CoChat

The DOI matches its Crossref record; 51 of the 51 references it cites were found in a registry or an index.

Predicting transcriptional outcomes of novel multigene perturbations with GEARS

Open access
Roohani, Yusuf, Huang, Kexin, Leskovec, Jure · 2024 · Nature Biotechnology

Abstract

Understanding cellular responses to genetic perturbation is central to numerous biomedical applications, from identifying genetic interactions involved in cancer to developing methods for regenerative medicine. However, the combinatorial explosion in the number of possible multigene perturbations severely limits experimental interrogation. Here, we present graph-enhanced gene activation and repression simulator (GEARS), a method that integrates deep learning with a knowledge graph of gene–gene relationships to predict transcriptional responses to both single and multigene perturbations using single-cell RNA-sequencing data from perturbational screens. GEARS is able to predict outcomes of perturbing combinations consisting of genes that were never experimentally perturbed. GEARS exhibited 40% higher precision than existing approaches in predicting four distinct genetic interaction subtypes in a combinatorial perturbation screen and identified the strongest interactions twice as well as prior approaches. Overall, GEARS can predict phenotypically distinct effects of multigene perturbations and thus guide the design of perturbational experiments.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41587-023-01905-6

Cites

51 of 51 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 GEARS combines prior knowledge with deep learning to predict postperturbation gene expression. a, Problem formulation: given unperturbed gene expression (green) and applied perturbation (red), predict the gene expression outcome (purple). Each box corresponds to an individual gene. Arrows indicate change in expression. b, GEARS model architecture. (i) For each gene in the unperturbed state, GEARS initializes a gene embedding vector (green) and a gene perturbation embedding vector (red) (ii). These embedding vectors are assigned as node features in the gene relationship graph and the perturbation relationship graph (iii). A GNN is used to combine information between neighbors in each graph. Each resulting gene embedding is summed with the perturbation embedding of each perturbation in the perturbation set (iv). The output is combined across all genes using the cross-gene layer and fed into gene-specific output layers (v). The final result is postperturbation gene expression; MLP, multilayer perceptron.
Fig. 1 GEARS combines prior knowledge with deep learning to predict postperturbation gene expression. a, Problem formulation: given unperturbed gene expression (green) and applied perturbation (red), predict the gene expression outcome (purple). Each box corresponds to an individual gene. Arrows indicate change in expression. b, GEARS model architecture. (i) For each gene in the unperturbed state, GEARS initializes a gene embedding vector (green) and a gene perturbation embedding vector (red) (ii). These embedding vectors are assigned as node features in the gene relationship graph and the perturbation relationship graph (iii). A GNN is used to combine information between neighbors in each graph. Each resulting gene embedding is summed with the perturbation embedding of each perturbation in the perturbation set (iv). The output is combined across all genes using the cross-gene layer and fed into gene-specific output layers (v). The final result is postperturbation gene expression; MLP, multilayer perceptron.CC BY · PubMed Central
Fig. 2 GEARS outperforms alternative approaches in predicting postperturbation gene expression. a, Train–test data split for single-gene perturbations. b, The m.s.e. in predicted postperturbation gene expression for single-gene perturbations normalized to the no perturbation case. For each perturbation, the 20 most differentially expressed (DE) genes were considered; perturb, perturbation; GRN, gene regulatory network. c, Pearson correlation between mean predicted postperturbation differential gene expression over control and true values across all genes. d, Fraction of the top 20 differentially expressed genes where the predicted postperturbation differential expression is in the opposite direction of the ground truth. e, Train–test data split categories for two-gene perturbations. f, Normalized m.s.e. in predicted postperturbation gene expression for two-gene perturbations. g, Boxes indicate experimentally measured differential gene expression after perturbing the gene combination FOSB and CEBPB (n = 85). The red symbol shows the mean change in gene expression predicted by GEARS when it has only seen FOSB experimentally perturbed at the time of training. The green dotted line shows mean unperturbed control gene expression. Whiskers represent the last data point within 1.5× interquartile range. h, Jaccard similarity between model-predicted differentially expressed genes and true differentially expressed genes. Throughout the figure, markers correspond to the mean and error bars correspond to 95% confidence intervals computed over predictions made by five models trained using different data splits (n = 5).
