AbstractSingle-cell RNA sequencing significantly deepened our insights into complex tissues and latest techniques are capable processing ten-thousands of cells simultaneously. With bigSCale, we provide an analytical framework being scalable to analyze millions of cells, addressing challenges of future large datasets. Unlike previous methods, bigSCale does not constrain data to fit an a priori-defined distribution and instead uses an accurate numerical model of noise. We evaluated the performance of bigSCale using a biological model of aberrant gene expression in patient derived neuronal progenitor cells and simulated datasets, which underlined its speed and accuracy in differential expression analysis. We further applied bigSCale to analyze 1.3 million cells from the mouse developing forebrain. Herein, we identified rare populations, such as Reelin positive Cajal-Retzius neurons, for which we determined a previously not recognized heterogeneity associated to distinct differentiation stages, spatial organization and cellular function. Together, bigSCale presents a perfect solution to address future challenges of large single-cell datasets.Extended AbstractSingle-cell RNA sequencing (scRNAseq) significantly deepened our insights into complex tissues by providing high-resolution phenotypes for individual cells. Recent microfluidic-based methods are scalable to ten-thousands of cells, enabling an unbiased sampling and comprehensive characterization without prior knowledge. Increasing cell numbers, however, generates extremely big datasets, which extends processing time and challenges computing resources. Current scRNAseq analysis tools are not designed to analyze datasets larger than from thousands of cells and often lack sensitivity and specificity to identify marker genes for cell populations or experimental conditions. With bigSCale, we provide an analytical framework for the sensitive detection of population markers and differentially expressed genes, being scalable to analyze millions of single cells. Unlike other methods that use simple or mixture probabilistic models with negative binomial, gamma or Poisson distributions to handle the noise and sparsity of scRNAseq data, bigSCale does not constrain the data to fit an a priori-defined distribution. Instead, bigSCale uses large sample sizes to estimate a highly accurate and comprehensive numerical model of noise and gene expression. The framework further includes modules for differential expression (DE) analysis, cell clustering and population marker identification. Moreover, a directed convolution strategy allows processing of extremely large data sets, while preserving the transcript information from individual cells.We evaluate the performance of bigSCale using a biological model for reduced or elevated gene expression levels. Specifically, we perform scRNAseq of 1,920 patient derived neuronal progenitor cells from Williams-Beuren and 7q11.23 microduplication syndrome patients, harboring a deletion or duplication of 7q11.23, respectively. The affected region contains 28 genes whose transcriptional levels vary in line with their allele frequency. BigSCale detects expression changes with respect to cells from a healthy donor and outperforms other methods for single-cell DE analysis in sensitivity. Simulated data sets, underline the performance of bigSCale in DE analysis as it is faster and more sensitive and specific than other methods. The probabilistic model of cell-distances within bigSCale is further suitable for unsupervised clustering and the identification of cell types and subpopulations. Using bigSCale, we identify all major cell types of the somatosensory cortex and hippocampus analyzing 3,005 cells from adult mouse brains. Remarkably, we increase the number of cell population specific marker genes 4-6-fold compared to the original analysis and, moreover, define markers of higher order cell types. These include CD90 (Thy1), a neuronal surface receptor, potentially suitable for isolating intact neurons from complex brain samples.To test its applicability for large data sets, we apply bigSCale on scRNAseq data from 1.3 million cells derived from the pallium of the mouse developing forebrain (E18, 10x Genomics). Our directed down-sampling strategy accumulates transcript counts from cells with similar transcriptional profiles into index cell transcriptomes, thereby defining cellular clusters with improved resolution. Accordingly, index cell clusters provide a rich resource of marker genes for the main brain cell types and less frequent subpopulations. Our analysis of rare populations includes poorly characterized developmental cell types, such as neuron progenitors from the subventricular zone and neocortical Reelin positive neurons known as Cajal-Retzius (CR) cells. The latter represent a transient population which regulates the laminar formation of the developing neocortex and whose malfunctioning causes major neurodevelopmental disorders like autism or schizophrenia. Most importantly, index cell cluster can be deconvoluted to individual cell level for targeted analysis of populations of interest. Through decomposition of Reelin positive neurons, we determined a previously not recognized heterogeneity among CR cells, which we could associate to distinct differentiation stages as well as spatial and functional differences in the developing mouse brain. Specifically, subtypes of CR cells identified by bigSCale express different compositions of NMDA, AMPA and glycine receptor subunits, pointing to subpopulations with distinct membrane properties. Furthermore, we found Cxcl12, a chemokine secreted by the meninges and regulating the tangential migration of CR cells, to be also expressed in CR cells located in the marginal zone of the neocortex, indicating a self-regulated migration capacity.Together, bigSCale presents a perfect solution for the processing and analysis of scRNAseq data from millions of single cells. Its speed and sensitivity makes it suitable to the address future challenges of large single-cell data sets.
Probabilistic Harmonization and Annotation of Single-cell Transcriptomics Data with Deep Generative Models
