Exploring genetic interaction manifolds constructed from rich single-cell phenotypes

May combine two claims.

The CRISPRa genetic interaction map contained fewer clusters than expected in a random map, consistent with the typical low-rank structure of genetic interaction maps.

“In addition, the map contained fewer clusters than would be expected in a random map (fig. S5C–D), consistent with the typical low-rank structure of GI maps (i.e. groups of genes interact similarly so that there are fewer overall degrees of freedom than total genes).”

May combine two claims.

The variation in sensitivity did not appear to result from stable differences in MAPK1 and DUSP9 expression, suggesting possible roles for historical differences or stochastic gene expression.

“This variation did not appear to be the result of stable differences in the expression of MAPK1 and DUSP9 (fig. S12B), suggesting a possible role either for historical differences or stochastic gene expression.”

May combine two claims.

In the DUSP9/ETS2 interaction, the DUSP9 phenotype dominated, suggesting that DUSP9 overexpression antagonized ETS2.

“In the DUSP9/ETS2 interaction (Fig. 5A), the DUSP9 phenotype dominated, suggesting that DUSP9 overexpression antagonized ETS2.”

May combine two claims.

There was modest concordance between genetic interaction profile and Perturb-seq profile correlations (R=0.29, p<10−103), suggesting that transcriptional data provide a complementary scalable means of comparing genes that might inform genetic interaction prediction.

“We examined the Perturb-seq profiles of single gene overexpression (fig. S13A) and found that there was a modest concordance between GI profile and Perturb-seq profile correlations (R=0.29, p<10−103; fig. S13B), suggesting that the transcriptional data provided a complementary, scalable means of comparing genes that might inform GI prediction.”

• AssumptionRests on an assumption: Predicting genetic interactions is substantially similar to predicting shopping preferences, so recommender system algorithms are applicable.

May combine two claims.

The approach was substantially better than random sampling at predicting the top 10% of interactions and reasonably preserved rank order (Spearman ρ≈0.5 at 10% sampling) and pairwise similarities.

“Our approach was also substantially better than random sampling at predicting the top 10% of interactions and reasonably preserved both the rank order of all interactions (Spearman ρ ≈ 0.5 at 10% sampling; Fig. 6D–E and fig S14C) and the pairwise similarities between GI profiles (Fig. 6F).”

• AssumptionRests on an assumption: Predicting genetic interactions is substantially similar to predicting shopping preferences, so recommender system algorithms are applicable.

May combine two claims.

The two-dimensional visualization and clustering identified numerous distinct categories of interaction, showing that the model generalizes the one-dimensional buffering versus synthetic lethal paradigm.

“The resulting figure (Fig. 4F) identified numerous distinct categories of interaction, showing that the model can serve as a generalization of the one-dimensional “buffering vs. synthetic lethal” paradigm that has typically been used to categorize genetic interactions.”

• cites [25] Cheng Y, Wong MT, van der Maaten L, Newell EW, Categorical Analysis of Human T Cell Heterogeneity with One-Dimensional Soli-Expression by Nonlinear Stochastic Embedding. J. Immunol 196, 924– 932 (2016). 26667171 PMC4705595 sourceNot checked yet

  Not checked in the free run.

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

The matrix factorization model constrained by transcriptional similarity predicted unobserved genetic interactions and preserved much of the large-scale structure of the map.

“The end result (Fig. 6B–D) preserved much of the large–scale structure of the map as seen through block averaging of GI scores (Fig. 6C–D; fig. S14B and materials and methods).”

• AssumptionRests on an assumption: Predicting genetic interactions is substantially similar to predicting shopping preferences, so recommender system algorithms are applicable.

May combine two claims.

Levels of target gene activation spanned a broad range, with a general trend that poorly expressed targets were more highly induced.

“Levels of target gene activation spanned a broad range (Fig. 1G and fig. S6B–C), with a general trend that poorly expressed targets were more highly induced.”

May combine two claims.

Transcriptional readouts were obtained for 287 perturbations across approximately 110,000 single cells, with a median of 273 cells per condition.

“In total, we obtained transcriptional readouts for 287 perturbations measured across ~110,000 single cells (median 273 cells per condition, Methods, fig S1, and Table S7) in one pooled experiment.”