Fig. 2 GEARS outperforms alternative approaches in predicting postperturbation gene expression. a, Train–test data split for single-gene perturbations. b, The m.s.e. in predicted postperturbation gene expression for single-gene perturbations normalized to the no perturbation case. For each perturbation, the 20 most differentially expressed (DE) genes were considered; perturb, perturbation; GRN, gene regulatory network. c, Pearson correlation between mean predicted postperturbation differential gene expression over control and true values across all genes. d, Fraction of the top 20 differentially expressed genes where the predicted postperturbation differential expression is in the opposite direction of the ground truth. e, Train–test data split categories for two-gene perturbations. f, Normalized m.s.e. in predicted postperturbation gene expression for two-gene perturbations. g, Boxes indicate experimentally measured differential gene expression after perturbing the gene combination FOSB and CEBPB (n = 85). The red symbol shows the mean change in gene expression predicted by GEARS when it has only seen FOSB experimentally perturbed at the time of training. The green dotted line shows mean unperturbed control gene expression. Whiskers represent the last data point within 1.5× interquartile range. h, Jaccard similarity between model-predicted differentially expressed genes and true differentially expressed genes. Throughout the figure, markers correspond to the mean and error bars correspond to 95% confidence intervals computed over predictions made by five models trained using different data splits (n = 5).CC BY · PubMed Central
Fig. 3 GEARS accurately predicts non-additive combinatorial effects and genetic interaction subtypes. a, Illustration of an additive interaction between two genes after perturbation. X and Y represent change over the unperturbed state caused by single-gene perturbations. Z is a combinatorial perturbation of both genes. b, Definition of genetic interaction subtypes. c, Mean precision@10 in predicting genetic interactions from 131 two-gene combinations (error bars represent s.d.). A random model performs 1,000 random draws; other models perform three predictions (n = 3). d, Change in gene expression after perturbing the combination PTPN12 and ZBTB25. The gray bars show the true mean postperturbation gene expression change (n = 257). The hatched gray bars show the true change for each of the two single-gene perturbations performed individually (PTPN12 n = 164 and ZBTB25 n = 247), which are summed by the naive additive model. The red bar indicates the prediction made by GEARS (n = 3 trained models). Error bars correspond to 95% confidence interval.
Fig. 3 GEARS accurately predicts non-additive combinatorial effects and genetic interaction subtypes. a, Illustration of an additive interaction between two genes after perturbation. X and Y represent change over the unperturbed state caused by single-gene perturbations. Z is a combinatorial perturbation of both genes. b, Definition of genetic interaction subtypes. c, Mean precision@10 in predicting genetic interactions from 131 two-gene combinations (error bars represent s.d.). A random model performs 1,000 random draws; other models perform three predictions (n = 3). d, Change in gene expression after perturbing the combination PTPN12 and ZBTB25. The gray bars show the true mean postperturbation gene expression change (n = 257). The hatched gray bars show the true change for each of the two single-gene perturbations performed individually (PTPN12 n = 164 and ZBTB25 n = 247), which are summed by the naive additive model. The red bar indicates the prediction made by GEARS (n = 3 trained models). Error bars correspond to 95% confidence interval.CC BY · PubMed Central
Fig. 4 GEARS can predict new biologically meaningful phenotypes. a, Workflow for predicting all pairwise combinatorial perturbation outcomes of a set of genes. b, Low-dimensional representation of postperturbation gene expression for 102 one-gene perturbations and 128 two-gene perturbations used to train GEARS. A random selection is labeled. c, GEARS predicts postperturbation gene expression for all 5,151 pairwise combinations of the 102 single genes seen experimentally perturbed. Predicted postperturbation phenotypes (non-black symbols) are often different from phenotypes seen experimentally (black symbols). Colors indicate Leiden clusters labeled using marker gene expression (Supplementary Information).
Fig. 4 GEARS can predict new biologically meaningful phenotypes. a, Workflow for predicting all pairwise combinatorial perturbation outcomes of a set of genes. b, Low-dimensional representation of postperturbation gene expression for 102 one-gene perturbations and 128 two-gene perturbations used to train GEARS. A random selection is labeled. c, GEARS predicts postperturbation gene expression for all 5,151 pairwise combinations of the 102 single genes seen experimentally perturbed. Predicted postperturbation phenotypes (non-black symbols) are often different from phenotypes seen experimentally (black symbols). Colors indicate Leiden clusters labeled using marker gene expression (Supplementary Information).CC BY · PubMed Central
Fig. 5 GEARS can search perturbational space for novel genetic interactions of different subtypes. a, Workflow for predicting genetic interaction (GI) scores. b, Multidimensional genetic interaction map generated by GEARS for all pairwise combinations of 102 single genes perturbed in Norman et al.9. For each combination, GEARS predicted genetic interaction scores for five different genetic interactions: synergy and suppression (red to blue), neomorphism (green), redundancy (orange) and epistasis (purple).
