Benchmarking of computational demultiplexing methods for single-nucleus RNA sequencing data

Probabilistic method-based variant calling tools performed best in previous benchmarking studies.

“These tools were chosen since probabilistic method-based tools performed best in previous benchmarking studies [33, 37].”

• cites [33] Andreu-Sánchez, 2021, A benchmark of genetic variant calling pipelines using metagenomic short-read sequencing, Front Genet sourceSupported
• cites [37] Weber, 2021, Genetic demultiplexing of pooled single-cell RNA-sequencing samples in cancer facilitates effective experimental design, GigaScience sourcePartly supported

  The Discussion (page unavailable) reports that a combination using bcftools to generate a SNP reference and cellSNP/Vireo for demultiplexing performed best in this paper’s benchmark. It does not establish that probabilistic variant-calling tools, as a class, performed best in previous benchmarking studies. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Cell hashing accuracy is lower on nuclei isolated from archived/frozen samples (snRNA-Seq).

“accuracy is lower on nuclei isolated from archived/frozen samples (snRNA-Seq) [15, 19, 20].”

• cites [15] Wang, 2024, scCDC: A computational method for gene-specific contamination detection and correction in single-cell and single-nucleus RNA-seq data, Genome Biol sourceCould not verify

  The Methods section (“Single-nuclei RNA-seq of mammary glands”; page unavailable) says the tissues were snap-frozen before nuclei extraction. The Background discusses ambient RNA contamination in snRNA-seq, but neither passage assesses cell hashing accuracy. Searches for cell hashing, frozen nuclei, and archived samples found no such comparison in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [19] Mylka, 2022, Comparative analysis of antibody- and lipid-based multiplexing methods for single-cell RNA-seq, Genome Biol sourcePartly supported

  The Results section ('Lipid (CMO)-based hashing outperforms antibody hashing on nuclei'; page unavailable) reports lower signal-to-noise ratios for nuclei hashing than for antibody cell hashing, and lower accuracy for antibody hashing of nuclei than for lipid hashing of nuclei. The nuclei came from human cancer cell lines; the paper does not establish that accuracy is lower specifically for nuclei isolated from archived or frozen samples, or for every hashing method. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [20] Hoeft, 2024, Label-free single-cell RNA multiplexing leveraging genetic variability, Nat Commun sourcePartly supported

  In Results (Benchmarking of SoupLadle framework in different tissues; page unavailable), hashtag-based labeling was tested on nuclei from snap-frozen heart tissue, and the two cell-labeling strategies assigned only about 75% of nuclei, versus nearly all with SoupLadle. But the paper reports greater than 95% agreement among methods for nuclei that were assigned, including Hashtag. It supports lower assignment coverage than the comparator method, not a general claim that cell hashing is less accurate on frozen nuclei than on cells. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Cell hashing methods can achieve >90% demultiplexing accuracy when applied to live cells.

“although when applied to live cells these methods can achieve >90% demultiplexing accuracy (defined as the number of all matching singlets plus the number of matching non-singlets divided by the total number of cells), accuracy is lower on nuclei isolated from archived/frozen samples (snRNA-Seq) [15, 19, 20].”

• cites [15] Wang, 2024, scCDC: A computational method for gene-specific contamination detection and correction in single-cell and single-nucleus RNA-seq data, Genome Biol sourceCould not verify

  The Results and Discussion passages address ambient RNA contamination and correction in cells and nuclei, not cell-hashing demultiplexing. Searches for cell hashing, demultiplexing, and the >90% accuracy figure found no passage in our copy that addresses the claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [19] Mylka, 2022, Comparative analysis of antibody- and lipid-based multiplexing methods for single-cell RNA-seq, Genome Biol sourceSupported
• cites [20] Hoeft, 2024, Label-free single-cell RNA multiplexing leveraging genetic variability, Nat Commun sourcePartly supported

  The Results section, “Benchmarking of scRNA-seq multiplexing methods” (page unavailable), reports ~98.5% precision for assigned single cells, including those assigned by CellPlex in a PBMC experiment. That measure excludes doublets and counts only assigned cells; it does not establish >90% accuracy under the citing sentence’s definition, which includes matching non-singlets and divides by all cells. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Mixing species is an effective strategy to demultiplex samples.