• AssumptionRests on an assumption: Many genetic interactions within blocks of the fitness genetic interaction map are likely explained by similar mechanisms, so sampling 132 gene pairs broadly samples the biological processes represented.

May combine two claims.

Highly correlated genes in the genetic interaction map were enriched for genes with the same DAVID term annotations, allowing unbiased annotation of clusters.

“Highly correlated genes were enriched for genes with the same DAVID term annotations (13), allowing for unbiased annotation of clusters (Fig. 1B and fig. S5A–B, Table S6).”

• cites [13] Huang DW, Sherman BT, Lempicki RA, Bioinformatics enrichment tools: paths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Res 37, 1– 13 (2009). 19033363 PMC2615629 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A cluster of perturbations induced CD41 but the cells did not adopt characteristic megakaryocyte morphology, suggesting at best priming towards megakaryocytic differentiation.

“Finally, a cluster of perturbations induced expression of the canonical megakaryocyte marker CD41, but these cells did not adopt the characteristic morphological features of megakaryocytes by microscopy (fig. S9G), suggesting that they are at best primed towards megakaryocytic differentiation (16, 17).”

• cites [16] Alitalo R, Induced differentiation of K562 leukemia cells: a model for studies of gene expression in early megakaryoblasts. Leuk. Res 14, 501– 514 (1990). 2197510 sourceNot checked yet

  Not checked in the free run.

• cites [17] Orkin SH, Zon LI, Hematopoiesis: An Evolving Paradigm for Stem Cell Biology. Cell 132, 631– 644 (2008). 18295580 PMC2628169 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The PTPN12/SNAI1 genetic interaction resulted from genetic suppression, with two smaller coefficients (c1=0.60, c2=0.57).

“By contrast, the PTPN12/SNAI1 GI resulted from genetic suppression (i.e. when combined overexpression of two genes attenuated each other’s individual phenotypes), resulting in two smaller coefficients (e.g. PTPN12/SNAI1, c1 = 0.60, c2 = 0.57, Fig. 4E and fig. S11B).”

May combine two claims.

A cluster of mean expression profiles labeled G1 cell cycle arrest contained canonical cell cycle regulators, and single-cell data confirmed that these perturbations induced cell cycle arrest in the expected stages, explaining the growth defect.

“For example, one cluster of mean expression profiles (“G1 cell cycle arrest”) contained canonical regulators of the cell cycle. In this case, the underlying single-cell data confirmed that these perturbations induced cell cycle arrest in the expected cell cycle stages (Fig. 2E–F), explaining the growth defect.”

May combine two claims.

Down-sampling analyses indicated that as few as 50 cells per perturbation could be sufficient, meaning approximately 10^6 cells to collect side information for the entire set of protein coding sequences.

“By down-sampling our measured perturbations and re-performing our analyses, we observed that as few as 50 cells per perturbation could be sufficient, meaning ~106 cells to collect side information for the entire set of protein coding sequences (Fig. 6G).”

May combine two claims.

Both fitness and genetic interaction scores were distributed throughout the genetic interaction manifold, consistent with scalar fitness measurements collapsing a larger landscape of transcriptional states.

“Both fitness and GI scores were distributed throughout the GI manifold (Fig. 2C–D), in accordance with the idea that scalar fitness measurements collapse the much larger landscape of transcriptional states (Fig. 1A).”

May combine two claims.

ETS2 and MAPK1 induced similar phenotypes, and ETS2 transcription was activated in all backgrounds (9.3-fold in MAPK1, 9.2-fold in ETS2, and 35.8-fold in MAPK1/ETS2 overexpression).

“Finally, ETS2 and MAPK1 induced similar phenotypes, and ETS2 transcription was activated in all backgrounds (9.3-fold in MAPK1, 9.2-fold in ETS2, and 35.8-fold in MAPK1/ETS2 overexpression, Fig. 5C).”

May combine two claims.

Gene-level genetic interaction scores followed a bell-shaped distribution with interactions ranging beyond the expectation from negative control sgRNAs.

“and followed a bell-shaped distribution but with GIs ranging well beyond the expectation from negative control sgRNAs (Fig. 1F and fig. S3B).”

May combine two claims.

