Evaluation of genetic demultiplexing of single-cell sequencing data from model species

DBA/1J mice have ∼5 × 10^6 total SNPs and indels.

“We also attempted SNP-based demultiplexing of in silico–pooled data from DBA/1J mice (Shinozaki et al, 2022) which have ∼5 × 106 total SNPs and indels (Doran et al, 2016).”

• cites Doran et al, 2016 sourceDisputed

  Table 1 in “Identification of SNPs and indels” reports 5,129,920 SNPs and 955,621 indels for DBA/1J. That is approximately 6.1 million combined; the cited figure of approximately 5 million counts the SNPs alone. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

A recent study collected single animal scRNA-seq datasets from the thymus of zebrafish.

“A recent study collected single animal scRNA-seq datasets from the thymus of zebrafish (Rubin et al, 2022).”

• cites Rubin et al, 2022 sourceSupported

The lack of biological replicates is a major cause of false discoveries in scRNA-seq.

“This is a serious drawback as the lack of biological replicates has also been shown to be a major cause of false discoveries (Hicks et al, 2018; Squair et al, 2021).”

• cites Hicks et al, 2018 sourceNot checked — no full text of the source
• cites Squair et al, 2021 sourceSupported

A published dataset of Xenopus laevis scRNA-seq data contains eight experimentally pooled samples from three Xenopus transgenic lines that each overexpress a different fluorescent gene.

“We analyzed a published dataset of Xenopus laevis scRNA-seq data containing eight experimentally pooled samples from three Xenopus transgenic lines that each overexpress a different fluorescent gene (Lin et al, 2021) (Fig 2A).”

• cites Lin et al, 2021 sourcePartly supported

  The STAR★Methods (Method details; page unavailable) describe Xenopus scRNA-seq data with two three-sample blastema pools and a two-sample limb-bud/mature-tissue pool, accounting for eight pooled sample preparations. They say the three animal batches in the blastema pools each expressed a different fluorescent color signal, but do not establish that all eight samples came from only three transgenic lines or that each line overexpressed a different fluorescent gene. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

SNP-based demultiplexing was unsuccessful on an in silico-pooled mixture of three C57BL/6 mice samples, with 99.7% of cells being unassigned by souporcell or Vireo.

“Our efforts to apply SNP-based demultiplexing to an in silico–pooled mixture of three C57BL/6 mice samples (Crowl et al, 2022) were unsuccessful, with 99.7% of cells being unassigned by souporcell or Vireo (data not shown).”

• cites Crowl et al, 2022 sourceSupported

Low detection of fluorescent gene transcripts is a common problem when using fluorescent marker genes for demuxing pooled data.

“This low detection of fluorescent gene transcripts is a common problem when using fluorescent marker genes for demuxing pooled data (Lin et al, 2021) and underscores the importance of alternative means to demultiplex pooled data.”

• cites Lin et al, 2021 sourcePartly supported

  The STAR★Methods section, under “Identifying samples in pooled scRNA-seq,” reports that weak expression of some fluorescent markers caused poor detection and hampered sample assignment in this study. It does not establish that low detection is a common problem across pooled datasets. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

C57BL/6 mice are nearly genetically identical with only ∼15,000 SNPs and insertion-deletion mutations (indels).

“C57BL/6 mice are nearly genetically identical with only ∼15,000 SNPs and insertion–deletion mutations (indels) (Doran et al, 2016), providing evidence that SNP-based demuxers are unable to separate biological replicates from highly inbred mice.”

• cites Doran et al, 2016 sourcePartly supported

  The Results and discussion section (‘Novel and private variation’; page unavailable) describes very little private variation in C57BL/10J compared with the C57BL/6J reference, supporting the general point that these related strains are genetically similar. But ‘Identification of SNPs and indels’ reports 349,702 SNPs and 83,930 indels for that comparison, not approximately 15,000, and does not establish that figure for C57BL/6 mice or biological replicates. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

SNP-based demultiplexers are unable to separate samples within the same mouse strain.

“These findings are in line with other claims that SNP-based demultiplexers are unable to separate samples within the same mouse strain (Mylka et al, 2022).”

