Benchmarking bulk and single-cell variant-calling approaches on Chromium scRNA-seq and scATAC-seq libraries

Benchmark data sets

Three tumor samples derived from a genetically engineered mouse model of epithelioid sarcoma (EPSRC) featuring an induced adulthood deletion of Smarcb1 were acquired, and each sample underwent Chromium scRNA-seq and scATAC-seq, as well as Illumina WGS, library generation (see Supplemental Table S1). Additionally, a control sample underwent Chromium scRNA-seq and scATAC-seq. Single-cell sequencing was performed at loadings of around 7000–15,000 cells per sample. This loading is typical for Chromium pipelines and serves to reproduce the low per-cell coverage typical of this single-cell methodology.

Benchmark metrics

Our unique experimental setting allowed us to capture both real variants that would be present in many cells as well as rare variants that would be present in a small fraction of cells. To identify real variants, the consensus of multiple variant callers (HaplotypeCaller [Poplin et al. 2018], Strelka2 [Kim et al. 2018], SAMtools [Li 2011], and VarScan [Koboldt et al. 2009]) on the matched WGS libraries was utilized (referred to as common ground-truth variants) (Table 1). Given that WGS is limited in its capability to detect variants present in a small number of cells, compared with other modalities, we would expect that such rare variants would most often be undetected or indistinguishable from noise in WGS data. As such, variant calls identified in scRNA-seq or scATAC-seq data without significant evidence in WGS were treated as potential rare variants (referred to as rare positive variants). We define a number of variant categories and metrics (Table 1; Supplemental Fig. S1) to facilitate comparison of various variant-calling algorithms. These are described in more detail in Methods (see subsection “Variant categories and metrics”).

As our primary metric to compare various variant-calling approaches on single-cell sequencing, we use a precision-recall measure that evaluates missed variants (pooled false negatives) and extra variants that have been called (rare positives) for each caller. We used the quality value for each variant generated by each caller as the score threshold to generate precision and recall across different levels. We further computed the area under the precision-recall curve (AUPRC) to compare the overall performance of each caller. Precision and recall were computed for variants across multiple samples to provide an overall score. We additionally use metrics termed the common true-positive rate (cTPR) and common false-discovery rate (cFDR) to evaluate the performance of variant callers with different filtering approaches. For more details on these metrics, see Methods (subsection “Variant categories and metrics”).

Benchmarking variant callers

We consider variant-calling approaches for single-cell data sets as falling into two categories: pooled and individual-cell methods. Bulk callers have been tested previously on Chromium scRNA-seq and scATAC-seq libraries at the pooled read level (Liu et al. 2019; Massarat et al. 2021). However, no studies have tested the application of bulk callers to individual cells for Chromium libraries. Given that such approaches are increasingly being used (Ding et al. 2019; Lu et al. 2021), it is essential to benchmark their performance against bulk methods and confirm their utility.

Typically, for individual-cell calling, the intersection of variants called across multiple cells is used to derive a consensus list that can be filtered by quality. To investigate the potential of these approaches, we apply Strelka2 (Kim et al. 2018) and GATK HaplotypeCaller (Poplin et al. 2018) on individual BAM files for reads corresponding to a single cell. We then take the intersection of variants from individual cells and define the quality of each variant as the maximum quality value of the variant over the cells in which it was found.

Based on previous benchmarking studies applying bulk callers to Chromium single-cell pooled reads (Liu et al. 2019; Massarat et al. 2021), we test Strelka2, SAMtools (Danecek et al. 2021), GATK HaplotypeCaller, and FreeBayes (Garrison and Marth 2012) as the most promising options, given their previously tested performance. We run them on the pooled reads for each single-cell library according to their recommended best-practice approach for each modality.

It is notable that there are many purpose-built options for individual-cell variant calling, such as Monovar (Zafar et al. 2016), SCcaller (Dong et al. 2017), SciPhi (Singer et al. 2018), ProSolo (Lähnemann et al. 2021), ScanSNV (Luquette et al. 2019), scAllele (Quinones-Valdez et al. 2022), and Monopogen (Dou et al. 2024). Many of these callers were developed to address allelic bias owing to multiple-displacement amplification used in scDNA-seq methodologies. As a result of their tailored statistical modeling approaches and traditional application to high-coverage single cells, they have been dismissed as inapplicable to Chromium single-cell methodologies (Lu et al. 2021). However, the ability of these tools to adaptively model changing allele frequency might be helpful to correct for allelic imbalance owing to preferential visibility of one chromosome in scRNA-seq and scATAC-seq, thus improving variant detection. The major question is whether they will transfer to the low read coverage per cell of Chromium-based sequencing. To verify this viability, we benchmark SCcaller, which among the caller options has a varying window allelic modeling approach suitable for RNA sequencing.