“Although limited to two samples, as shown in our analysis, sex chromosomes are also an effective strategy to demultiplex 2 samples [43], as can the mixing of species—as in a barnyard experiment [11].”

• cites [11] Macosko, 2015, Highly parallel genome-wide expression profiling of individual cells using Nanoliter droplets, Cell sourceSupported

Without identification and removal of doublets, cell type clustering, differential gene expression analysis, and inference of cell developmental trajectories are severely compromised.

“Without identification and removal of doublets, cell type clustering, differential gene expression analysis, and inference of cell developmental trajectories are severely compromised [26, 27].”

• cites [26] Luecken, 2019, Current best practices in single-cell RNA-seq analysis: A tutorial, Mol Syst Biol sourcePartly supported

  The Quality control section recommends filtering potential doublets before analysis, and the Downstream analysis section identifies clustering and trajectories as later steps. The paper does not say that failing to remove doublets severely compromises clustering, differential expression analysis, or developmental trajectory inference. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [27] Wolock, 2019, Scrublet: Computational identification of cell doublets in single-cell transcriptomic data, Cell Syst sourcePartly supported

  The Introduction (page unavailable) says doublets can appear as distinct cell types, bridge cell states, and interfere with differential gene-expression tests. The Results section on cell-state continua describes artifactual bridges between progenitor types. These passages support risks to clustering, expression analysis, and trajectory inference, but do not establish that all three are necessarily severely compromised without doublet removal. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Sex chromosomes are an effective strategy to demultiplex 2 samples.

“Although limited to two samples, as shown in our analysis, sex chromosomes are also an effective strategy to demultiplex 2 samples [43], as can the mixing of species—as in a barnyard experiment [11].”

• cites [43] Nassiri, 2024, Demultiplexing of single-cell RNA-sequencing data using interindividual variation in gene expression, Bioinforma Adv sourceCould not verify

  The Results (p. 4) describe demultiplexing by interindividual gene-expression patterns, and the Discussion (p. 16) discusses SNP- and barcode-based methods. Searches for sex chromosomes, male–female demultiplexing, and XIST found no passage evaluating sex chromosomes as a way to separate two samples. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

The prevalence of doublets is particularly increased in protocols involving increased cell/nuclear analysis (overloading) during sequencing.

“The prevalence of doublets is particularly increased in protocols involving increased cell/nuclear analysis (overloading) during sequencing [25].”

• cites [25] Bernstein, 2020, Solo: Doublet identification in single-cell RNA-Seq via semi-supervised deep learning, Cell Syst sourceSupported

Inappropriate use of published methods or incorrect selection of user settings can result in bias in data analysis and false conclusions.

“Inappropriate use of these published methods or incorrect selection of user settings can result in a bias in data analysis (such as by introducing artificial, non-biological signals) and false conclusions [41, 42].”

• cites [41] Andrews, 2019 sourceSupported
• cites [42] Weber, 2019, Essential guidelines for computational method benchmarking, Genome Biol sourcePartly supported

  The Introduction’s “Parameters and software versions” section says parameter settings affect method performance and that uneven tuning can bias benchmark results. Its “Interpretation, guidelines, and recommendations” section warns that poorly explained limitations can mislead readers. These passages concern benchmarking; they do not establish that inappropriate use of published methods introduces artificial, non-biological signals into data analyses. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Doublets can confound gene expression analysis, cell clustering, and cell type annotation if not considered in downstream analysis.

“Doublets arising from the inclusion of either two independent nuclei or more nuclei that have become attached during isolation being encapsulated within a GEM resulting in the generation of a single transcriptome can, if not considered in downstream analysis, confound gene expression analysis as well as cell clustering and cell type annotation [26, 27].”

