Choice of pre-processing pipeline influences clustering quality of scRNA-seq datasets

Abstract

Background: Single-cell RNA sequencing (scRNA-seq) has quickly become one of the most dominant techniques in modern transcriptome assessment. In particular, 10X Genomics Chromium system, with its high throughput approach, turn key and thorough user guide made this cutting-edge technique accessible to many laboratories using diverse animal models. However, standard downstream processing, including the alignment and cell filtering pipelines might not be ideal for every organism or tissue. Here we applied an alternative strategy, based on the pseudoaligner kallisto, on twenty-two publicly available single cell sequencing datasets from a wide range of tissues of eight organisms and compared the results with the standard 10X Genomics Cell Ranger pipeline. Results: In most of the tested samples, kallisto outperformed Cell Ranger in sequencing read alignment rates and total gene detection rates. Although datasets processed with Cell Ranger had higher cell counts, outside of human and mouse datasets, these additional cells were routinely of low quality, containing low gene detection rates. Thorough downstream analysis of one kallisto processed dataset, obtained from the zebrafish pineal gland, revealed clearer clustering, allowing the identification of an additional photoreceptor cell type that previously went undetected. The finding of the new cluster suggests that the photoreceptive pineal gland is essentially a bi-chromatic tissue containing both green and red cone-like photoreceptors and implies that the alignment and processing pipeline can affect the discovery of biologically-relevant cell types. Conclusion: While Cell Ranger favors higher cell numbers, using kallisto results in datasets with higher median gene detection per cell. We could demonstrate that cell type identification was not hampered by the lower cell count, but in fact improved as a result of the high gene detection rate and the more stringent filtering. It is thus beneficial to favor high quality cells and accept a lower cell count, leading to an improved classification of cell types.

Identification of enamel knot gene signature within the developing mouse molar

“…Previously published raw reads from four E14 mouse molars (Hallikas et al ., 2021) were aligned to the mouse genome (mm10) with pseudoaligner kallisto, based on improved performance relative to other approaches (Shainer and Stemmer, 2020). A total of 33,886 cells passed quality control filters for number of genes detected, total counts, and percent of reads derived from mitochondrial gene expression (Fig.…”

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

“…Initially, we noticed that, despite the fact that Shainer and Stemmer (28) 1: The performance of the examined tools under various accuracy metrics on the simulated data. All metrics are measures on the subset of genes and cells defined by all tested methods, and are taken with respect to ground-truth abundances.…”

“…Taken together, these results suggest that the main factor in the separation of these clusters during processing is a combination of the specific filtering parameters used to retain cell barcodes, in conjunction with what specific UMI deduplication strategy was used. Of the methods tested, alevin-fry, kallisto|bustools, and STARsolo (with the 1MMDir and exact UMI resolution strategies) recovered the distinct clusters under the initial filtering described by Shainer and Stemmer (28) and under the kneedistance filtering of alevin-fry (though not under the filtering that directly made use of emptyDrops). Conversely, STARsolo (with the default 1MM UMI resolution strategy) produced counts under which these clusters were only discovered when evaluating the subset of cells returned by alevin-fry's knee-distance filtering procedure.…”

“…To explore the performance of alevin-fry in species other than human and mouse, and in an experimental sample where the alignment pipeline has previously been hypothesized to be of importance for downstream analysis, we analyzed the Danio rerio pineal gland dataset introduced in Shainer et al (27), and subsequently re-analyzed in (28).…”

“…To further investigate the differences reported in (28), we processed this data with STARsolo, kallisto|bustools, and alevin-fry using USA mode and the splici reference. Further, we ran STARsolo with 1MM (default), 1MMDir and exact UMI deduplication strategies.…”