It should be mentioned that there are recently developed callers developed specifically for improving sensitivity in single-cell modalities through the use of additional population-level information to overcome the limitations of single-cell sequencing. Examples include Monopogen (Dou et al. 2024), which leverages population-level statistics such as linkage disequilibrium to help differentiate real variants from noise, and scAllele (Quinones-Valdez et al. 2022), which leverages splicing patterns to help inform variant confidence. Although we do not benchmark them in this paper, approaches like these may have significant promise for Chromium-based single-cell modalities.

Given that single-cell data have the potential to capture variants that show evidence in a low number of cells and thus have a low presentation frequency, we first investigated the performance of variant callers at different variant allele frequencies (VAFs; the fraction of alternate-allele reads over total reads at a variant site). We can see from Figure 1, C and D, that, for both scATAC-seq and scRNA-seq, the majority of variants have either very low VAF (less than 0.01) or very high VAF (equal to 1.0). As such, a failure to detect low-VAF variants will miss many WGS variant sites being lost. All callers suffer a drop in sensitivity at low VAF, especially below VAF of 0.01 (Fig. 1A,B). However, two of the individual-cell callers, SCcaller and Strelka, perform quite well at low VAF compared with all other callers, possibly highlighting the potential of individual-cell callers to pick up lower-frequency variants. Based on the results of Figure 1, in our further analyses, we kept only those variants with VAF of 0.01 or greater to filter out the large number of very-low-frequency variants missed by all callers.

Figure 2, A and B, demonstrates the performance of variant-calling approaches on scRNA-seq and scATAC-seq, respectively, based on precision-recall curves. We can see that pooled read calling can achieve good sensitivity in both cases, reaching up to 90.5% and 76.9% recall in each modality, respectively. Based on AUPRC, we find that Strelka demonstrates the best performance on scATAC-seq, whereas SAMtools demonstrates the best performance on scRNA-seq (see Table 2). Notably, in scRNA-seq, most callers have lower precision than in scATAC-seq, indicating that extra variants are consistently called in scRNA-seq that are not present in the common ground-truth set. This is especially notable when variants are called in nonexonic and exonic regions (see Supplemental Fig. S2) and raises the importance of limiting scRNA-seq variant calling to exonic regions. In later sections, we investigate what proportion of these rare positive variants are noise/artifacts, biological factors unique to sampling RNA, or true biological variants. The fact that callers perform well at detecting common ground-truth variants indicates that the WGS variants are well represented at the pooled read level in both modalities.

Contrary to bulk callers applied to the pooled reads, the same callers applied at the individual-cell level do not effectively resolve common ground-truth variants. Both Strelka SC and HaplotypeCaller SC demonstrate low recall at most filtering thresholds, only detecting common ground-truth variants when quality filters are so stringent that only a handful of variants remain (corresponding to the left side of the graph in Figs. 2A,B). On scATAC-seq data sets, the single-cell callers do manage to match pooled read calling for precision at the very extremum of their quality threshold, indicating that variant calling at the single-cell level is able to resolve common ground-truth variants but at the cost of extremely low recall. HaplotypeCaller, on the other hand, performs poorly at any filtering threshold, indicating a very low sensitivity to common ground-truth variants.

A common expectation of variant callers applied at the individual-cell level is the ability to resolve rare variants that have low alternate allele evidence. To determine the ability of variant callers to detect low-read-support variants, we look at the variants called on the single-cell libraries without applying any filtering to the variant sites determined by each caller. As we can see in Figure 3, A and B, certain bulk callers considerably outperform others, with Strelka consistently demonstrating the highest sensitivity for both scRNA-seq and scATAC-seq. Additionally, the individual-cell application of Strelka (Strelka SC) appears to match the performance of Strelka when called in bulk. On the other hand, HaplotypeCaller calling on individual cells (HaplotypeCaller SC) demonstrates poor performance at detecting common ground-truth variants across all read coverage levels, indicating that the caller has decreased sensitivity when applied at the single-cell level.