• cites [26] Luecken, 2019, Current best practices in single-cell RNA-seq analysis: A tutorial, Mol Syst Biol sourcePartly supported

  The Quality control section says that unusually high counts and gene numbers may indicate doublets and recommends filtering potential doublets before analysis. The Cluster analysis section explains how expression profiles inform clustering and annotation, but does not specifically say that doublets confound those steps. The cited passages also do not describe nuclei becoming attached or their encapsulation within a GEM; no page numbers are available in our copy. (LLM, conf=0.80)

  Experimental — accuracy not yet measured
• cites [27] Wolock, 2019, Scrublet: Computational identification of cell doublets in single-cell transcriptomic data, Cell Syst sourceSupported

Recall for all methods except scSplit was close to 100%, even with 30% doublets, in the four and six sample datasets.

“The recall value obtained by all methods except for scSplit was close to 100%, even with 30% doublets.”

At 0% doublets, the demultiplexing tools exhibited precision ranging from 100% to 98%.

“At 0% doublets, the demultiplexing tools exhibited precision ranging from 100% to 98%, which gradually decreased with increasing doublets to the lowest precision of 70% obtained with Scsplit.”

The best-performing pipelines for two samples were also best at demultiplexing 4 and 6 samples.

“As a general pattern, the best-performing pipelines for two samples were also best at demultiplexing 4 and 6 samples.”

Vireo achieved the best demultiplexing accuracy among the evaluated tools.

“Vireo achieved the best accuracy.”

• Also stated in: § Computational efficiency, robustness, scalability, and software implementation of demultiplexing methods · plain

  “while Vireo achieved the greatest accuracy (precision and recall) and was the least computationally intensive method.”

scSplit had a median recall of 52%.

“which had a median of 52% (Fig. 1A).”

Increasing sample number did not exacerbate the negative influence of doublets on precision.

“Moreover, increasing sample number did not exacerbate the negative influence of doublets on precision.”

Using array-derived SNPs, Vireo, Souporcell, and Freemuxlet successfully disaggregated each sample within the real-world two-sample dataset.

“As shown in Fig. 1B, using the array-derived SNP for demultiplexing, Vireo, Souporcell, and Freemxulet successfully disaggregated each sample within this real-world dataset.”

All pipelines that included cellSNP had inferior performance to those that included other variant calling tools.

“All pipelines that included cellSNP had an inferior performance to those which included the other variant calling tools.”

The Freemuxlet-BCFtools pipeline showed poorer performance as sample number increased.

“The Freemuxlet BCFtools pipeline was an exception to this finding, showing a poorer performance as the sample number increased.”

There is a strong correlation between the number of variants detected by the calling tools and demultiplexing performance.

“Indeed, the strong correlation between the number of variants detected by the calling tools with demultiplexing performance supports the importance of variant number in the capacity of the demultiplexing tool to perform its function.”

As sample number increased from two to four and six, the performance of cellSNP combined with Vireo decreased further, particularly in recall.

“As the sample number increased from two to four and 6, the performance of cellSNP in combination with Vireo decreased further, particularly in recall (Fig. 3B, C).”

Recall across all doublet percentages was 100% for all tools other than scSplit.

“Recall that all doublet percentages was 100% for all tools other than scSplit, which had a median of 52% (Fig. 1A).”

The precision obtained using Vireo increased as the sample number increased from two to six.

“Notably, the precision obtained using Vireo increased as the sample number increased from two to six (Fig. 4A).”

At 0% doublets, all tools successfully demultiplexed the four and six sample datasets, achieving close to 100% precision.

“At 0% doublets, as observed for two samples, all tools successfully demultiplexed the four and six samples, achieving close to 100% precision.”

At 0% doublets, cellSNP combined with Vireo achieved precision and recall of 90% and 88%, respectively.

“While at 0% doublets, cellSNP combined with Vireo performed with a precision and recall of 90% and 88% respectively (Fig. 3A), when used in combination with Souporcell and Freemuxlet, precision and recall were both 0.”

For demultiplexing using genetic variation from SNP array, with the exception of scSplit, increasing sample number did not significantly affect the accuracy of each pipeline.