• cites Mylka et al, 2022 sourcePartly supported

  The Discussion says the SNP-based tool freemuxlet fails to separate C57BL6N from C57BL6J, two closely related but distinct mouse strains. It does not directly report testing samples within the same strain, and its observation about freemuxlet does not establish that all SNP-based demultiplexers are unable to separate them. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Animal-to-animal variability in gene expression patterns has been observed for decades and is expected in many tissues, especially in immune cells, of animals that encounter different stressors or pathogens.

“Animal-to-animal variability in gene expression patterns has been observed for decades and is expected in many tissues, especially in immune cells, of animals that encounter different stressors or pathogens (Boedigheimer et al, 2008; Schokker et al, 2015; Paunovska et al, 2018; Reid et al, 2018).”

• cites Boedigheimer et al, 2008 sourcePartly supported

  The Results (Variability analysis; page unavailable) report gene-expression differences between control animals, and the Background notes earlier studies of individual-animal variability. But this paper examines rat liver and kidney, not immune cells, and the cited passages do not establish that such variability has been observed for decades or link it to animals encountering different stressors or pathogens. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites Schokker et al, 2015 sourcePartly supported

  The Results (Transcriptomic analyses; page not provided) report substantial variation within treatment groups in intestinal-tissue gene expression in pigs. The Discussion (Comparison of gene expression in intestinal tissues between treatment groups; page not provided) also reports treatment-related differences in immune processes. The paper does not establish that such variability has been observed for decades, occurs in many tissues or specifically in immune cells, or is expected after pathogen exposure. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites Paunovska et al, 2018 sourceCould not verify

  The Introduction discusses changes in gene expression when cells are cultured, not variation between animals. The Animal Experiments section describes the mice used, but neither it nor the searched passages addresses whether stressors or pathogens cause animal-to-animal differences in gene expression, particularly in immune cells. No page numbers were provided in our copy. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

A thorough benchmarking in human samples of transcriptional and SNP-based doublet detectors suggest that an intersectional approach to doublet detection is superior to any one single method.

“In line with this, a thorough benchmarking in human samples of transcriptional and SNP-based doublet detectors suggest that an intersectional approach to doublet detection is superior to any one single method (Neavin et al, 2022 Preprint).”

• cites Neavin et al, 2022 Preprint sourceSupported

Multiple independent programs designed specifically for doublet detection using transcriptomic instead of genotypic information are available for scRNA-seq data including DoubletDetection, Solo, and Scds.

“Multiple independent programs designed specifically for doublet detection using transcriptomic instead of genotypic information are available for scRNA-seq data including DoubletDetection (Gayoso & Shor, 2022), Solo (Bernstein et al, 2020), and Scds (Bais & Kostka, 2020).”

• cites Gayoso & Shor, 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Bernstein et al, 2020 sourceSupported
• cites Bais & Kostka, 2020 sourceSupported

SNP-based demuxers can successfully demultiplex single-nuclei RNA-seq and scRNA-seq datasets without any prior information about the samples and with only a de novo transcriptome as a reference.

“We found that SNP-based demuxers can successfully demultiplex single-nuclei RNA-seq and scRNA-seq datasets without any prior information about the samples and with only a de novo transcriptome as a reference.”

Most available SNP-based demultiplexing programs are highly accurate on model organism datasets.

“We benchmarked available SNP-based demultiplexing programs and found that most are highly accurate on model organism datasets.”

Both inbred mouse strains had less than 0.2 SNPs/kilobase (0.013 for C57BL/6 and 0.19 for DBA1/J).

“We found that both inbred mouse strains had less than 0.2 SNPs/kilobase (kb) (0.013 and 0.19 for C57BL/6 and DBA1/J, respectively).”

Souporcell failed to identify a large proportion of the synthetic heterotypic doublets in the axolotl dataset.

“Similar to the three sample zebrafish datasets, we found that souporcell failed to identify a large proportion of the synthetic heterotypic doublets (Fig S5C).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

Compared with demultiplexing of three zebrafish samples, the doublet detection observed with the pool of five monkey samples was improved (82% versus 55%).