Overexpressing either CBL or CNN1 induced similar transcriptional changes, and single-cell analysis revealed progression from unperturbed through singly perturbed to doubly perturbed CBL/CNN1 cells.

“Overexpressing either gene induced similar transcriptional changes, and single-cell analysis revealed an apparent progression of phenotypes from unperturbed through singly-perturbed to doubly-perturbed CBL/CNN1 cells (Fig. 3C,D).”

May combine two claims.

Independent replicate experiments showed high concordance for sgRNA-level genetic interaction scores and genetic interaction profile correlations, and independent sgRNAs targeting the same gene were more similar than the background (median R=0.50 versus 0.04).

“Independent replicate experiments showed high levels of concordance for sgRNA-level GI scores and GI profile correlations, and independent sgRNAs targeting the same gene were much more similar than the background of all sgRNA GI correlations (median R=0.50 compared to 0.04; Fig. 1C,D and fig. S3A, Table S4).”

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

Perturb-seq pairs CRISPR-based screens with single-cell RNA sequencing, allowing each cell to be an independent experiment connecting a genetic perturbation to its transcriptional consequences.

“For example, Perturb-seq pairs CRISPR-based screens with single-cell RNA sequencing (5–8). Each individual cell is in effect an independent experiment connecting a genetic perturbation to its transcriptional consequences, allowing hundreds of thousands of parallel measurements (9, 10).”

• cites [5] Jaitin DA, Weiner A, Yofe I, Lara-Astiaso D, Keren-Shaul H, David E, Salame TM, Tanay A, van Oudenaarden A, Amit I, Dissecting Immune Circuits by Linking CRISPR-Pooled Screens with Single-Cell RNA-Seq. Cell 167, 1883– 1896. e15 (2016). 27984734 sourceNot checked yet

  Not checked in the free run.

• cites [6] Adamson B, Norman TM, Jost M, Cho MY, Nuñez JK, Chen Y, Villalta JE, Gilbert LA, Horlbeck MA, Hein MY, Pak RA, Gray AN, Gross CA, Dixit A, Parnas O, Regev A, Weissman JS, A Multiplexed Single-Cell CRISPR Screening Platform Enables Systematic Dissection of the Unfolded Protein Response. Cell 167, 1867– 1882. e21 (2016). 27984733 PMC5315571 sourceNot checked yet

  Not checked in the free run.

• cites [7] Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, Marjanovic ND, Dionne D, Burks T, Raychndhury R, Adamson B, Norman TM, Lander ES, Weissman JS, Friedman N, Regev A, Perturb-seq: Dissecting molecular circuits with scalable single cell RNA profiling of pooled genetic screens. Cell 167, 1853– 1866. e17 (2016). 27984732 PMC5181115 sourceNot checked yet

  Not checked in the free run.

• cites [8] Datlinger P, Rendeiro AF, Schmidl C, Krausgruber T, Traxler P, Klughammer J, Schuster LC, Kuchler A, Alpar D, Bock C, Pooled CRISPR screening with single-cell transcriptome readout. Nat. Methods 14, 297– 301 (2017). 28099430 PMC5334791 sourceNot checked yet

  Not checked in the free run.

• cites [9] Gasperini M, Hill AJ, McFaline-Figueroa JL, Martin B, Kim S, Zhang MD, Jackson D, Leith A, Schreiber J, Noble WS, Trapnell C, Ahituv N, Shendure J, A Genome-wide Framework for Mapping Gene Regulation via Cellular Genetic Screens. Cell 176, 377– 390. e19 (2019). 30612741 PMC6690346 sourceNot checked yet

  Not checked in the free run.

• cites [10] Replogle JM, Xu A, Norman TM, Meer EJ, Terry JM, Riordan D, Srinivas N, Mikkelsen TS, Weissman JS, Adamson B, Direct capture of CRISPR guides enables scalable, multiplexed, and multi-omic Perturb-seq. bioRxiv, 503367 (2018).Not checked yet

  Not checked in the free run.

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

Genetic interactions can reveal synthetic lethal vulnerabilities in tumors, identify suppressors of inherited and acquired disorders, and guide the design of gene cocktails to drive differentiation between cell types.

“GIs can reveal synthetic lethal vulnerabilities in tumors, identify suppressors of inherited and acquired disorders, and guide the design of cocktails of genes to drive differentiation between cell types (3, 4).”