“For demultiplexing using genetic variation obtained from the SNP array, with the exception of scSplit, increasing sample number did not significantly affect the accuracy of each pipeline.”

The precision of all pipelines decreased to 75% when doublets were set to 30%.

“Moreover, the precision of all pipelines decreased to 75% when doublets were set to 30% (Fig. 3C).”

Demultiplexing tools were differentially affected by the variant calling tool with which they were paired.

“Demultiplexing tools were differentially affected by the variant calling tool with which it was paired.”

For all demultiplexing tools, accuracy decreased with increasing percentage of doublets.

“For all tools, accuracy decreased with the increasing percentage of doublets.”

• Also stated in: § Use of sample-specific genetic variation obtained from SNP array for demultiplexing of pooled snRNA-Seq data · plain

  “With an increasing percentage of doublets, demultiplexing precision decreased for all tools, with scSplit showing the greatest decline in performance, achieving a median precision of ~70% at 30% doublets (Fig. 2A, B).”

Freemuxlet combined with FreeBayes had precision and recall of 0 for the two sample pool.

“For the two sample pool, Freemuxlet in combination with FreeBayes also had a precision and recall of 0 (Fig. 3A).”

cellSNP combined with Souporcell or Freemuxlet had precision and recall of 0.

“when used in combination with Souporcell and Freemuxlet, precision and recall were both 0.”

The proportion of cardiomyocytes in each donor increased after removal of misannotated cell types.

“After the removal of misannotated cell types, non-cardiomyocyte nuclei were absent from the cardiomyocyte cluster (Fig. 8J) and the proportion of cardiomyocytes in each donor was increased (Fig. 8H, K).”

After doublet removal, Azimuth projection identified 13 cell types, mainly cardiomyocytes, fibroblasts, and endothelial cells.

“Projection of the data after doublet removal onto a reference again, using the Azimuth package, identified 13 cell types, which mainly comprised cardiomyocytes, fibroblasts, and endothelial cells (Fig. 8J, 8K).”

Doublets including heterogenic and heterotypic were identified for human and sheep datasets using demultiplexing information and DoubletFinder.

“After removing low-quality nuclei, doublets including heterogenic and heterotypic were identified for the human (Fig. 9D, E) and sheep (Fig. 9F, G) datasets using the demultiplexing information and DoubletFinder, respectively.”

Vireo identified 1097 additional doublets (10.12%) compared to 813 heterotypic doublets identified by DoubletFinder.

“Illustrating the added advantage of sample pooling and demultiplexing in doublet identification, 1097 additional doublets were identified by Vireo (10.12%) to 813 heterotypic doublets identified by DoubletFinder (Fig. 8D).”

For pipelines relying on SNPs extracted from bulk RNA-Seq, variation in mean precision and recall was high due to some failed scenarios.

“For the pipelines relying on SNPs extracted from bulk-RNA-Seq, due to some failed scenarios, which unsuccessfully assigned the cells to their original samples, the variation in their mean precision and recall was high (Fig. 6C, D).”

For bulk RNA-Seq-based demultiplexing, BCFtools combined with Vireo or Freemuxlet were among the fastest pipelines, taking less than 2 hours to generate output.

“For demultiplexing using SNPs from bulk RNA-Seq, BCFtools combined with Vireo or Freemuxlet were two of the fastest pipelines, taking less than 2 hours to generate output (Fig. 5B).”

Cell types absent in sheep included adipocytes, mesothelial cells, mast cells, and B cells.

“Cell types absent in the sheep included adipocytes, mesothelial cells, mast cells, and B cells.”

The relative proportions of these cell types varied among samples.

“While all cell types were identified in each sample, reflecting variability in the human population, the relative proportions of these cell types varied.”

Vireo had a higher run speed than Souporcell.

“While these tools were scored similarly for most criteria, Vireo had a higher run speed than Souporcell.”

Tools with the highest demultiplexing precision (i.e., Vireo) showed the least sensitivity to sample number and doublet percentage.