“Compared with demultiplexing of three zebrafish samples, the doublet detection observed with this pool of five monkey samples was improved (82% versus 55%) (Fig 1E and F).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

95-99% of the cells from each individual axolotl dataset were correctly demultiplexed from the pool.

“95–99% of the cells from each individual axolotl dataset were correctly demultiplexed from the pool (Fig S5B and C).”

SNP-based demultiplexing of the synthetically pooled zebrafish scRNA-seq data was highly accurate, resulting in correct cell assignments of 99-100% of cells based on their known animal origin.

“SNP-based demultiplexing of this synthetically pooled zebrafish scRNA-seq data was highly accurate, resulting in correct cell assignments of 99–100% of cells based on their known animal origin (Fig 1C).”

Freemuxlet and Vireo misassigned many of the same doublets as souporcell in the zebrafish dataset.

“Souporcell was not alone in its poor doublet detection performance, with Freemuxlet and Vireo misassigning many of the same doublets as souporcell (Fig S1B).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

Results from demultiplexing the five African green monkey samples closely mirrored the zebrafish dataset in terms of accuracy and similarity of results from independent SNP-based demultiplexing programs.

“Our results from this demultiplexing closely mirrored that of the zebrafish dataset in terms of accuracy and the similarity of results from independent SNP-based demultiplexing programs (Fig S3).”

Souporcell only called 66% of true heterotypic doublets from a pool of three monkey samples.

“We found that souporcell only called 66% of true heterotypic doublets from this pool of three monkey samples (Fig S4).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

Souporcell missed almost half of the synthetic heterotypic true doublets in the pooled zebrafish dataset.

“We found that souporcell missed almost half of the synthetic heterotypic true doublets in the pooled dataset (Fig 1C).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

There is a strong agreement between souporcell assignments and ground truth animal identities for the synthetically pooled zebrafish data.

“We found a strong agreement between souporcell assignments and ground truth animal identities (Fig 1B and C).”

99.7% of cells from in silico-pooled DBA/1J mice data were unable to be assigned via souporcell.

“Again, 99.7% of cells were unable to be assigned via souporcell (data not shown), suggesting that SNP-based demultiplexing is incompatible with highly inbred animals.”

For the zebrafish dataset, the use of the common VCF file had minimal overall effect on the souporcell demultiplexing results, with less than 1% of cell assignments differing (11/1,312 cells).

“We found that for this zebrafish dataset, the use of the common VCF file had minimal overall effect on the souporcell demultiplexing results, with less than 1% of cell assignments differing (11/1,312 cells).”

Many cells with high CMO reads per cell have relatively low biological reads per cell.

“One important note, it appeared that many cells with high CMO reads per cell have relatively low biological reads per cell (Fig 4B), suggesting that although these cells can be CMO demultiplexed they would be unlikely to provide adequate biological information.”

Doublet assignments were the biggest discrepancy in the Xenopus experiment.

“As with the in silico experiment, doublet assignments were the biggest discrepancy.”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

Fluorescent-based and souporcell assignments show a high degree of correlation in the Pleurodeles dataset.

“We found that fluorescent-based and souporcell assignments show a high degree of correlation (Fig 3C and D).”

There was a remarkable similarity in demultiplexing the eight-animal Xenopus dataset with fluorescent and SNP-based methods.

“We observed a remarkable similarity in demultiplexing the eight-animal dataset with these two orthologous methods (Fig 2C–E).”

When the two methods disagree, the most prevalent occurrence is "negative" fluorescent-based cell assignments that souporcell assigned to one of the transgenic animals.

“When the two methods disagree, the most prevalent occurrence is “negative” fluorescent–based cell assignments that souporcell assigned to one of the transgenic animals (i.e., the rescue we also found in the Xenopus data).”

The dataset of a pool of 30 zebrafish embryos contains 25,279 cells.

“We found this dataset to contain 25,279 cells, which according to the manufacturer would result in ∼20% of droplets being doublets.”

There is a high degree of agreement between fluorescent-based cell assignments and SNP-based assignments from Pleurodeles scRNA-seq data.

