Transcriptome-wide analysis of differential expression in perturbation atlases

May combine two claims.

TI estimates were robust with slight downward bias.

“Estimates of the transcriptome-wide impact (Fig. 1C) were robust, with slight downward bias.”

May combine two claims.

Downward bias in TI results from inadequate power to detect very small effects at finite sample size; bias disappears at increased sample size.

“The downward bias results from inadequate power to detect very small effects at finite sample size, even aggregating power across genes; at increased sample size, the bias disappears (Supplementary Fig. 1).”

May combine two claims.

TI correlations were weaker between K562 and HepG2 (0.64), Jurkat and HepG2 (0.69), and substantially weaker between K562 and RPE1 (0.40) and Jurkat and RPE1 (0.56).

“There were weaker correlations between K562 and HepG2 (median: 0.64) and between Jurkat and HepG2 (median: 0.69), and substantially weaker correlations between K562 and RPE1 (0.40) and Jurkat and RPE1 (0.56).”

May combine two claims.

Fraction of TI explained by FDR-significant effects: K562-GenomeWide 36%, K562-Essential 18%, RPE1-Essential 35%, Jurkat-Essential 13%, HepG2-Essential 14%.

“In the K562-GenomeWide experiment, 36% of transcriptome-wide impact was explained by FDR-significant effects (Fig. 2C); in K562-Essential, 18%; in RPE1-Essential, 35%; in Jurkat-Essential: 13%; and in HepG2-Essential, 14%.”

May combine two claims.

In downsampling K562-GenomeWide to 50% gemgroups, signal in significant genes decreased substantially but TI estimate decreased only slightly.

“We conducted a downsampling analysis, repeating our analysis of the K562-GenomeWide experiment using only 50% of the 10x Genomics gemgroups. Whereas the signal in significant genes decreased substantially, our estimate of the transcriptome-wide impact decreased only slightly (Fig. 2C), consistent with the slight downward biased observed in simulations (Supplementary Fig. 1).”

May combine two claims.

TRADE estimates GATA1 knockdown has large TI (0.4, p=9e-12) while EIF4A4 knockdown lacks evidence of broad TI (0.004, p=0.12).

“However, TRADE accounts for the difference in sample size and estimates that whereas GATA1 knockdown has a large transcriptional effect (transcriptome-wide impact = 0.4, p = 9e-12), EIF4A4 knockdown lacks evidence of broad transcriptome-wide effects (transcriptome-wide impact = 0.004, p = 0.12) (Fig. 2A,B).”

May combine two claims.

Genes highly expressed in K562 are enriched for both perturbation impact (2.3x) and perturbation response (4.4x).

“In contrast, genes that are highly expressed in K562 cells are strongly enriched for both perturbation impact (2.3x) and perturbation response (4.4x), supporting a correlation between absolute expression and functional importance.”

May combine two claims.

Weak-to-moderate perturbations of Sox9 and Polycomb caused relatively small TI, whereas strong perturbations caused disproportionately large TI.

“Weak-to-moderate perturbations of these proteins caused relatively small transcriptome-wide effects, whereas strong perturbations caused disproportionately large transcriptome-wide effects.”

May combine two claims.

Some identified cell-type-specific perturbations are known indispensable for that cell type (GATA1 for K562, HMGCR for Jurkat), but most had no previously documented explanation.

“Some of these perturbations are known to be indispensable for their corresponding cell type, including GATA1 for erythroid cells such as K56221 and HMGCR for T-cells such as Jurkat22, but most had no previously documented explanation for their cell-type dependent effects.”

• cites [21] Weiss, M. J., Keller, G. & Orkin, S. H. Novel insights into erythroid development revealed through in vitro differentiation of GATA-1 embryonic stem cells. Genes Dev. 8, 1184–1197 (1994). sourceNot checked yet

  Not checked in the free run.