“We generally observed that the tools (i.e., Vireo) with the highest demultiplexing precision showed the least sensitivity to sample number and doublet percentage (Fig. 6A, B).”

All methods that used genetic variants from the SNP array had run times that scaled linearly with the number of samples.

“All methods that used genetic variants from the SNP array had run times that scaled linearly with the number of samples (Fig. 6E).”

After doublet removal, nuclei from each sample were present in each of the clusters.

“After doublet removal, the data were re-clustered and clusters were annotated according to cell types. Nuclei from each sample were present in each of the clusters (Fig. 8I).”

cellSNP combined with Vireo or Souporcell took the longest time to demultiplex.

“cellSNP combined with Vireo or Souporcell took the longest time to demultiplex.”

The run time of scSplit increased more rapidly with sample number than for the other methods.

“Notably, the run time of scSplit increased more rapidly with sample number than for that of the other methods.”

548 nuclei were assigned as heterogenic doublets.

“The remainder of the 548 nuclei were assigned as heterogenic doublets.”

When doublets were retained, different cell types annotated by Azimuth mapping onto a reference were clustered together.

“When the doublets were retained in the dataset, the different cell types annotated by using Azimuth mapping onto a reference were clustered together (Fig. 8F, G).”

After removal of misannotated cell types, non-cardiomyocyte nuclei were absent from the cardiomyocyte cluster.

“After the removal of misannotated cell types, non-cardiomyocyte nuclei were absent from the cardiomyocyte cluster (Fig. 8J) and the proportion of cardiomyocytes in each donor was increased (Fig. 8H, K).”

scSplit generated the greatest number of SNPs from RNA-Seq data among the variant calling tools.

“Moreover, scSplit, which had the worst performance when included in the demultiplexing pipeline, generated the greatest number of SNPs from RNA-Seq data of the variant calling tools (Fig. S3 available online at http://bib.oxfordjournals.org/).”

All cell types were identified in each sample.

“While all cell types were identified in each sample, reflecting variability in the human population, the relative proportions of these cell types varied.”

The species-mixed dataset was successfully demultiplexed when mapping to the genome based on the proportion of reads of each origin.

“Based on the proportion of reads of each origin, the species-mixed dataset was successfully demultiplexed when mapping to the genome.”

Vireo achieved the highest usability score among the four demultiplexing methods.

“Among the four methods, Vireo achieved the highest usability score owing to its excellent implementation.”

For both sheep and human, cardiomyocytes, endothelial cells, and fibroblasts were the most abundant cell types, making up 75% of total cells between them.

“For both sheep and human, cardiomyocytes, endothelial cells, and fibroblasts were the most abundant cell types, making up 75% of the total cells between them.”

Vireo combined with BCFtools was the most computationally efficient method in the bulk RNA-Seq-based pipeline.

“Of note, Vireo combined BCFtools was the computational efficiency method in bulk RNA-Seq-based pipeline.”

The cardiomyocyte population included cells labeled as neuronal cells and fibroblasts when doublets were retained.

“We noticed the cardiomyocyte population also included cells that were labeled as neuronal cells and fibroblasts (Fig. 8G).”

The doublet proportion in each sample in the pool can be as high as 32.5%.

“The doublet proportion in each sample in the pool can be as high as 32.5% (Fig. 8E).”

scSplit was the worst performing tool, inferior across all measures than the other methods.

“scSplit was the worst performing tool being inferior across all measures than the other methods.”

Pipelines that included cellSNP data had the lowest accuracy.

“Although the concordance of SNPs between the variant caller and array were highest with cellSNP regardless of total SNP or genic SNP, pipelines that included cellSNP data had the lowest accuracy.”

Concordance of SNPs between variant caller and array was highest with cellSNP regardless of total SNP or genic SNP.

“Although the concordance of SNPs between the variant caller and array were highest with cellSNP regardless of total SNP or genic SNP, pipelines that included cellSNP data had the lowest accuracy.”

Kallisto-bustools can successfully demultiplex cells from different species based on proportion of UMI.

