcome: held Quotation withheld because this record is not cleared for redistribution. normalization normalization Quotation withheld because this record is not cleared for redistribution. size factor size factor Quotation withheld because this record is not cleared for redistribution. Pearson residuals Pearson residuals Quotation withheld because this record is not cleared for redistribution. overfitting overfitting Quotation withheld because this record is not cleared for redistribution. regularization regularization Quotation withheld because this record is not cleared for redistribution. Chromium control dataset Chromium control dataset Quotation withheld because this record is not cleared for redistribution. kernel bandwidth kernel bandwidth Quotation withheld because this record is not cleared for redistribution. zero-inflated negative binomial zero-inflated negative binomial Quotation withheld because this record is not cleared for redistribution. Poisson Poisson Quotation withheld because this record is not cleared for redistribution. sequencing depth sequencing depth Quotation withheld because this record is not cleared for redistribution. PPBP PPBP Quotation withheld because this record is not cleared for redistribution. regularized NB regularized NB Quotation withheld because this record is not cleared for redistribution. parameter heterogeneity across genes - READING 1: true biological variation in the distribution of single-cell gene expression Quotation withheld because this record is not cleared for redistribution. parameter heterogeneity across genes - READING 2: irreproducible variation driven by overfitting. Resolved in favour of this one by a control dataset with no biological variation in it, where the heterogeneity persisted Quotation withheld because this record is not cleared for redistribution. "effectively normalized" - UNDER LOG-NORMALIZATION: true only of the bottom three abundance tiers; high-abundance genes remain confounded with sequencing depth Quotation withheld because this record is not cleared for redistribution. "effectively normalized" - UNDER PEARSON RESIDUALS: level and variance independent of sequencing depth across the full range of expression Quotation withheld because this record is not cleared for redistribution. critiques → size-factor normalisation for single-cell data - including recent methods that compute the factor more cleverly, since the problem is the single factor and not its estimation Quotation withheld because this record is not cleared for redistribution. critiques → unconstrained negative binomial and zero-inflated negative binomial error models as used in current statistical and deep-learning methods for scRNA-seq Quotation withheld because this record is not cleared for redistribution. builds-on → regularisation of error-model parameters across genes of similar abundance, standard in bulk RNA-seq differential expression and not previously used for single-cell normalisation Quotation withheld because this record is not cleared for redistribution. qualifies → SCnorm's conclusion that gene groups need separate treatment - reached for non-UMI data by a method that ignores the zero values which dominate droplet-based assays Quotation withheld because this record is not cleared for redistribution. size factor —distinct-from→ normalization size factor normalization differs in: whether one constant per cell can serve genes across the whole range of abundance · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether one constant per cell can serve genes across the whole range of abundance · the relation stated: a single per-cell scaling factor introduces different effects on genes of different average abundance, so it is not normalization for all of them Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span says a single size factor has distinct effects; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. Poisson —distinct-from→ regularized NB Poisson regularized NB differs in: whether the error model can represent overdispersion without absorbing the biology · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the error model can represent overdispersion without absorbing the biology · the relation stated: the single-parameter Poisson cannot represent overdispersion and leaves residual variance above one for highly expressed genes; the unconstrained NB overfits in the other direction and flattens the markers Quotation withheld because this record is not cleared for redistribution. a single per-cell scaling factor cannot normalise genes of different abundance, whatever method computes the factor a single per-cell scaling factor cannot normalise genes of different abundance, whatever method computes the factor Quotation withheld because this record is not cleared for redistribution. log-normalisation effectively normalises only the bottom three abundance tiers; high-abundance genes stay confounded with depth log-normalisation effectively normalises only the bottom three abundance tiers; high-abundance genes stay confounded with depth Quotation withheld because this record is not cleared for redistribution. an unconstrained negative binomial model overfits and removes biological variation along with technical variation an unconstrained negative binomial model overfits and removes biological variation along with technical variation Quotation withheld because this record is not cleared for redistribution. the parameter heterogeneity persists in a dataset containing no biological variation, which is what identifies it as overfitting the parameter heterogeneity persists in a dataset containing no biological variation, which is what identifies it as overfitting Quotation withheld because this record is not cleared for redistribution. after normalisation the genes with high residual variance are exclusively markers of known immune cell subsets after normalisation the genes with high residual variance are exclusively markers of known immune cell subsets Quotation withheld because this record is not cleared for redistribution. both non-regularised models fit a negative depth-count slope for a megakaryocyte marker, which cannot be true of a single cell type both non-regularised models fit a negative depth-count slope for a megakaryocyte marker, which cannot be true of a single cell type Quotation withheld because this record is not cleared for redistribution. removing the influence of technical effects in the underlying molecular counts while preserving true biological variation normalization Quotation withheld because this record is not cleared for redistribution. response residuals divided by the expected standard deviation, effectively a variance-stabilizing transformation putting lowly and highly expressed genes on a common scale Pearson residuals Quotation withheld because this record is not cleared for redistribution. droplets each representing a technical replicate of a bulk RNA sample - no biological variation in it by construction Chromium control dataset Quotation withheld because this record is not cleared for redistribution. the unit residual variance expected for a gene with no biological heterogeneity 1 Quotation withheld because this record is not cleared for redistribution. negative binomial negative binomial Quotation withheld because this record is not cleared for redistribution. kernel bandwidth kernel bandwidth Quotation withheld because this record is not cleared for redistribution. generalized linear model generalized linear model Quotation withheld because this record is not cleared for redistribution. sequencing depth sequencing depth Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1186/s13059-019-1874-1.