• cites [22] Lacher, S. M. et al. HMG-CoA reductase promotes protein prenylation and therefore is indispensible for T-cell survival. Cell Death Dis. 8, e2824 (2017). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A few perturbations had low between-experiment TI correlations; most had very low TI, so their correlations are expected to be noisy.

“A few perturbations did have low between-experiment transcriptome-wide impact correlations; most of these had very low transcriptome-wide impact, and thus, their correlations are expected to be noisy (Supplementary Fig. 18).”

May combine two claims.

Among 1660 perturbations with significant TI in all four cell types, TI correlation was highest between K562 and Jurkat (median 0.74) and between HepG2 and RPE1 (0.75).

“The transcriptome-wide impact correlation was highest between K562 and Jurkat (median: 0.74) and between HepG2 and RPE1 (median: 0.75) (Fig. 4C; Supplementary Fig. 19).”

May combine two claims.

Clustering perturbations with GMM found 44% in Cluster1 with selectively higher correlations between similar cell types (mean within similar 0.61; outside 0.35) and 56% in Cluster2 with high correlations across all cell types (mean within similar 0.75; outside 0.66).

“Clustering these perturbations with a Gaussian mixture model (Methods), we found that 44% of perturbations belonged to a cluster with selectively higher correlations between more similar cell types (Cluster1, mean correlation within similar cell type pairs: 0.61; outside similar pairs: 0.35), whereas 56% of the perturbations belonged to a cluster with high correlations across all cell types (Clust”

May combine two claims.

Genes with more efficient CRISPRi knockdown were not enriched for perturbation impact, suggesting enrichments not driven by guide differences.

“Genes with more efficient CRISPRi knockdown were not enriched for perturbation impact, suggesting that enrichments are not driven by differences between guides (Supplementary Fig. 15).”

May combine two claims.

For dTag depletion of Sox9 and Polycomb, TI correlations decayed smoothly with dosage difference, and smallest and largest perturbations were only moderately correlated (r=0.60, 0.48).

“For dTag depletion of Sox9 (Supplementary Table 12) and Polycomb (Supplementary Table 14), transcriptome-wide impact correlations decayed smoothly with the difference in dosage, and the smallest and largest perturbations were only moderately correlated (r=0.60, 0.48), implying that weak and strong perturbations have qualitatively different transcriptional consequences.”

May combine two claims.

Integrating all diagnosis pairs and technologies, TI correlations were substantially larger than sample Spearman correlations, with increase for 9/9 psychiatric trait pairs.

“Integrating data from all diagnosis pairs and technologies, we found that transcriptome-wide impact correlations were substantially larger than sample Spearman correlations, with an increase for 9/9 psychiatric trait pairs.”

May combine two claims.

Inferred bivariate effect size distributions showed effects either null in both conditions or positively correlated; little evidence of negatively correlated component or large numbers of condition-specific genes.

“Visualizing the inferred bivariate effect size distributions revealed that in general, effects are either null in both conditions or positively correlated; there was little evidence of a negatively correlated component or large numbers of condition-specific genes (Supplementary Fig. 24).”

May combine two claims.

Dosage affects not only magnitude of transcriptome-wide effect but also which genes are affected.

“In dose-response experiments, we found that dosage affects not only the magnitude of the transcriptome-wide effect, but also the genes that are affected.”

May combine two claims.

Response to GATA1 knockdown was highly correlated among all but the weakest perturbation, which was only moderately correlated with the strongest.

“The response to GATA1 knockdown was highly correlated among all but the weakest perturbation, which was only moderately correlated with the strongest perturbation (Fig. 5B).”

May combine two claims.

For autism, bipolar disorder, and schizophrenia, TI correlation between microarray and RNA-seq cohorts was 0.96, 0.91, and 0.78 respectively.

“The transcriptome-wide impact correlation was 0.96, 0.91, and 0.78 for autism, bipolar disorder, and schizophrenia respectively (Supplementary Fig. 25).”

May combine two claims.

The bivariate Gaussian mixture model identified two components: one with mean within-similar-cell-type correlation 0.75 and outside 0.66, and another with mean within 0.61 and outside 0.35.