Individual-cell application of bulk callers demonstrate extremely high numbers of rare positive variants, as is evidenced by the cFDR statistics in Figure 3, C and D. HaplotypeCaller SC demonstrates lower FDR than Strelka SC, possibly because of its decreased sensitivity. Pooled read callers (with the exception of FreeBayes) demonstrate similar performance in cFDR across allele coverage levels, indicating more rare positives for variants with low alternate allele coverage.

By comparison to other calling approaches, SCcaller significantly underperforms, mostly owing to a high cFDR, especially in scRNA-seq. Although the cTPR of the caller is decent, being outperformed only by Strelka, it appears that SCcaller is indiscriminate in its sensitivity. Given that SCcaller has been shown to obtain good results on MDA scDNA-seq data sets, this may be primarily owing to the low coverage present in the cells. Additionally, Strelka SC also generates a very high cFDR across different allele coverages in the case of scRNA-seq. Overall, this result supports the suspicion of others that existing single-cell variant callers are not suitable for Chromium data, although this may not be the case for all such callers.

Overall, these results agree with previous studies demonstrating that a wide range of variants present in a sample are visible in libraries generated by Chromium scRNA-seq and scATAC-seq modalities (Liu et al. 2019; Massarat et al. 2021). This enables such modalities to supplement, or act independent of, DNA sequencing to discover biologically relevant variant sites. The choice of caller has a significant impact on the ability to resolve these variants. Applying bulk callers to individual cells results in extremely high rare positive rates likely owing to a large number of false-positive calls, making such approaches impractical for calling biologically relevant variants. The potential veracity of such rare positives is discussed in later sections. Applying bulk callers to pooled reads can provide quite useful results, even at standard-quality levels (see dots on Fig. 2). Strelka, HaplotypeCaller, and SAMtools perform well on both scRNA-seq and scATAC-seq, with small differences such that it is difficult to say which one is best. Depending on whether precision or recall is more important and depending on which modality is used, Strelka or SAMtools might give optimal performance. There appear to be trade-offs between callers, and we hope the data we have presented here will help inform researchers in the field as to which approach they would prefer.

Investigating rare positive variants in pooled reads

Previously, we noted that bulk callers applied to pooled reads detect variants in addition to those found in the common ground-truth set. It is possible that some of these rare positive variants are either real and not detected in WGS or owing to other biological factors. ATAC-seq and RNA-seq modalities frequently sample specific regions to a depth much higher than the 40× coverage we obtained for our WGS. This deeper coverage may enable higher alternate allele visibility and greater confidence in variant calling at a given locus. However, rare positive variants may just as well be the result of artifacts owing to noise/error or modality-specific sequence alterations. Considering that explicit verification of these variants would likely require high-depth sequencing to confirm, we use indirect approaches to draw conclusions about the portion of rare positive variants that are potentially real.

To examine the composition of rare positive variants in general, we defined union rare positive (URP) variants as the union of rare positive variant calls for all callers performed on pooled scRNA-seq or scATAC-seq reads. We quantified the number of URP variants without any filtering except that which demonstrate evidence (at least one alt-allele read) in WGS (Table 3). This is to account for any variants that did not pass filters for calling in our common ground-truth set (i.e., WGS) but that gave sufficient evidence in scRNA-seq or scATAC-seq to be called. We find that, on average, 9.8% of URP variants called in ATAC-seq (scATAC-RPs) are present in at least five reads in WGS, with 6.8% present in at least 10 reads. On the other hand, only 2.2% and 1.2% of scRNA-seq URP variants (scRNA-RPs) show evidence of being present in WGS at depths greater than five or 10 reads, respectively. Thus, it appears that a portion of scATAC-RPs may be real and possibly missed in the common ground-truth set, whereas this is significantly less so for scRNA-RP variants. This raises a concern that scRNA-RPs are enriched for noisy or artifactual variants.

It should be noted that scRNA-seq appears to generate more URPs than scATAC-seq. If we normalize by the total coverage of each modality (see Table 3), we find that scRNA-seq has 2.9× more URP variants per megabase than scATAC-seq (P = 0.087 by two-tailed paired t-test, n = 3), despite scATAC-seq detecting similar or more (∼1.8×) common ground-truth variants than scRNA-seq (P = 0.068 by two-tailed paired t-test, n = 3). The extra variants we see may either be real and owing to differences in coverage location or owing to higher error rates in scRNA-seq.