Fig. 5 GEARS can search perturbational space for novel genetic interactions of different subtypes. a, Workflow for predicting genetic interaction (GI) scores. b, Multidimensional genetic interaction map generated by GEARS for all pairwise combinations of 102 single genes perturbed in Norman et al.9. For each combination, GEARS predicted genetic interaction scores for five different genetic interactions: synergy and suppression (red to blue), neomorphism (green), redundancy (orange) and epistasis (purple).CC BY · PubMed Central
Extended Data Fig. 1 GEARS identifies groups of genes inducing similar perturbation effect, even when not seen perturbed previously. Each plot presents a low-dimensional (UMAP) representation of postperturbation gene expression following genetic perturbations that were held out in the test set. Each column corresponds to a different split of the experimental data into training and test sets. a, Each panel corresponds to true postperturbational transcriptional state measured using a Perturb-Seq assay. Colors correspond to distinct clusters identified using Leiden clustering set to a constant resolution across all panels. The largest cluster is assumed to show minimal perturbation effect and is colored grey. b, Each panel corresponds to postperturbation state predicted by GEARS. Colors correspond to the true labels identified when clustering the true experimental data, thus each point is labeled the same as in a. Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) were used to compare clusters identified by GEARS to those observed in true postperturbation expression for each data split. Average values for each metric across splits shown on left. c, Same as b using a baseline model that predicts no perturbation effect. d, Same as b using a baseline model that predicts mean perturbation effect.
Extended Data Fig. 1 GEARS identifies groups of genes inducing similar perturbation effect, even when not seen perturbed previously. Each plot presents a low-dimensional (UMAP) representation of postperturbation gene expression following genetic perturbations that were held out in the test set. Each column corresponds to a different split of the experimental data into training and test sets. a, Each panel corresponds to true postperturbational transcriptional state measured using a Perturb-Seq assay. Colors correspond to distinct clusters identified using Leiden clustering set to a constant resolution across all panels. The largest cluster is assumed to show minimal perturbation effect and is colored grey. b, Each panel corresponds to postperturbation state predicted by GEARS. Colors correspond to the true labels identified when clustering the true experimental data, thus each point is labeled the same as in a. Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) were used to compare clusters identified by GEARS to those observed in true postperturbation expression for each data split. Average values for each metric across splits shown on left. c, Same as b using a baseline model that predicts no perturbation effect. d, Same as b using a baseline model that predicts mean perturbation effect.CC BY · PubMed Central
Extended Data Fig. 2 Identifying significant enrichment for true differentially expressed genes in GEARS predictions. a, Hypergeometric distribution used to model the probability of obtaining a random overlap between the differentially expressed genes predicted by GEARS and the true significantly differentially expressed genes following a perturbation. In this example, 142 genes were shared between GEARS and the true prediction. A p-value is calculated for each perturbation in the held out set. b, Box-plot showing the log (base 10) of the p-value for all held-out perturbations in the Norman et al. 2019 dataset. To account for multiple hypothesis testing (561 tests), a Bonferroni correction was applied, using a significance threshold of 0.05. A black dashed line represents the adjusted threshold. GEARS was trained on 5 different data splits (n=5). Number of data points for each bar are listed above it. Whiskers represent last data point within 1.5x interquartile range below the first quantile and above the third quantile.
Extended Data Fig. 2 Identifying significant enrichment for true differentially expressed genes in GEARS predictions. a, Hypergeometric distribution used to model the probability of obtaining a random overlap between the differentially expressed genes predicted by GEARS and the true significantly differentially expressed genes following a perturbation. In this example, 142 genes were shared between GEARS and the true prediction. A p-value is calculated for each perturbation in the held out set. b, Box-plot showing the log (base 10) of the p-value for all held-out perturbations in the Norman et al. 2019 dataset. To account for multiple hypothesis testing (561 tests), a Bonferroni correction was applied, using a significance threshold of 0.05. A black dashed line represents the adjusted threshold. GEARS was trained on 5 different data splits (n=5). Number of data points for each bar are listed above it. Whiskers represent last data point within 1.5x interquartile range below the first quantile and above the third quantile.CC BY · PubMed Central