“We found a high degree of agreement between fluorescent-based cell assignments and SNP-based assignments from Pleurodeles scRNA-seq data (Fig 3E).”

Of cells assigned by fluorescent-based demuxing to one of the three transgenic animals, 78-93% were correctly identified by souporcell.

“Of cells assigned by fluorescent-based demuxing to one of the three transgenic animals, 78–93% of those cells were correctly identified by souporcell (Fig 3F).”

There was high agreement between souporcell and fluorescent-based assignments, with souporcell assignment agreement on 86-92% of cells in each sample.

“We found high agreement between the two methods, as demonstrated by the souporcell assignment agreement on 86–92% of cells in each sample (Fig 2F).”

Cells with more biological read depth showed improved percent agreement between souporcell- and Cellplex-based assignments.

“Interestingly, we did see that cells with more biological read depth showed improved percent agreement between souporcell- and Cellplex-based assignments, which was not a given because the Cellplex label is a distinct library (Fig S10).”

At the cluster level, there was wide variation in assigned doublets ranging from 0.8 to 30.2% of cells within a cluster.

“At the cluster level, there was wide variation in assigned doublets ranging from 0.8 to 30.2% of cells within a cluster (Fig S7B).”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

The next most common discrepancy between the two demultiplexing approaches were contested doublets, cells labeled as "doublet" by fluorescent-based demuxing, for which souporcell disagrees.

“The next most common discrepancy between the two demultiplexing approaches were contested doublets, cells labeled as “doublet” by fluorescent-based demuxing, for which souporcell disagrees.”

In the Pleurodeles dataset, cells sequenced to a higher depth showed higher agreement between souporcell cell assignments and the fluorescent-based ground truth.

“In this dataset, we observed that cells sequenced to a higher depth showed higher agreement between souporcell cell assignments and the fluorescent-based ground truth, which is expected because both methods should improve with more information (Fig S8I).”

Many cells assigned as "negative" by the fluorescent-based approach could be computationally rescued and assigned a transgenic line by souporcell.

“Furthermore, many cells were assigned as “negative” by the fluorescent-based approach, which could then be computationally rescued and assigned a transgenic line by souporcell.”

souporcell sample assignments correlated tightly with CMO-based assignments.

“We found that souporcell sample assignments correlated tightly with CMO-based assignments (Fig 4C–E).”

souporcell and scds agreed on 68% of assignments, including singlets and doublets.

“Overall, souporcell and scds agreed on 68% of assignments, including singlets and doublets.”

The fluorescent-based approach assigns many cells as "negative" (63% of cells pre-filter, 29% of cells after filtering) in the Pleurodeles dataset.

“The fluorescent-based approach assigns many cells as “negative” (63% of cells pre-filter, 29% of cells after filtering) (Figs 3C and S8H).”

When CMO-based assignments were interpreted as ground truth, the accuracy of souporcell assignments was clear, with 73-96% agreement between the two methods for assignment to the three CMO-based assignments.

“When CMO-based assignments were interpreted as ground truth, the accuracy of souporcell assignments was clear, with 73–96% agreement between the two methods for assignment to the three CMO-based assignments (Fig 4F).”

Souporcell gave assignments for 95.1% of cells (24,047 of 25,279) in the 30 zebrafish embryo dataset.

“We applied souporcell to this dataset and found that it gave assignments for 95.1% of cells (24,047 of 25,279).”

SNP-based demultiplexing and barnyard style analysis showed general agreement in the dual-species pooled salamander dataset.

“To investigate SNP-based assignments and doublets in this dataset, we applied souporcell and a barnyard style analysis to this dataset and found a general agreement between these two approaches (Fig S9A–G).”

The common assumption that each cluster is composed of cells from all biological replicates held true for nine of the identified 21 clusters in the 30 zebrafish sample, predominantly those with the most cells.

“This assumption held true for nine of the identified 21 clusters in this sample, predominantly those with the most cells.”

The largest difference in assignments between these methods was that souporcell identified many doublets in the data from one CMO-labeled sample.

“The largest difference in assignments between these methods was that souporcell identified many doublets in the data from one CMO-labeled sample.”