“including one component with relatively high correlations within and outside of similar cell types (mean correlation within = 0.75, mean correlation outside = 0.66) and one component with lower correlations outside of similar cell types (mean correlation within = 0.61, mean correlation outside = 0.35).”

Probably not a substantial claim: may not pass S3 (load-bearing).

In simulations, TRADE-generated posterior-mean effect size estimates had lower mean squared error than DESeq2 effect size estimates.

“We briefly explored TRADE-generated posterior-mean effect size estimates in simulations, finding that these posterior-mean estimates had lower mean squared error than DESeq2 effect size estimates (Supplementary Figure 5)”

May combine two claims.

Existing methods for capturing aspects of the DE effect distribution have conceptual limitations, including lack of clear estimand, ambiguous units, or finite-sample bias.

“However, while these approaches are useful, they have conceptual limitations. Some lack a clear estimand (such as the running-sum GSEA statistic); some have ambiguous units (such as the energy distance); some are biased at finite sample size (such as the proportion of DE genes π in iDEA, see O’Connor et al11).”

• cites [6] Replogle, J. M. et al. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell 185, 2559–2575.e28 (2022). sourceNot checked yet

  Not checked in the free run.

• cites [7] Peidli, S. et al. scPerturb: harmonized single-cell perturbation data. Nat. Methods 21, 531–540 (2024). sourceNot checked yet

  Not checked in the free run.

• cites [8] Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA 102, 15545–15550 (2005). sourceNot checked yet

  Not checked in the free run.

• cites [9] Plaisier, S. B., Taschereau, R., Wong, J. A. & Graeber, T. G. Rank–rank hypergeometric overlap: identification of statistically significant overlap between gene-expression signatures. Nucleic Acids Res. 38, e169 (2010). sourceNot checked yet

  Not checked in the free run.

• cites [10] Ma, Y. et al. Integrative differential expression and gene set enrichment analysis using summary statistics for scRNA-seq studies. Nat. Commun. 11, 1585 (2020). sourceNot checked yet

  Not checked in the free run.

• cites [11] O’Connor, L. J. The distribution of common-variant effect sizes. Nat. Genet. 53, 1243–1249 (2021). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Perturb-seq data are noisy and many true effects may go undetected because the number of cells per perturbation is limited.

“However, the data produced by these screens are noisy, and many effects may go undetected.”

• Also stated in: § Introduction · hedged

  “Perturb-seq allows thousands of genetic perturbations to be tested in a single pool of cells, but the number of cells per perturbation (the sample size) is limited, such that many true effects may not be detected.”

Classified as probably not a substantial claim. May combine two claims.

GWAS power is limited, but statistical methods can estimate SNP-heritability and characterize genetic architecture even when few associations are significant.

“Even when genome-wide significant associations are few, statistical methods can be used to estimate “SNP-heritability,” a measure of the total genetic effect5, and to characterize nearly every aspect of genetic architecture.”

• cites [5] Yang, J. et al. Common SNPs explain a large proportion of the heritability for human height. Nat. Genet. 42, 565–569 (2010). sourceNot checked yet

  Not checked in the free run.

May combine two claims. The quote only partly states this claim.

TRADE-derived estimators of transcriptome-wide impact, effective number of DE genes, gene-set enrichments, and correlation are unbiased at finite sample size.

“TRADE also estimates gene-set enrichments, correlations between different experiments, and the effective number of differentially expressed genes, all of which are unbiased at finite sample size (Methods; Supplementary Fig. 1,12,14,16).”

Classified as probably not a substantial claim.

Perturb-Seq combines single-cell RNA sequencing with high-throughput CRISPR screens.

“A particularly scalable approach to differential expression is Perturb-Seq, which combines single-cell RNA sequencing with a high-throughput CRISPR screen1-4.”

• cites [1] Dixit, A. et al. Perturb-seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens. Cell 167, 1853–1866.e17 (2016). sourceNot checked yet

  Not checked in the free run.

