SCelVis: exploratory single cell data analysis on the desktop and in the cloud

Introduction

Single-cell omics technologies, in particular single-cell RNA sequencing (scRNA-seq), allow for the high-throughput profiling of gene expression in thousands to millions of cells with unprecedented resolution. Recent large-scale efforts are underway to catalogue and describe all human cell types (Regev et al., 2017) and to study cells and tissues in health and disease (https://lifetime-fetflagship.eu). Single-cell sequencing could therefore become a routine tool in the clinic for comprehensive assessments of molecular and physiological alterations in diseased organs as well as systemic responses, e.g., of the immune system. The enormous scale and high-dimensional nature of the resulting data presents an ongoing challenge for computational analysis (Stegle, Teichmann & Marioni, 2015). Ever more sophisticated methods, e.g., deep learning frameworks (Eraslan et al., 2019), extract multiple layers of information from cell types to lineages and differentiation programs. Many of these methods, their mathematical background, and the underlying assumptions will remain opaque to users without specific bioinformatics expertise. At the same time, an in-depth understanding of the relevant biology is often beyond the know-how of typical bioinformatics researchers. More than ever, single-cell omics requires close communication and collaboration from wet and dry lab experts. Due to the large amount of data, communication needs to be based on interactive channels (e.g., web-based apps) rather than static tables. Further, as single-cell omics moves towards the clinic, FAIR (Wilkinson et al., 2016) data management, data privacy, and data security issues need to be handled appropriately. All employed methods should be able to scale towards handling a large number of users and even larger numbers of samples.

Materials & Methods

SCelVis is based on Dash by Plotly (Plotly Technologies Inc., 2015) and accepts data in HDF5 format as AnnData objects. These objects can be created using Scanpy (Wolf, Angerer & Theis, 2018), provide a scalable and memory-efficient data format for scRNA-seq data and integrate naturally into python environments. SCelVis also provides conversion functionality to AnnData from raw text, loom format or 10X Genomics CellRanger output. The built-in converter is accessible from the command line and a web-based user interface (Fig. 1). One HDF5 file or a folder containing multiple such files can then be provided to SCelVis for visualization, and data sets can be selected for exploration on the graphical web interface. To enable both local and cloud access, data can be read from the file system or remote data sources via the standard internet protocols FTP, SFTP, and HTTP(S). SCelVis also provides data access through the open source iRODS protocol (Rajasekar et al., 2010) or the widely-used Amazon S3 object storage protocol. The data sources can be given on the command line and as environment variables as is best practice for cloud deployments (Adam Wiggins, 2011). The latter allows for easy “serverless” and cloud deployments.

SCelVis is built around two viewpoints on single-cell data (Fig. 1). On the one hand, it provides a cell-based view, where users can browse and investigate cell annotations (e.g., cell type) and cell-specific statistics (sequencing depth, cell type proportions, etc.) in multiple visualizations, e.g., on a t-SNE or UMAP embedding, as violin or box plots or bar charts. Cells to be displayed can be filtered by various criteria, and groups of cells can be defined manually on a scatter plot as input for on-the-fly differential gene expression analysis. On the other hand, SCelVis provides a gene-based view that lets users explore gene expression in multiple visualizations on embeddings or as violin or dot plots. Relevant genes can be specified by hand or selected directly from lists of marker or differential genes.

The source code is available under the permissive MIT license on the GitHub repository at https://github.com/bihealth/scelvis, which also contains a tutorial movie and a link to a public demonstration instance. The software can be run both in the cloud and on workstation desktops via Docker. Documentation and tutorials are provided on https://scelvis.readthedocs.io.

Usage example

We provide three example datasets within our GitHub repository or via figshare. First, a small synthetic simulated dataset created for testing and illustration purposes, and secondly a publicly available processed scRNA-seq dataset from 10X Genomics containing ∼1,000 cells of a mix of human HEK293T and murine NIH3T3 cells. Finally, we reanalyzed a published data set of stimulated and control peripheral blood mononuclear cells (PBMCs; Kang et al., 2018) with the Seurat “data integration” workflow (Stuart et al., 2019) and made it accessible via https on figshare (https://files.figshare.com/18037739/pbmc.h5ad; Fig. 2). With the species-mix dataset from 10X, the relevant plot to demonstrate a low doublet rate can be readily re-created (Fig. 2A left; compare to Fig. 2A in Zheng et al. (2017), which shows data obtained with a previous version of the 10X chemistry), and the species composition of the different clusters found by CellRanger can be easily interrogated (Fig. 2A right). For the PBMC dataset, it is straightforward to perform differential gene expression analysis, e.g., between stimulated and control monocytes by using the “filter” and “differential gene expression” buttons (Fig. 2B). Summarized gene expression for cell-type marker genes as well as for general (e.g., IFI6) or cell-type specific (e.g., CXCL10) differential genes can be displayed in a split dot plot as in Fig. 2D of Stuart et al. (2019). Hence, our visualizations for the published datasets are equivalent to those obtained from other visualization tools, e.g., Seurat.

Conclusions

In this manuscript, we have presented SCelVis, a method for the interactive visualization of single-cell RNA-seq data. It provides easy-to-use yet flexible means of scRNA-seq data exploration for researchers without computational background. SCelVis takes processed data, e.g., provided by CellRanger or a bioinformatics collaboration partner, as input, and focuses solely on visualization and explorative analysis. Great care has been taken to make the method flexible in usage and deployment. It can be used both on a researcher’s desktop with minimal training yet its usage scales up to a cloud deployment. Data can be read from local file systems but also from a variety of remote data sources, e.g., via the widely deployed (S)FTP, S3, and HTTP(S) protocols. This allows for deploying it in a Docker container on “serverless” cloud systems. As both the application and data can be hosted on the network or cloud systems, the application facilitates cross-institutional research. For example, a sequencing or bioinformatics core unit can use it for giving access to non-computational collaboration partners over the internet. This is particularly relevant as it comes with no dependency on any vendor-specific technology such as the Google or Facebook authentication that appears to become pervasive in today’s life science.