“By providing a pipeline applied to the human-sheep dataset, we showed that kallisto-bustools can successfully demultiplex the cells from different species based on the proportion of UMI.”

Removal of heterogenic doublets formed by different species improved performance of the analysis.

“The removal of heterogenic doublets formed by different species also improved the performance of the analysis.”

Overlap of SNPs identified by each variant caller and SNP array varied substantially.

“The overlap of SNPs identified in the RNA-Seq data by each variant caller and those identified by SNP array varied substantially.”

Demultiplexing performance was significantly affected by the variant caller irrespective of demultiplexing tool.

“Demultiplexing performance was however significantly affected by the variant caller irrespective of the demultiplexing tool used.”

Precision of nuclear identification for all methods was high, average 85%.

“Importantly, the precision of nuclear identification for all methods was high, achieving an average value of 85%, which indicates all methods have the capacity to identify doublets and singlets in real-world datasets.”

Computing run time varied substantially between pipelines, ranging from 2 hours for the most rapid to 12 hours for the slowest.

“This varied substantially between pipelines ranging from 2 hours for the most rapid pipeline to 12 hours for the slowest.”

Effect of doublet proportion was greater than increased sample number within the pool.

“This effect of doublet proportion was greater than increased sample number within the pool.”

Removing both types of doublets improves cell type annotation.

“By removing both types of doublets, the cell type annotation can be improved, which was shown in our application dataset.”

Run time scaled quite linearly with sample number in the pool.

“Run time also scaled quite linearly with the sample number in the pool to be demultiplexed, again supporting this approach.”

Souporcell had good performance, particularly with SNP array SNPs.

“Souporcell also had a good performance, particularly when using SNPs from the SNP Array.”

Performance of each tool/pipeline decreased with increasing doublet numbers.

“Notably, the performance of each tool/pipeline decreased with increasing doublet numbers.”

Appending variant caller onto each demultiplexing tool did not substantially add to run time.

“While it was expected that the appending of the variant caller onto each demultiplexing tool would substantially add to this run time, thus making this approach less attractive, this did not appear to be the case.”

Regardless of SNP source, Vireo performed best in accuracy (recall and precision).

“Regardless of whether SNPs were obtained from RNA-Seq or SNP arrays, Vireo performed the best in terms of accuracy (recall and precision).”

Freemuxlet performance was particularly poor when combined with variant calling tools.

“The performance of Freemuxlet was particularly poor when combined with variant calling tools.”

scSplit requires significant memory (~100 GB), especially with >10k nuclei.

“This long run time and the significant memory (~100 GB) it requires for use, especially with a higher number of nuclei (more than 10 k nuclei), would suggest that it should be avoided.”

Vireo and Souporcell showed superior performance on sample demultiplexing using genetic information obtained from SNP arrays.

“Together, these analyses of demultiplexing precision and recall over 2–6 samples and on real-world data indicate a superior performance for Vireo and Souporcell on sample demultiplexing using genetic information obtained from SNP arrays.”

The number of variants called by BCFtools was insufficient to discriminate between samples in the six sample dataset.

“The number of variants called by BCFtools was insufficient to discriminate the difference between samples in the six samples dataset.”

Variant number is important for the capacity of the demultiplexing tool to perform its function.

“Indeed, the strong correlation between the number of variants detected by the calling tools with demultiplexing performance supports the importance of variant number in the capacity of the demultiplexing tool to perform its function.”

Some samples in the six sample dataset were assigned to the wrong donor.

“Therefore, some samples in the dataset were assigned to the wrong donor.”

The performance of the demultiplexing method is more sensitive to doublet number than to the number of samples multiplexed.

“We observed only subtle differences between median precision values for each method when applied to four and six samples, suggesting the performance of the demultiplexing method is more sensitive to doublet number rather than the number of samples multiplexed (Fig. 2A, B).”

The four tissue samples analysed are similar.

“This not only indicates the similarity of the four tissue samples analysed but also highlights the advantage of sample pooling prior to nuclear isolation, which minimizes the risk of batch effects.”