• cites [2] Adamson, B. et al. A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response. Cell 167, 1867–1882.e21 (2016). sourceNot checked yet

  Not checked in the free run.

• cites [3] Jaitin, D. A. et al. Dissecting immune circuits by linking CRISPR-pooled screens with single-cell RNA-seq. Cell 167, 1883–1896.e15 (2016). sourceNot checked yet

  Not checked in the free run.

• cites [4] Datlinger, P. et al. Pooled CRISPR screening with single-cell transcriptome readout. Nat. Methods 14, 297–301 (2017). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Because dataset targeted common essential genes, many more cell-type-specific effects are expected in a larger atlas.

“As this dataset targeted primarily common essential genes which are expected to be important for growth across most cell types, there are expected to be many more examples of cell-type-specific effects in a larger cellular perturbation atlas.”

May combine two claims.

When a gene has a cell type specific function, perturbing it in the relevant cell type causes both larger magnitude and different pattern of effects across genes.

“This indicates that when a gene has a cell type specific function, perturbing that gene in the relevant cell type causes not only a larger magnitude of effect, but also a different pattern of effects across genes (Supplementary Fig. 21).”

May combine two claims.

πDEG estimation procedure was validated in simulations and downsampling.

“We validated our estimation procedure for πDEG in simulations (Supplementary Fig. 12) and in a downsampling analysis (Supplementary Fig. 13).”

Probably not a substantial claim: may not pass S3 (load-bearing).

One explanation for higher correlations among neuropsychiatric conditions is that transcriptomic effects are often downstream of condition liability and shared.

“One explanation for this difference is that transcriptomic effects are often downstream of condition liability, and these downstream effects are often shared between neuropsychiatric conditions.”

May combine two claims.

TI may be a suitable distance metric for cell-type variation because it is unbiased at finite sample size and computable from summary statistics.

“Transcriptome-wide impact may be a suitable distance metric for such analyses, as it is unbiased at finite sample size (unlike the commonly used Euclidean distance, Supplementary Note 6) and can be computed from differential expression summary statistics.”

May combine two claims.

Across cell types or replicate experiments, concordance between estimated effect sizes is attenuated due to sampling variation, but true effect sizes are often highly concordant.

“Across cell types or even replicate experiments, the concordance between estimated effect sizes is attenuated due to sampling variation, but in many cases, the true effect sizes are highly concordant.”

May combine two claims.

Gandal et al found DE effects correlated between conditions, paralleling genetic correlations.

“They found that differential expression effects were correlated between conditions, and that these correlations paralleled those between the genetic effects on those conditions (the genetic correlation).”

• cites [29] Gandal, M. J. et al. Shared molecular neuropathology across major psychiatric disorders parallels polygenic overlap. Science 359, 693–697 (2018). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Sox9 and Polycomb are haploinsufficient (pLI=117), indicating 50% depletion is deleterious.

“These genes are haploinsufficient (pLI = 117) indicating that 50% depletion is deleterious; our results suggest that stronger depletions nonetheless produce progressively larger cellular effects.”

May combine two claims.

Genetic correlation was estimated with REML accounting for sampling variation, whereas transcriptomic effect correlation was estimated as sample Spearman correlation between DE point estimates, which is downwardly biased in a sample size dependent manner.

“A notable difference between the transcriptomic and genetic analyses in this study is that the genetic correlation was estimated with a REML approach that accounts for sampling variation30, whereas the transcriptomic effect correlation was estimated as the sample Spearman correlation between differential expression point estimates, which is downwardly biased in a sample size dependent manner (”

• cites [30] Lee, S. H., Yang, J., Goddard, M. E., Visscher, P. M. & Wray, N. R. Estimation of pleiotropy between complex diseases using single-nucleotide polymorphism-derived genomic relationships and restricted maximum likelihood. Bioinformatics 28, 2540–2542 (2012). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Naqvi et al found a subset of Sox9 targets sensitive at partial dosage depletions, whereas a larger set affected only at full dosage depletions.