To determine if rare positive variants tend to have a lower read coverage or lower VAF distribution, potentially indicating noise as a cause for these variants, we broke down scRNA-RP and scATAC-RP variants by VAF (Fig. 4A,C) and read depth (Fig. 4B,D) in comparison with pooled TP variants for control. Notably, although URP variants in both modalities appear to be slightly more distributed toward lower VAFs compared with TP variants, there appears to be no meaningful difference that would indicate VAF as a major correlating factor for rare positive variants. Likewise, the read depth distribution of URP for both modalities is relatively similar to that of TP variants, with TP variants presenting in a higher portion of low-read-coverage sites than URP variants, especially for scATAC-seq. The tendency of URPs to occur in regions of higher coverage for scATAC-seq may be indicative of a systematic source for their generation, which we attempt to explore more in the next section.

Analysis of the mutational signatures of variants, specifically the base sequence in the context of a SNV or indel locus, is a common approach for resolving the mechanisms underlying a particular variant (Omichessan et al. 2019). Such analyses have been used to identify signatures corresponding to mutations resulting from artifacts in sequencing and variant calling (Alexandrov et al. 2020). Here we used the SigProfiler tool developed by Bergstrom et al. (2019) to extract and fit known signatures for scATAC-seq and scRNA-RP variants, as well as common ground-truth variants. The COSMIC (Alexandrov et al. 2020) signatures serve as a reference of well-known mutational signatures that might compose some part of the variants seen in our samples.

We find that scATAC-RP variants feature a trinucleotide mutational breakdown similar to that seen in WGS (Fig. 5A), namely, with a high proportion of variants that may be explained as a composition of the COSMIC signature SBS5 (Fig. 6A). scATAC-RPs in EPSRC1 also appear to match the SBS45 and SBS46 signatures, which have been linked to sequencing errors. scATAC-seq also features a prominent CTG→CCG rate (Fig. 5A), which does not match any known signatures that we have reviewed and may be because of a preparation/sequencing error. As such, it appears that, although scATAC-RPs may contain real variants, there is a substantial risk of including noise based on the signatures observed.

On the other hand, the trinucleotide mutational breakdown of scRNA-RPs variants differs substantially from that of WGS, demonstrating an extremely high incidence of repeat extension mutations of the form YY[X]YY > YY[Y]YY, especially in poly(T)/poly(C) regions (Fig. 5A,B). These mutations are present even if we reduce the list of variants to those found in exonic regions, which we would expect to have higher quality owing to increased coverage by scRNA-seq (Fig. 5B). According to SigProfiler, the signature of these mutations does not match any yet established signature, although they are somewhat similar to SBS58, another COSMIC potential sequencing artifact signature (Fig. 6A,B).

These previous results suggest that many URP variants are caused by noise. One major advantage of our data set is the matched scRNA-seq and scATAC-seq, which allows rare variants to be validated by confirming their presence in both modalities. We use the intersection of scRNA-RP and scATAC-RP variants as a validation set for rare positive variants, which we term multimodal rare positive (MRP) variants. We use estimated MRP and URP variant statistics (for definition, see Methods, subsection “Variant categories and metrics”; a minimum threshold of five supporting reads) (Table 4) alongside pileup overlap statistics to extrapolate the number of valid URP variants we would expect to see in each modality. We find that the rate of MRP variants per megabase is extremely small. If we extrapolate this rate to the URP variants unique to each modality, we find that they would make up ∼2% and 14% of unique URP variants in scRNA-seq and scATAC-seq, respectively. This agrees well with the portion of rare positive variants seen with evidence of the alternate allele in WGS discussed previously. It should be noted that the true number of overlapping rare variants may be larger owing to regions where one modality has very low coverage and thus does not sample the rare alternate allele. Overall, these results suggest that rare variants detectable through pooled-read variant calling in scRNA-seq and scATAC-seq make up only a small portion of rare positive variants.