• cites [3] Hartman JL, Garvik B, Hartwell L, Principles for the buffering of genetic variation. Science 291, 1001– 1004 (2001). 11232561 sourceNot checked yet

  Not checked in the free run.

• cites [4] Takahashi K, Yamanaka S, Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. Cell 126, 663– 676 (2006). 16904174 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The genetic interaction manifold carries more information about interactions than a traditional genetic interaction map and is intrinsically more interpretable, including resolving distinct outcomes and modeling how perturbations combine.

“We argue that this GI manifold carries much more information about interactions than a traditional GI map, and is intrinsically more interpretable in several ways, including the ability to resolve the distinct outcomes underlying GIs and to model the different ways genetic perturbations combine to yield new phenotypes.”

Classified as probably not a substantial claim.

Mapping pairwise interactions among approximately 10,000 transcribed genes in a human cell would require measuring approximately 50 million double mutants.

“For example, mapping pairwise interactions among the ~10,000 transcribed genes in a human cell would require measuring ~50 million double mutants.”

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

Pioneering yeast genetic interaction maps between all gene pairs enabled systematic determination of gene function, identification of protein complexes, and definition of gene regulatory networks.

“Pioneering efforts in yeast to construct systematic GI maps between all gene pairs have enabled systematic determination of gene function, identification of protein complexes and definition of gene regulatory networks in a principled and unbiased manner (1, 2).”

• cites [1] Costanzo M, Kuzmin E, van Leeuwen J, Mair B, Moffat J, Boone C, Andrews B, Global Genetic Networks and the Genotype-to-Phenotype Relationship. Cell 177, 85– 100 (2019). 30901552 PMC6817365 sourceNot checked yet

  Not checked in the free run.

• cites [2] Domingo J, Baeza-Centurion P, Lehner B, The Causes and Consequences of Genetic Interactions (Epistasis). Annu Rev Genomics Hum Genet (2019), 10.1146/annurev-genom-083118-014857. 31082279 sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

A split experimental and computational approach, in which a subset of interactions are sampled and the remainder are predicted computationally, is a possible strategy to address the scale of genetic interactions.

“One possible strategy to address the scale of genetic interactions is to pursue a split experimental/computational approach, in which only a subset of interactions are sampled (either randomly or through “compressed” experimental designs (28)) and the remainder are predicted computationally (Fig. 6A) (7, 29, 30).”

• cites [7] Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, Marjanovic ND, Dionne D, Burks T, Raychndhury R, Adamson B, Norman TM, Lander ES, Weissman JS, Friedman N, Regev A, Perturb-seq: Dissecting molecular circuits with scalable single cell RNA profiling of pooled genetic screens. Cell 167, 1853– 1866. e17 (2016). 27984732 PMC5181115 sourceNot checked yet

  Not checked in the free run.

• cites [29] Madhukar NS, Elemento O, Pandey G, Prediction of Genetic Interactions Using Machine Learning and Network Properties. Front Bioeng Biotechnol 3, 172 (2015). 26579514 PMC4620407 sourceNot checked yet

  Not checked in the free run.

• cites [30] Žitnik M, Zupan B, Data Imputation in Epistatic MAPs by Network-Guided Matrix Completion. Journal of Computational Biology 22, 595– 608 (2015). 25658751 PMC4449711 sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Predicting genetic interactions is substantially similar to predicting shopping preferences, so recommender system algorithms are applicable.

May combine two claims.

The geometric genetic interaction model identified different ways in which genetic perturbations combine to yield new phenotypes, allowing ordering of genes into linear pathways.

“We derived a simple, geometric GI model and used it to identify the different ways in which genetic perturbations combine to yield new phenotypes, for example allowing us to order genes into linear pathways.”

May combine two claims.

These results suggested a linear regulatory pathway in which DUSP9 inhibits MAPK1, which activates ETS2, consistent with known biology.

“Taken together, these results suggested a linear regulatory pathway in which DUSP9 (a phosphatase) inhibits MAPK1 (a kinase) that activates ETS2 (a transcription factor), which is consistent with the known biology of these gene families (Fig. 5D) (26).”