SNP-based demultiplexing on a pool of 30 zebrafish samples produced assignments.

“In line with this, we obtained assignments when performing SNP-based demultiplexing on a pool of 30 zebrafish samples.”

One of the eight transgenic animals was almost completely absent from the Xenopus dataset, only displaying dozens of total cells.

“Also of interest, we found that one of the eight transgenic animals was almost completely absent from the dataset, only displaying dozens of total cells.”

We demonstrated the successful SNP-based demuxing of a pooled two-species experiment.

“Finally, though pooling of multiple species into one experiment is not currently commonplace, we demonstrated the successful SNP-based demuxing of a pooled two-species experiment.”

We successfully applied SNP-based methods to demultiplex pooled single-cell data from multiple species and a two species mix.

“Overall, we successfully applied SNP-based methods to demultiplex pooled single-cell data from multiple species and a two species mix.”

The potential for heterogeneity in sample representation across cell clusters can be clearly seen in Figs 3 and S8, where some cell clusters are from one or two of the three individuals present in the pool.

“The potential for heterogeneity in sample representation across cell clusters can be clearly seen in Figs 3 and S8, where some cell clusters are from one or two of the three individuals present in the pool.”

Our results from synthetic pooled samples saw varied success in heterotypic doublet detection by SNP-based demultiplexers.

“However, our results from synthetic pooled samples saw varied success in heterotypic doublet detection by SNP-based demultiplexers.”

souporcell stood out as the only tool that was feasible to use with only a de novo transcriptome and no VCF input.

“This is where souporcell stood out as the only tool that was feasible to use with only a de novo transcriptome and no VCF input.”

• AssumptionRests on an assumption: Notophthalmus viridescens has only a de novo transcriptome and no published common SNP VCF file.

Freemuxlet was successful when inputting the VCF and minimap.bam generated via souporcell when running souporcell without a VCF.

“Important note: as with the zebrafish data, we found that Freemuxlet failed when using the downloaded VCF file but was successful when inputting the VCF and minimap.bam generated via souporcell when running souporcell without a VCF.”

Freemuxlet failed when using the downloaded VCF file.

“Important note: as with the zebrafish data, we found that Freemuxlet failed when using the downloaded VCF file but was successful when inputting the VCF and minimap.bam generated via souporcell when running souporcell without a VCF.”

In the authors' analysis, the Seurat MULTIseqDemux function works well with data from overexpressed fluorescent gene mRNAs.

“Although the MULTIseqDemux function is written to assign cell identities based on CMO labels, we found that it works well with data from overexpressed fluorescent gene mRNAs.”

SNP-based demuxers can facilitate effective experimental design via the demultiplexing of pooled scRNA-seq data from a vast range of species.

“This suggests that SNP-based demuxers can facilitate effective experimental design via the demultiplexing of pooled scRNA-seq data from a vast range of species.”

scRNA-seq sample pooling followed by genetics-based separation of individuals is an effective means to identify individual samples in a variety of commonly studied species.

“The authors demonstrate that scRNA-seq sample pooling followed by genetics-based separation of individuals is an effective means to identify individual samples in a variety of commonly studied species.”

• AssumptionRests on an assumption: Sample pooling is an effective means to increase biological replicate throughput while simultaneously decreasing batch effects and costs.

The individually sequenced libraries provide a ground truth for definitive benchmarking of SNP-based assignments.

“Although synthetically pooled data are less challenging for an SNP-based demultiplexing algorithm than experimentally pooled samples, the individually sequenced libraries provide a ground truth for definitive benchmarking of SNP-based assignments.”

SNP-based demuxers are unable to separate biological replicates from highly inbred mice.

“C57BL/6 mice are nearly genetically identical with only ∼15,000 SNPs and insertion–deletion mutations (indels) (Doran et al, 2016), providing evidence that SNP-based demuxers are unable to separate biological replicates from highly inbred mice.”

The in silico-pooled scRNA-seq setup lacks challenges of physically pooled scRNA-seq data, like ambient RNA and real heterotypic doublets.