Overall, it appears that many of the URP variants may be explained as owing to noise. Some of the URP variants show evidence of being real rare mutations, indicating that scATAC-seq and scRNA-seq de novo calling may yield variants not visible in typical WGS experiments. However, they make up such a small portion of the URP variant set that they would be infeasible to separate out from the majority of noisy variants in any practical scenario. The increased incidence of scRNA-RPs compared with scATAC-seq, as well as characteristic repeat extension mutations in scRNA-seq but not in scATAC-seq, strongly suggests that most of the scRNA-RPs are false positives generated by RNA-modality-specific effects. These effects may be the result of scRNA-seq-specific qualities, such as high coverage in low-complexity regions (3′ UTR, 5′ UTR and intronic regions) and poly-adenylation, although further investigation would be required to confirm this. In scRNA-seq, although these false positives are frequent, it may be possible to reduce their incidence. We recommend others consider removing repeat-extension variants of the forms TT[C]TT > TT[T]TT, GG[C]NG > GG[G]NG, and CC[T]CN > CC[C]CN. On the other hand, scATAC-RPs seem to contain a more prominent mixture of real variants among the noisy variants. As we have not discovered an easy way to separate these, it appears that noise must simply be accepted.

Single-cell-unique variants

As described previously, there is an abundance of variants that are unique to single-cell sequencing modalities, which tend to present in a low number of reads and hence a low number of cells. Previous studies have attributed these variants to real mutations unique to small-cell populations caused by high mutation rates in cancer. Given the incidence of systematic errors in regions of high coverage or motif repeats (Li 2014; Tørresen et al. 2019; Stoler and Nekrutenko 2021), variant loci called on loci with a low number of supporting reads should be treated with caution. Before any claims of rare single-cell-unique (scu) variants may be made, it is essential to determine what portion of these variants may be the result of noise or error.

To estimate the potential of rare single-cell-unique variants in our single-cell libraries, we quantify scu-variants in scRNA-seq and scATAC-seq. In line with previous results, we find abundance of scu-variants (average of 1180 per sample at this filtering level) for scRNA-seq (see Table 3).

To evaluate whether these scu-variants are the result of rapid mutations characteristic of cancer, we extract the scu-variants from an unmatched control for which we have scRNA-seq and scATAC-seq, but no WGS. To account for the lack of matched WGS, we filter out all variants seen in the intersection of the common ground-truth set from each of the sarcoma samples. We find considerably more scu-variants are generated (6600 for CTRL vs. average of 1170 for EPSRC). Repeating the analysis performed by Prashant et al. (2021), if we examine the breakdown of scu-variant locations and impact annotated by SnpEff (Cingolani et al. 2012), as seen in Figure 7, we find a breakdown very similar to that of the original paper, with 3′ UTR variants being extremely common and an abundance of missense and intergenic variants (see Fig. 7A). The rare positive variants called on pooled reads (referred to as URP variants in Table 1) also demonstrate breakdowns similar to those of Prashant et al. (2021; see Fig. 7B). Although our breakdowns are not directly comparable to previous results owing to a difference in the annotation tool used (SnpEff rather than SeattleSeq), these similarities give confidence that we have reproduced the scu-variants seen in previous studies.

Notably, the distribution of both scu-variants and rare positive variants in our wild-type control closely follows that of the sarcomas. A chi-squared test of the category counts for scu-variants across samples results in a P-value of 0.0004 (see Table 5), which is mostly owing to the decreased incidence of intergenic and noncoding variants seen in EPSRC1. If intergenic and noncoding categories are removed, or EPSRC1 is removed from comparison, the P-value shifts to nonsignificant values of 0.357 and 0.359, respectively, indicating strong agreement. Because intergenic and noncoding regions typically have poor coverage in scRNA-seq, we are inclined to conclude these extra variants in EPSRC1 are caused by noise, possibly resulting from differences in the coverage profile of the scRNA-seq library sampled from EPSRC1. This overall result demonstrates that scu-variants are found in both cancerous and healthy tissue consistently across the same genomic locations and therefore cannot be easily attributed to low frequency somatic mutations associated with cancer development as previously described.

The abundance of scu-variants seen in Chromium scRNA-seq appears to be modality specific, as performing the same analysis on scATAC-seq returns, on average, ∼230× fewer variants for the same filtering levels (average, 27 scu-variants) (see Table 3). This is surprising considering the significantly higher breadth of coverage seen in scATAC-seq versus scRNA-seq. Although the two modalities cover different genomic regions, with scRNA-seq focusing on the exome, we would not expect the background somatic mutation rate to vary so drastically by region. This discrepancy further implies that the abundant scu-variants seen in scRNA-seq are the result of RNA-specific artifacts.