Through introducing additional genetically distinct nuclei by sample pooling prior to GEM formation, the proportion of heterogenic doublets increases proportionately with the number of samples pooled.

“Through introducing additional genetically distinct nuclei into the sequencing pipeline by sample pooling prior to GEM formation, the proportion of doublets that are heterogenic in composition is increased proportionately with increased number of samples pooled (from genetically different samples).”

The optimal pipeline was scalable.

“Further, this pipeline was computationally efficient, statistically robust, and was scalable.”

Apart from scSplit, each demultiplexing method performed well across all evaluation criteria relative to the best-case scenario with species or sex differences.

“Relative to the best-case scenario with species or sex differences, we found that apart from scSplit, each demultiplexing method performed well across all evaluation criteria.”

The optimal pipeline was statistically robust.

“Further, this pipeline was computationally efficient, statistically robust, and was scalable.”

Vireo and Souporcell are the top two methods in the benchmark.

“This analysis is summarized in Fig. 7 and shows Vireo and Souporcell to be the top two methods.”

Sample pooling prior to nuclear isolation minimizes the risk of batch effects.

“This not only indicates the similarity of the four tissue samples analysed but also highlights the advantage of sample pooling prior to nuclear isolation, which minimizes the risk of batch effects.”

In the absence of original samples, single-cell data can be demultiplexed into anonymous individual contributors using the demultiplexing tools employed.

“In the absence of these original samples, single-cell data can be demultiplexed into anonymous individual contributors using the demultiplexing tools that we employed.”

Vireo is the best overall method for demultiplexing.

“Our analysis established Vireo as the best overall method for demultiplexing (Fig. 7).”

The optimal pipeline was computationally efficient.

“Further, this pipeline was computationally efficient, statistically robust, and was scalable.”

The optimal pipeline provides high precision and recall for demultiplexing up to six samples over doublet percentage ranging from 0-30%.

“Here, through benchmarking pipelines for demultiplexing of 10x snRNA-Seq data, we identify an optimal pipeline that provides high precision and recall for demultiplexing up to six samples over doublet percentage ranging from 0–30%.”

Real-world datasets present biological doublets and ambient RNAs, which are lacking in simulated dataset.

“The advantage of the use of real-world datasets is the challenge offered by the presence of biological doublets and ambient RNAs, which are lacking in the simulated dataset.”

All demultiplexing tools achieved high precision and recall.

“All demultiplexing tools achieved high precision and recall, with Vireo combined with genetic information from SNP array or matched RNA-Seq exhibiting the highest accuracy.”

The simulated dataset did not take heterotypic doublets from different cell types into account.

“We also did not take the heterotypic doublets that are from different cell types into account in our simulated dataset.”

Doublet removal is a key step in bioinformatic pipelines for snRNA-Seq data.

“Doublet removal is thus a key step in bioinformatic pipelines for snRNA-Seq data.”

All methods have capacity to identify doublets and singlets in real-world datasets.

“Importantly, the precision of nuclear identification for all methods was high, achieving an average value of 85%, which indicates all methods have the capacity to identify doublets and singlets in real-world datasets.”

Using expressed variants in RNA-Seq for demultiplexing requires additional bioinformatical steps and computing power involving variant calling tools.

“When using expressed variants in RNA-Seq for demultiplexing, additional bioinformatical steps and computing power are required, involving variant calling tools.”

Cell doublets may be identified if genetically distinct using demultiplexing tools.

“Cell doublets may also be identified if they are genetically distinct using demultiplexing tools.”

Vireo combined with genetic information from SNP array or matched RNA-Seq exhibited the highest accuracy.

“All demultiplexing tools achieved high precision and recall, with Vireo combined with genetic information from SNP array or matched RNA-Seq exhibiting the highest accuracy.”

In certain situations, such as highly synthetic or polyploid cells, nuclei are more likely to be inappropriately removed.

“Since doublet removal tools generally remove doublets based on the relative expression of transcripts compared to an expected distribution, in certain situations, such as for highly synthetic or polyploid cells, nuclei are more likely to be inappropriately removed.”