• cites [26] Caunt CJ, Keyse SM, Dual-specificity MAP kinase phosphatases (MKPs). FEBS J 280, 489– 504 (2013). 22812510 PMC3594966 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The genetic interaction map robustly identified many strong interactions, but the origins of specific interactions were difficult to deduce because each interaction was characterized only by a scalar value.

“However, though the GI map robustly identified many strong GIs, the origins of specific interactions were difficult to deduce as each GI was characterized only by a single scalar value (i.e. deviation from expected growth rate).”

May combine two claims.

Functionally related genes interact similarly, making the genetic interaction manifold highly constrained and enabling imputation strategies.

“We and others have established that functionally related genes interact similarly—in geometric terms the GI manifold is therefore highly constrained, enabling imputation strategies (1, 2, 7, 12, 29).”

• cites [1] Costanzo M, Kuzmin E, van Leeuwen J, Mair B, Moffat J, Boone C, Andrews B, Global Genetic Networks and the Genotype-to-Phenotype Relationship. Cell 177, 85– 100 (2019). 30901552 PMC6817365 sourceNot checked yet

  Not checked in the free run.

• cites [2] Domingo J, Baeza-Centurion P, Lehner B, The Causes and Consequences of Genetic Interactions (Epistasis). Annu Rev Genomics Hum Genet (2019), 10.1146/annurev-genom-083118-014857. 31082279 sourceNot checked yet

  Not checked in the free run.

• cites [7] Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, Marjanovic ND, Dionne D, Burks T, Raychndhury R, Adamson B, Norman TM, Lander ES, Weissman JS, Friedman N, Regev A, Perturb-seq: Dissecting molecular circuits with scalable single cell RNA profiling of pooled genetic screens. Cell 167, 1853– 1866. e17 (2016). 27984732 PMC5181115 sourceNot checked yet

  Not checked in the free run.

• cites [12] Horlbeck MA, Xu A, Wang M, Bennett NK, Park CY, Bogdanoff D, Adamson B, Chow ED, Kampmann M, Peterson TR, Nakamura K, Fischbach MA, Weissman JS, Gilbert LA, Mapping the Genetic Landscape of Human Cells. Cell 174, 953– 967. e22 (2018). 30033366 PMC6426455 sourceNot checked yet

  Not checked in the free run.

• cites [29] Madhukar NS, Elemento O, Pandey G, Prediction of Genetic Interactions Using Machine Learning and Network Properties. Front Bioeng Biotechnol 3, 172 (2015). 26579514 PMC4620407 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Perturb-seq provides a scalable means of constraining searches for genetic interactions, enabling exploration of the genetic interaction landscape.

“Perturb-seq provides a scalable means of constraining these types of searches (7), enabling exploration of the GI landscape.”

• cites [7] Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, Marjanovic ND, Dionne D, Burks T, Raychndhury R, Adamson B, Norman TM, Lander ES, Weissman JS, Friedman N, Regev A, Perturb-seq: Dissecting molecular circuits with scalable single cell RNA profiling of pooled genetic screens. Cell 167, 1853– 1866. e17 (2016). 27984732 PMC5181115 sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Predicting genetic interactions is substantially similar to predicting shopping preferences, so recommender system algorithms are applicable.

Classified as probably not a substantial claim.

Because the approach is sensitive to single-cell phenotypes, incomplete differentiation, and multiple differentiation states, it could facilitate higher-order combinatorial perturbation screens for driving cells into distinct differentiation states.

“More generally, because our approach was sensitive to single-cell phenotypes, incomplete differentiation and could simultaneously detect signatures of multiple differentiation states, it could facilitate higher-order combinatorial perturbation screens aimed at improving protocols for driving cells into distinct differentiation states.”

May combine two claims.

Buffering interactions travel less far along the genetic interaction manifold, with smaller coefficients, while synthetic sick or lethal interactions travel further, with bigger coefficients.

“An intuitive explanation was that buffering interactions travel less “far” (smaller coefficients) along the GI manifold while SSL interactions travel further (bigger coefficients) (Fig. 4A).”

Probably not a substantial claim: may not pass S2 (checkable).

By intelligently measuring and exploring the genetic interaction manifold, one can start to create a global view of the nonlinear mapping between genotype and phenotype.

“By intelligently measuring and exploring the GI manifold, one can start to create a global view of the nonlinear mapping between genotype and phenotype.”