Extended Data Fig. 3 Model ablation study highlights relative importance of GEARS components under different generalization conditions. The ‘No Graph’ condition removes both the gene ontology graph and co-expression graph; ‘No GO Graph’ removes the gene ontology graph; ‘No Co-Express Graph’ removes the co-expression graph; ‘No Cross-gene’ removes the cross-gene MLP layer; ‘No Gene-specific Decoder’ removes the gene specific decoder MLP and uses a shared MLP instead; ‘MSE Loss’ switches from the auto-focus loss to the regular L2 loss. Four generalization conditions are considered: a, (1/1 Unseen) single-gene perturbations not seen experimentally perturbed at the time of training. b–d, (2/2 Unseen) two-gene perturbations in which both genes were not seen experimentally perturbed individually at the time of training (b), (1/2 Unseen) one of the two genes was not seen experimentally perturbed (c) or (0/2 Unseen) both genes have been seen experimentally perturbed (d). Performance is measured using the mean squared error in predicted postperturbation gene expression for the top 20 most differentially expressed genes. For all panels (a–d) the marker indicates the mean MSE over predictions made by models trained using 5 different training data splits (n=5). The error bars represent bootstrapped 95% CI.
Extended Data Fig. 3 Model ablation study highlights relative importance of GEARS components under different generalization conditions. The ‘No Graph’ condition removes both the gene ontology graph and co-expression graph; ‘No GO Graph’ removes the gene ontology graph; ‘No Co-Express Graph’ removes the co-expression graph; ‘No Cross-gene’ removes the cross-gene MLP layer; ‘No Gene-specific Decoder’ removes the gene specific decoder MLP and uses a shared MLP instead; ‘MSE Loss’ switches from the auto-focus loss to the regular L2 loss. Four generalization conditions are considered: a, (1/1 Unseen) single-gene perturbations not seen experimentally perturbed at the time of training. b–d, (2/2 Unseen) two-gene perturbations in which both genes were not seen experimentally perturbed individually at the time of training (b), (1/2 Unseen) one of the two genes was not seen experimentally perturbed (c) or (0/2 Unseen) both genes have been seen experimentally perturbed (d). Performance is measured using the mean squared error in predicted postperturbation gene expression for the top 20 most differentially expressed genes. For all panels (a–d) the marker indicates the mean MSE over predictions made by models trained using 5 different training data splits (n=5). The error bars represent bootstrapped 95% CI.CC BY · PubMed Central
Extended Data Fig. 4 Model performance relationship with network connectivity. Each point in the scatter plot corresponds to a prediction made for a novel single-gene perturbation not seen at the time of training. The y-axis plots the pearson correlation between the true mean postperturbation differential expression over unperturbed control and the same predicted by GEARS. The x-axis measures the number of connections between the novel perturbed gene and other genes in the network that had been seen at the time of training. Error band corresponds to 95% CI.
Extended Data Fig. 4 Model performance relationship with network connectivity. Each point in the scatter plot corresponds to a prediction made for a novel single-gene perturbation not seen at the time of training. The y-axis plots the pearson correlation between the true mean postperturbation differential expression over unperturbed control and the same predicted by GEARS. The x-axis measures the number of connections between the novel perturbed gene and other genes in the network that had been seen at the time of training. Error band corresponds to 95% CI.CC BY · PubMed Central
Extended Data Fig. 5 Model performance at predicting genetic interaction (GI) scores. a, GI scores for the set of combinatorial perturbations that were defined as expressing a specific GI subtype phenotype in Norman et al. 2019. The gray dots correspond to GI scores computed using true postperturbation gene expression. The colored dots were computed using predicted postperturbation gene expression under three different models: GEARS, CPA and Naive models. The naive model here simply sums together the effects of single-gene perturbations. The metrics on the y-axis correspond to different GI scores and the colored dotted lines indicate the defined thresholds for determining if a combination is exhibiting a specific GI subtype phenotype. Both GEARS and CPA were trained using a leave-one-out testing approach for each of the 131 combinations. The black dashed line represents the minimum and maximum of all 131 values and the black solid line represents the mean. b, Scatter plots of GI scores for all 131 two-gene combinatorial perturbations from that dataset. The x-axis shows GI scores computed using true postperturbation gene expression and the y-axis shows scores computed using predicted postperturbation gene expression. The top row shows predictions made by GEARS and the bottom row shows predictions made by CPA. R2 refers to the coefficient of determination.
