Systematic measurement biases make data normalization an essential preprocessing step in single-cell RNA sequencing (scRNA-seq) analysis. There may be multiple, competing considerations behind the assessment of normalization performance, some of them study-specific. Because normalization can have a large impact on downstream results (e.g., clustering and differential expression), it is critically important that practitioners assess the performance of competing methods.We have developed scone -a flexible framework for assessing normalization performance based on a comprehensive panel of data-driven metrics. Through graphical summaries and quantitative reports, scone summarizes performance trade-offs and ranks large numbers of normalization methods by aggregate panel performance. The method is implemented in the open-source Bioconductor R software package scone. We demonstrate the effectiveness of scone on a collection of scRNA-seq datasets, generated with different protocols, including Fluidigm C1 and 10x platforms. We show that top-performing normalization methods lead to better agreement with independent validation data. * These authors contributed equally. † These authors contributed equally.scRNA-seq data such as: zero inflation, i.e., an artifactual excess of zero read counts observed in some single-cell protocols (e.g., SMART-seq) [3,4]; transcriptome-wide nuisance effects (e.g., batch), comparable in magnitude to the biological effects of interest [5]; uneven sample quality, e.g., in terms of alignment rates and nucleotide composition [6]. In particular, widely-used global-scaling methods, such as reads per million (RPM) [7], trimmed mean of M values (TMM) [8], and DESeq [9], are not well suited to handle large or complex batch effects and may be biased by low counts and zero inflation [2]. Other more flexible methods, such as remove unwanted variation (RUV) [10,11] and surrogate variable analysis (SVA) [12,13], depend on tuning parameters (e.g., the number of unknown factors of unwanted variation).A handful of normalization methods specifically designed for scRNA-seq data have been proposed. These include scaling methods [14,15], regression-based methods for known nuisance factors [16,17], and methods that rely on spike-in sequences from the External RNA Controls Consortium (ERCC) [18,19]. While these methods address some of the problems affecting bulk normalization methods, each suffers from limitations with respect to their applicability across diverse study designs and experimental protocols. Global-scaling methods define a single normalization factor per cell and thus are unable to account for complex batch effects. Explicit regression on known nuisance factors (e.g., batch, number of reads in a library) may miss unknown, yet unwanted variation, which may still confound the data [11]. Unsupervised normalization methods that regress gene expression measures on unknown unwanted factors may perform poorly with default parameters (e.g., number of factors adjusted for) and require tuning, while ERCC-based m...