“Indeed, Naqvi et al26 found that a subset of Sox9 targets are sensitive at partial dosage depletions, whereas a much larger set of targets are affected only at full dosage depletions.”

• cites [26] Naqvi, S. et al. Precise modulation of transcription factor levels identifies features underlying dosage sensitivity. Nat. Genet. 55, 841–851 (2023). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

TRADE relies on predefined labels and cannot cluster cells into cell types.

“However, a limitation of TRADE is that it relies upon predefined labels and cannot be used to cluster cells into cell types.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some biological processes are better assayed using alternative modalities (splicing, chromatin, protein, imaging).

“In addition, some biological processes are better assayed using alternative modalities, including mRNA splicing, chromatin state, protein level, and imaging, all of which are now being studied at scale with single cell CRISPR screens39-44.”

• cites [39] Rubin, A. J. et al. Coupled single-cell CRISPR screening and epigenomic profiling reveals causal gene regulatory networks. Cell 176, 361–376.e17 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [40] Feldman, D. et al. Optical pooled screens in human cells. Cell 179, 787–799.e17 (2019). sourceNot checked yet

  Not checked in the free run.

• cites [41] Gu, J. et al. CRISPRmap: sequencing-free optical pooled screens mapping multi-omic phenotypes in cells and tissue. Preprint at bioRxiv https://doi.org/10.1101/2023.12.26.572587 (2023). sourceNot checked yet

  Not checked in the free run.

• cites [42] Binan, L. et al. Simultaneous CRISPR screening and spatial transcriptomics reveals intracellular, intercellular, and functional transcriptional circuits. Preprint at bioRxiv https://doi.org/10.1101/2023.11.30.569494 (2023). sourceNot checked yet

  Not checked in the free run.

• cites [43] Xu, Z., Sziraki, A., Lee, J., Zhou, W. & Cao, J. Dissecting key regulators of transcriptome kinetics through scalable single-cell RNA profiling of pooled CRISPR screens. Nat. Biotechnol. 42, 1218–1223 (2024). sourceNot checked yet

  Not checked in the free run.

• cites [44] Kudo, T. et al. Multiplexed, image-based pooled screens in primary cells and tissues with PerturbView. Nat. Biotechnol. https://doi.org/10.1038/s41587-024-02391-0 (2024). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

For genetic perturbations, parameters such as TI are likely driven by the pattern of causal regulatory connections (GRN).

“For genetic perturbations, parameters such as transcriptome-wide impact are likely driven by the pattern of causal regulatory connections between genes, i.e. the gene regulatory network (GRN).”

Probably not a substantial claim: may not pass S3 (load-bearing).

On day 7 after transduction, cells were measured to be more than 90% GFP positive.

“On day 7 post-transduction, cells were measured to be >90% GFP positive”

Classified as probably not a substantial claim.

Cells were filtered for quality by removing cells with low UMI content (Jurkat: >1750 UMIs, HepG2: >3000 UMIs) and high mitochondrial RNA content (Jurkat: <14%, HepG2: <18%).

“Cells were filtered for quality to remove cells with low UMI content (Jurkat: > 1750 UMIs, HepG2: > 3000 UMIs) and high mitochondrial RNA content (Jurkat: < 14%, HepG2: <18%).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Estimates of the fraction of signal in FDR-significant genes are expected to be upwardly biased by winner's curse.

“we note that such estimates are expected to be upwardly biased by winner’s curse.”

Probably not a substantial claim: may not pass S3 (load-bearing).

GFP-positive singlets were isolated by FACS in order to decrease cell doublets in the single-cell RNA-sequencing.

“In order to decrease cell doublets in the single-cell RNA-sequencing, GFP positive singlets were isolated by FACS”

Classified as probably not a substantial claim.

Choosing Sg as a diagonal matrix with diagonal entries equal to the variance of the individual estimated effects is appropriate when each estimate is derived from a different experiment, as is the case in these analyses.