Extended Data Fig. 5 Model performance at predicting genetic interaction (GI) scores. a, GI scores for the set of combinatorial perturbations that were defined as expressing a specific GI subtype phenotype in Norman et al. 2019. The gray dots correspond to GI scores computed using true postperturbation gene expression. The colored dots were computed using predicted postperturbation gene expression under three different models: GEARS, CPA and Naive models. The naive model here simply sums together the effects of single-gene perturbations. The metrics on the y-axis correspond to different GI scores and the colored dotted lines indicate the defined thresholds for determining if a combination is exhibiting a specific GI subtype phenotype. Both GEARS and CPA were trained using a leave-one-out testing approach for each of the 131 combinations. The black dashed line represents the minimum and maximum of all 131 values and the black solid line represents the mean. b, Scatter plots of GI scores for all 131 two-gene combinatorial perturbations from that dataset. The x-axis shows GI scores computed using true postperturbation gene expression and the y-axis shows scores computed using predicted postperturbation gene expression. The top row shows predictions made by GEARS and the bottom row shows predictions made by CPA. R2 refers to the coefficient of determination.CC BY · PubMed Central
Extended Data Fig. 6 Model performance in predicting genetic interactions (GIs). a, Mean Square Error (MSE) in predicting non-additive combinatorial effects between the additive model which assumes that the effect of the combination is just the sum of the two known single-gene perturbation outcomes and GEARS predictions. MSE was measured on the 20 genes with the largest difference between true postperturbation expression following two-gene combinatorial perturbation and the additive prediction for that combination. GI subtypes (x-axis) were labelled without overlap as in Norman et al. 2019 (Synergy n=30, Suppression n=12, Redundancy n=8, Neomorphism n=13, Epistasis n=9). Bar plots represent the mean and error bars correspond to 95% CI. b, Top 10 accuracy in predicting GIs: Model accuracy in predicting the set of 10 strongest interactions for each GI subtype as determined using true expression. Marker represents mean and error bar represents 1SD for the random model which performs 1000 draws (n=1000). For other models, predictions from 3 trained models were used (n=3). c, Precision and recall in predicting GIs (n=3).
Extended Data Fig. 6 Model performance in predicting genetic interactions (GIs). a, Mean Square Error (MSE) in predicting non-additive combinatorial effects between the additive model which assumes that the effect of the combination is just the sum of the two known single-gene perturbation outcomes and GEARS predictions. MSE was measured on the 20 genes with the largest difference between true postperturbation expression following two-gene combinatorial perturbation and the additive prediction for that combination. GI subtypes (x-axis) were labelled without overlap as in Norman et al. 2019 (Synergy n=30, Suppression n=12, Redundancy n=8, Neomorphism n=13, Epistasis n=9). Bar plots represent the mean and error bars correspond to 95% CI. b, Top 10 accuracy in predicting GIs: Model accuracy in predicting the set of 10 strongest interactions for each GI subtype as determined using true expression. Marker represents mean and error bar represents 1SD for the random model which performs 1000 draws (n=1000). For other models, predictions from 3 trained models were used (n=3). c, Precision and recall in predicting GIs (n=3).CC BY · PubMed Central
Extended Data Fig. 7 Validation of GEARS predicted genetic interaction (GI) map using combinatorial cell fitness screen. a, Combinatorial cell fitness screen data was used for all pairwise combination of 92 genes leading to 4186 unique combinations. Using cell fitness, interactions were quantified as synergistic or suppressive. b, Combinations showing the strongest cell fitness effects were used to validate GEARS predictions. c, Combinatorial Perturb-seq data was available for 110 of these combinations. GEARS was trained on Perturb-Seq data to predict remaining 4076 perturbation outcomes. d, GEARS performs similar to experimental Perturb-Seq data in predicting strong genetic interaction outcomes for both strongly synergistic and suppressive interactions identified using cell fitness measurements. GI scores are z-normalized within each modality for comparison. Centreline represents mean. Whiskers represent last data point within 1.5x interquartile range below the first quantile and above the third quantile, outliers not shown. The p-values were computed using a one-sided t-test comparing the means of the two distributions.
Extended Data Fig. 7 Validation of GEARS predicted genetic interaction (GI) map using combinatorial cell fitness screen. a, Combinatorial cell fitness screen data was used for all pairwise combination of 92 genes leading to 4186 unique combinations. Using cell fitness, interactions were quantified as synergistic or suppressive. b, Combinations showing the strongest cell fitness effects were used to validate GEARS predictions. c, Combinatorial Perturb-seq data was available for 110 of these combinations. GEARS was trained on Perturb-Seq data to predict remaining 4076 perturbation outcomes. d, GEARS performs similar to experimental Perturb-Seq data in predicting strong genetic interaction outcomes for both strongly synergistic and suppressive interactions identified using cell fitness measurements. GI scores are z-normalized within each modality for comparison. Centreline represents mean. Whiskers represent last data point within 1.5x interquartile range below the first quantile and above the third quantile, outliers not shown. The p-values were computed using a one-sided t-test comparing the means of the two distributions.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 51 references