“Our previous results with in silico–pooled scRNA-seq data indicated high accuracy of SNP-based demultiplexers, but this setup lacks challenges of physically pooled scRNA-seq data, like ambient RNA and real heterotypic doublets.”

SNP-demultiplexing tools struggle with heterotypic doublet detection when only three animals are pooled.

“This suggests that SNP-demultiplexing tools struggle with heterotypic doublet detection when only three animals are pooled.”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

In these in silico pools, true doublets can be identified with absolute certainty because we have origin information.

“In these in silico pools, true doublets can be identified with absolute certainty because we have origin information.”

Genetic demultiplexing is a viable means to enable sample pooling and subsequent demultiplexing in zebrafish.

“This suggests that genetic demultiplexing is a viable means to enable sample pooling and subsequent demultiplexing in zebrafish.”

SNP-based demultiplexing can be a highly accurate and efficient method for demultiplexing pooled single-cell data from non-human primates.

“Overall, these results indicate that SNP-based demultiplexing can be a highly accurate and efficient method for demultiplexing pooled single-cell data from non-human primates.”

The production of synthetic doublets bolsters the ability to test and validate the accuracy of cell and doublet assignments.

“The production of synthetic doublets bolsters the ability to test and validate the accuracy of cell and doublet assignments.”

The number of samples in a pool contributes to efficiency of heterotypic doublet detection.

“This suggests that the number of samples in a pool contributes to efficiency of heterotypic doublet detection.”

• AssumptionRests on an assumption: Souporcell relies on intergenotypic doublet detection, so homotypic true doublets can be removed.

SNP-based demultiplexing is incompatible with highly inbred animals.

“Again, 99.7% of cells were unable to be assigned via souporcell (data not shown), suggesting that SNP-based demultiplexing is incompatible with highly inbred animals.”

SNP-based demultiplexing may accurately demultiplex pooled samples in any single-cell compatible organism with genetic heterogeneity.

“These results suggest that SNP-based demultiplexing may accurately demultiplex pooled samples in any single-cell compatible organism with genetic heterogeneity.”

A range of 0.2-0.34 SNPs/kb may be the minimum required within a sc-seq dataset for SNP-based demultiplexing.

“Although an exact quantitative analysis of this possible genetic cutoff would be useful, these results imply that a range of 0.2–0.34 SNPs/kb may be the minimum required within a sc-seq dataset for SNP-based demultiplexing.”

Transcriptional-based doublet detection potentially identifies doublets missed by souporcell and wet laboratory-based approaches.

“This suggests that transcriptional-based doublet detection potentially identifies doublets missed by souporcell and wet laboratory–based approaches.”

SNP-based demultiplexing is a powerful and near universal method to demultiplex pooled samples.

“Altogether, these data suggest that SNP-based demultiplexing is a powerful and near universal method to demultiplex pooled samples.”

Sample pooling is a useful and viable tool for empowering single-cell datasets.

“We demonstrate that sample pooling is a useful and viable tool for empowering single-cell datasets.”

The fluorescence-based "negative" assignments are attributed to the low capture of fluorescent gene reads.

“Similar to the Xenopus dataset, we attribute these fluorescence-based “negative” assignments to the low capture of fluorescent gene reads (Figs 3B and C and S8A and C).”

SNP-based demultiplexing produces highly accurate demuxing results in numerous species with varying genomic resources and genetic variability.

“We show with numerous species which have varying genomic resources and genetic variability that SNP-based demultiplexing produces highly accurate demuxing results.”

Cells with high CMO reads per cell but low biological reads per cell would be unlikely to provide adequate biological information.

“One important note, it appeared that many cells with high CMO reads per cell have relatively low biological reads per cell (Fig 4B), suggesting that although these cells can be CMO demultiplexed they would be unlikely to provide adequate biological information.”

SNP-based assignments display a high degree of accuracy in a complex, experimentally pooled mixture of Xenopus cells.

“Overall, these results display a high degree of accuracy for SNP-based assignments in a complex, experimentally pooled mixture of Xenopus cells.”

• AssumptionRests on an assumption: Fluorescent-based assignments provide a ground truth for benchmarking souporcell cell assignments.

