Bayesian inference of the gene expression states of single cells from scRNA-seq data

In spite of a large investment in the development of methodologies for analysis of single-cell RNA-seq data, there is still little agreement on how to best normalize such data, i.e. how to quantify gene expression states of single cells from such data. Starting from a few basic requirements such as that inferred expression states should correct for both intrinsic biological fluctuations and measurement noise, and that changes in expression state should be measured in terms of fold-changes rather than changes in absolute levels, we here derive a unique Bayesian procedure for normalizing single-cell RNA-seq data from first principles. Our implementation of this normalization procedure, called Sanity (SAmpling Noise corrected Inference of Transcription activitY), estimates log expression values and associated errors bars directly from raw UMI counts without any tunable parameters. Comparison of Sanity with other recent normalization methods on a selection of scRNA-seq datasets shows that Sanity outperforms other methods on basic downstream processing tasks such as clustering cells into subtypes and identification of differentially expressed genes. More importantly, we show that all other normalization methods present severely distorted pictures of the data. By failing to account for biological and technical Poisson noise, many methods systematically predict the lowest expressed genes to be most variable in expression, whereas in reality these genes provide least evidence of true biological variability. In addition, by confounding noise removal with lower-dimensional representation of the data, many methods introduce strong spurious correlations of expression levels with the total UMI count of each cell as well as spurious co-expression of genes.