Research works printed but not linked: 3

  1. Wang, Y., Solus, L., Yang, K. & Uhler, C. Permutation-based causal inference algorithms with interventions. In Proc. 31st International Conference on Neural Information Processing Systems (Ed. von Luxburg, U. & Guyon, I.) 5824–5833 (Association for Computing Machinery, 2017).
  2. Hamilton, W., Ying, Z. & Leskovec, J. Inductive representation learning on large graphs. In Proc. 31st International Conference on Neural Information Processing Systems (Ed. von Luxburg, U. & Guyon, I.) 1025–1035 (Association for Computing Machinery, 2017).
  3. Kendall, A. & Gal, Y. What uncertainties do we need in Bayesian deep learning for computer vision? In Proc. 31st International Conference on Neural Information Processing Systems (Ed. von Luxburg, U. & Guyon, I.) 5580–5590 (Association for Computing Machinery, 2017).

Cited by

See all 368

Metadata

Authors
Roohani, Y., Huang, K., & Leskovec, J.
Affiliations
Stanford University (US)
Corresponding author
Leskovec, Jure
Venue
Nature Biotechnology
Published
2023-08-17
Volume
42
Issue
6
Pages
927-935
ISSN
  • 0733-222X (ISSN-L)
  • 1087-0156
  • 1546-1696
OpenAlex
W4385955324
Cited by
355
References
51
Citations by year
2026: 153 · 2025: 118 · 2024: 77 · 2023: 7
FWCI
31.63
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Biotechnology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • U.S. Department of Health & Human Services | NIH | Center for Information Technology (3U54HG010426-04S1)
  • GlaxoSmithKline
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
Keywords
gene-gene interaction, genetic interactions, knowledge graphs, deep learning, single-cell RNA sequencing, gene activation and repression, gears
MeSH terms
Deep Learning, Computer Simulation, Humans, Transcription, Genetic, Sequence Analysis, RNA, Computational Biology, Gene Regulatory Networks, Single-Cell Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Predicting transcriptional outcomes of novel multigene perturbations with GEARS. https://app.cochat.ai/papers/10.1038/s41587-023-01905-6
The paper
Roohani, Y., Huang, K., & Leskovec, J. (2024). Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nature Biotechnology. https://doi.org/10.1038/s41587-023-01905-6

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.