“We choose Sg to be a diagonal matrix with diagonal entries equal to the variance of the individual estimated effects. This choice is appropriate when each estimate is derived from a different experiment, which is the case in our analyses.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The quality filters removed 7408 cells from the Jurkat-Essential experiment (262,956 retained) and 15,952 cells from the HepG2-Essential experiment (145,473 retained).

“These filters removed 7408 cells from the Jurkat-Essential experiment (262956 cells retained), and 15952 cells from the HepG2-Essential experiment (145473 cells retained).”

Probably not a substantial claim: may not pass S4 (based).

Removing the null-biased prior prevents bias in the distribution estimate.

“We removed the null-biased prior in order to prevent bias in our distribution estimate.”

• AssumptionRests on an assumption: Using a null-biased prior is crucial for accurately computing the local false sign rate, but is less important for estimating the effect size distribution itself.

Probably not a substantial claim: may not pass S3 (load-bearing).

On day 3 after transduction, the infection rate was measured as 7% using GFP as a transduction marker.

“On day 3 post-transduction, an infection rate of 7% was measured based on GFP as a marker of transduction”

Classified as probably not a substantial claim.

Stephens found that the half-uniform mixture model is sufficiently flexible to model realistic effect size distributions.

“Stephens 13 found that these this model is sufficiently flexible to model realistic effect size distributions.”

• cites [13] Stephens, M. False discovery rates: a new deal. Biostatistics 18, 275–294 (2016).Not checked yet

  Not checked in the free run.

May combine two claims.

The transcriptome-wide impact estimates have a slight downward bias, which may lead to slight over-shrinkage for underpowered perturbations.

“Our estimates could be used in this way, although we caution that slight downward bias in our transcriptome-wide impact estimates (Supplementary Figure 1) may lead to slight over-shrinkage for underpowered perturbations.”

Classified as probably not a substantial claim.

The targeted low infection rate of about 10% was chosen to reduce the chance that a single cell would be infected by multiple viruses.

“This low rate was chosen to reduce the chances of a single cell being infected by multiple viruses.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Cells were maintained at a coverage of more than 1000 cells per library element throughout the screen.

“Cells were maintained at a coverage of >1000 cells per library element for the duration of the screen.”

May combine two claims.

In our analyses, the influence of pseudoreplication bias is expected to be minimal because the pseudobulk observation units are different sequencing batches from the same cell line, with no variability from supra-cellular sampling units.

“Because there is no variability due to supra-cellular sampling unit (i.e. animal, donor, etc.), the influence of pseudoreplication bias is expected to be minimal.”

Classified as probably not a substantial claim.

The adaptive grid procedure produces a set of covariance matrices that attempt to tile all possible bivariate relationships between the two perturbations.

“This procedure produces a set of covariance matrices that attempt to tile all possible bivariate relationships between the two perturbations.”

• AssumptionRests on an assumption: mash was designed primarily for multivariate experiments with several conditions, where specifying all possible covariance patterns is not feasible.

Probably not a substantial claim: may not pass S4 (based).

The perturbation response enrichment approach is expected to be approximately unbiased for most gene sets.

“This approach is expected to be approximately unbiased for most gene sets.”

Classified as probably not a substantial claim.

Failure to account for biological variability across supra-cellular sampling units can lead to overestimation of effects, a phenomenon called pseudoreplication bias.

“In single cell differential expression analyses, a failure to account for biological variability across supra-cellular sampling units (i.e. animal, donor, etc.) can lead to overestimation of effects; this phenomenon is referred to as “pseudoreplication bias” 36,37.”

• cites [36] Squair, J. W. et al. Confronting false discoveries in single-cell differential expression. Nat. Commun. 12, 5692 (2021). sourceNot checked yet

  Not checked in the free run.

• cites [37] Soneson, C. & Robinson, M. D. Bias, robustness and scalability in single-cell differential expression analysis. Nat. Methods 15, 255–261 (2018). sourceNot checked yet

  Not checked in the free run.