Alevin-fry unlocks rapid, accurate and memory-frugal quantification of single-cell RNA-seq data

Abstract

The rapid growth of high-throughput single-cell and single-nucleus RNA-sequencing (scRNA-seq and snRNA-seq) technologies has produced a wealth of data over the past few years. The size, volume and distinctive characteristics of these data necessitate the development of new computational methods to accurately and efficiently quantify sc/snRNA-seq data into count matrices that constitute the input to downstream analyses. We introduce the alevin-fry framework for quantifying sc/snRNA-seq data. In addition to being faster and more memory frugal than other accurate quantification approaches, alevin-fry ameliorates the memory scalability and false-positive expression issues that are exhibited by other lightweight tools. We demonstrate how alevin-fry can be effectively used to quantify sc/snRNA-seq data, and also how the spliced and unspliced molecule quantification required as input for RNA velocity analyses can be seamlessly extracted from the same preprocessed data used to generate normal gene expression count matrices.

Fig. 1 – 2-column figure.. Overview of the alevin-fry pipeline

(operating in unspliced, spliced, ambiguous (USA) quantification mode). The arrows highlight the flow of data through the pipeline, whose output is a matrix specifying the expected counts of each of the considered splicing states of each gene within each quantified cell.

Fig. 2 – 2-column, 5-panel figure.. Comprehensive analysis of the performance pf alevin-fry on real and simulated datasets.

(a) The frequency distribution of the presence of genes across all shared cells for STARsolo, kallisto|bustools, and alevin-fry (including multiple index types for alevin-fry) on the simulated data. Different color lines represent the quantification methods. Within the variants of alevin-fry, txome stands for transcriptome reference (i.e., just indexing the annotated, spliced, transcriptome), and sketch (pseudoalignment with structural constraints) and sla (selective-alignment) label the mapping method used to obtain the result. Due to the similarity of the distributions, the line of STARsolo is occluded by the line of alevin-fry(splici, sla). (b) A visualization of the velocity estimation derived from alevin-fry counts in a UMAP-based embedding after assigning all ambiguous counts as spliced; the streamlines represent the direction of RNA velocity estimated by scVelo. Points (cells) are colored according to the cell type annotation. (c) The t-SNE embedding plot of an alevin-fry processed mouse placenta snRNA-seq dataset. The color of each nucleus represents the inferred cell-type annotation, which was learned from a reference dataset. (d) and (e) are timing and peak memory usage for all tools (run with 16 threads) on the different datasets evaluated in this manuscript. The x-axis of (d) and (e) represents the evaluated datasets. The y-axis of (d) represents the run time of each tool, measured in seconds. The y-axis of (e) denotes the peak memory usage — measured as the maximum resident set size (max rss) — during the execution of each tool. Dashed horizontal lines in (d) denote 15 minutes, 30 minutes, 60 minutes and 90 minutes, respectively. Dashed horizontal lines in (e) denote 4GB, 8GB, 16GB and 32GB, respectively.