Overall, although we do not dismiss the possibility that scu-variants may capture rare cancerous somatic mutations, the abundance of scu-variants in scRNA-seq seen in both normal and cancerous tissue suggests that isolating rare somatic variants from background noise in scRNA-seq would be a considerable challenge. The lack of scu-variants in scATAC-seq is further evidence that those seen in scRNA-seq are because of noise. As such, we conclude that the utility of scu-variants as candidate rare somatic variants is questionable at best.

RNA-editing events

RNA-editing events are a type of post-transcriptional modification that can alter the sequence of an mRNA transcript to modify the function or localization of the resulting protein (Nishikura 2010). RNA editing is widespread among eukaryotic organisms, and specific modifications are often tissue specific (Song et al. 2004; Picardi et al. 2015). There is a considerable interest in furthering our understanding of editing sites, such as the resulting impact on function and the context in which it occurs, as RNA editing has been suggested as a safer method of genomic alteration for cancer therapy (Reardon 2020). As a result, there is a demand for approaches that can resolve RNA editing at the cell type level and link this to cell activity.

RNA sequencing can capture RNA-editing events that occurs on mRNA strands as they are reverse-transcribed to cDNA for tagging and amplification. Moreover, scRNA-seq methods may resolve this at the single-cell level, enabling cell type–specific characterization of editing sites at a much finer level of detail than can be achieved with many bulk-RNA-seq libraries. Thus, scRNA-seq is uniquely suited to provide additional insight into RNA editing. However, some studies have questioned the ability to resolve RNA-editing events in tag-based scRNA-seq approaches such as Chromium (Chen et al. 2019) We set out to determine if RNA-editing events are visible in Chromium scRNA-seq libraries and to determine if they may resolve cell type–specific effects.

Prior work has amalgamated libraries of common RNA-editing sites in mice (Mansi et al. 2021). Using overlap between scRNA-RP variants derived from our scRNA-seq data sets and the known RNA-editing sites, we attempted to characterize the visibility of RNA edits in Chromium 3′ scRNA-seq. Across our four scRNA-seq samples, we find a total of 4942 unique variant sites (average, 1870 per sample) matching RNA-editing events in the REDIportal database, out of a total of 107,095 variants (average, 1.75% detection rate). For context, in scATAC-seq-derived common ground-truth and scATAC-RP variants, only two and 31 matches were discovered, respectively. The coverage of editing sites seen in scRNA-seq can vary, with most demonstrating low coverage. However, many editing sites appear in a sufficient quantity of reads such that many cells within a sample demonstrate visibility of the editing site (Fig. 8A). Given that editing alleles often appear on a fraction of reads (Fig. 8A), it is possible that a considerable number of editing sites may have relevance in cell type–specific analysis of editing events.

Most editing sites did not present in an adequate number of cells to infer cell type–specific editing. However, we found one site at Chr 7: 130,985,524 presenting as an A-to-G base substitution that demonstrated differential presentation of the RNA edit based on the expression profile of the cells in the sample. This was found in EPSRC1, and we display the presentation of the RNA-edit allele presentation frequency per cell, termed VAF, on the UMAP embedding of expression generated using the Seurat R package (Fig. 8D). For comparison, we used the expression profile of cells to perform unsupervised Louvain clustering for any cell type heterogeneity within our tumor sample (Fig. 8B). As we can see, the VAF varies considerably across the sample, with clusters 5 and 2 being enriched for the RNA edit and clusters 3 and 4 showing an absence of the edit (P = 4.0 × 10⁻¹⁴ by Fisher's exact test using alternate allele and reference allele counts across cluster categorization). As can be seen in Figure 8C, this is not because of differences in expression levels, as the gene on which the edit is located (Htra1) is expressed in all involved clusters.

Overall, this constitutes evidence that RNA-editing events may be seen in Chromium scRNA-seq libraries and used to infer heterogeneous presentation of RNA editing. Although only one differentially presented site was found in our samples, there are many other sites with high coverage. Given that our samples are relatively homogeneous with respect to cell type compared with other studies owing to all being sampled from a mesenchymal tumor phenotype, it is possible that performing the same analysis on Chromium scRNA-seq libraries composed of multiple distinct cell types would yield additional cell type–specific presentation of RNA edits. However, we realize that a fair number of the variants found to match known PTM sites may be false positives owing to noise, as discussed in previous sections. Furthermore, without a ground-truth data set of PTMs to compare, we cannot measure the accuracy of the approach. Thus, significant further study is required to confirm whether the results we have found represent real measurable RNA-editing events in Chromium scRNA-seq.