The heterogeneity of sample representation in different cell clusters highlights the critical need for demultiplexing of pooled scRNA-seq data.

“The heterogeneity of sample representation in different cell clusters highlights the critical need for demultiplexing of pooled scRNA-seq data (Fig S8A–C, J, and K).”

Animal dropout in the Xenopus dataset would not be identified without demultiplexing.

“Although individual animal assignments by souporcell could not be validated, this suggests animal dropout in this dataset that would not be identified without demultiplexing.”

SNP-based demuxers display a performance relationship with increased read depth.

“This suggests that SNP-based demuxers do display a performance relationship with increased read depth.”

Without demultiplexing, erroneous conclusions on novel cell states or types may arise.

“Without demultiplexing, erroneous conclusions on novel cell states or types may arise.”

This analysis of the 30 zebrafish dataset could not be validated because the dataset did not contain a distinct cell-labeling method to enable orthogonal non-SNP-based demultiplexing and thereby a ground truth for comparison.

“This analysis could not be validated because this dataset did not contain a distinct cell-labeling method to enable orthogonal non-SNP–based demultiplexing and thereby a ground truth for comparison.”

The Pleurodeles experiment was designed to only contain non-erythroid spleen cells from one individual of each transgenic newt line, making it technically feasible to benchmark souporcell cell assignments for individuals through comparisons to fluorescent-based assignments.

“We designed this experiment to only contain non-erythroid spleen cells from one individual of each transgenic newt line, making it technically feasible to benchmark souporcell cell assignments for individuals through comparisons to fluorescent-based assignments (Fig 3A).”

The low number of mapped fluorescent reads in these samples means that a "doublet" assignment by fluorescent-based demuxing likely indicates that a particular cell had one to two counts of two distinct fluorescent genes.

“The low number of mapped fluorescent reads in these samples means that a “doublet” assignment by fluorescent-based demuxing likely indicates that a particular cell had one to two counts of two distinct fluorescent genes.”

These analyses demonstrate high agreement between CMO-based and SNP-based demultiplexing of this data from a notably complex experiment using species with poor genomic resources.

“In general, these analyses demonstrate high agreement between CMO-based and SNP-based demultiplexing of this data from a notably complex experiment using species with poor genomic resources.”

The 30 zebrafish embryo dataset would be considered a superloaded sample as it is over the supported cell recovery recommendations.

“This would be considered a superloaded sample as it is over the supported cell recovery recommendations.”

Species that do not possess a VCF file can still do the in silico ground truth benchmarking performed in this study.

“This means that species that do not possess a VCF file can still do the in silico ground truth benchmarking we performed in this study.”

SNP demuxing can be performed without resources like a high quality, well annotated genome, or a population-wide common SNP genotypes VCF file.

“Fortunately, we demonstrate that SNP demuxing can be performed without resources like a high quality, well annotated genome, or a population-wide common SNP genotypes VCF file.”

• AssumptionRests on an assumption: Notophthalmus viridescens has only a de novo transcriptome and no published common SNP VCF file.

The existence of natural biological variability coupled with technical variability in single-cell methods poses problems for studies without replicates or those that do not demultiplex pooled datasets.

“The existence of this natural biological variability coupled with technical variability in single-cell methods poses problems for studies without replicates or those that do not demultiplex pooled datasets (Hicks et al, 2018).”

• cites Hicks et al, 2018 sourceNot checked — no full text of the source

This experiment demonstrates the success of SNP-based demuxing in a particularly challenging scenario.

“Despite these potential challenges, this experiment demonstrates the success of SNP-based demuxing in a particularly challenging scenario, and as more research focuses on cross-species comparisons, this is a clever means to increase sample throughput.”

SNP-based demultiplexing in these species is accurate relative to other available demultiplexing approaches.

“Our benchmarking results suggest that SNP-based demultiplexing in these species is accurate relative to other available demultiplexing approaches.”

Sample number in each pool may directly impact doublet-detection efficiency.

“Our analyses also indicate that sample number in each pool may directly impact doublet-detection efficiency, which should be considered in experimental design.”