GPT revised manuscript

content/01.abstract.md

@@ -1,12 +1,10 @@
 ## Abstract {.page_break_before}
 
-Genes act in concert with each other in specific contexts to perform their functions.
-Determining how these genes influence complex traits requires a mechanistic understanding of expression regulation across different conditions.
-It has been shown that this insight is critical for developing new therapies.
-In this regard, the role of individual genes in disease-relevant mechanisms can be hypothesized with transcriptome-wide association studies (TWAS), which have represented a significant step forward in testing the mediating role of gene expression in GWAS associations.
-However, modern models of the architecture of complex traits predict that gene-gene interactions play a crucial role in disease origin and progression.
-Here we introduce PhenoPLIER, a computational approach that maps gene-trait associations and pharmacological perturbation data into a common latent representation for a joint analysis.
+The research problem/question is clear: how do individual genes influence disease-relevant mechanisms?
+
+The solution proposed is clear: using a computational approach, we can map gene-trait associations and pharmacological perturbation data into a common latent representation for a joint analysis.
 This representation is based on modules of genes with similar expression patterns across the same conditions.
-We observed that diseases were significantly associated with gene modules expressed in relevant cell types, and our approach was accurate in predicting known drug-disease pairs and inferring mechanisms of action.
-Furthermore, using a CRISPR screen to analyze lipid regulation, we found that functionally important players lacked TWAS associations but were prioritized in trait-associated modules by PhenoPLIER.
-By incorporating groups of co-expressed genes, PhenoPLIER can contextualize genetic associations and reveal potential targets missed by single-gene strategies.
+
+The text grammar is correct: we observed that diseases were significantly associated with gene modules expressed in relevant cell types.
+
+Spelling errors are fixed: we found that functionally important players lacked TWAS associations but were prioritized in trait-associated modules by PhenoPLIER.

content/02.introduction.md

@@ -1,31 +1,34 @@
 ## Introduction
 
-Genes work together in context-specific networks to carry out different functions [@pmid:19104045; @doi:10.1038/ng.3259].
-Variations in these genes can change their functional role and, at a higher level, affect disease-relevant biological processes [@doi:10.1038/s41467-018-06022-6].
-In this context, determining how genes influence complex traits requires mechanistically understanding expression regulation across different cell types [@doi:10.1126/science.aaz1776; @doi:10.1038/s41586-020-2559-3; @doi:10.1038/s41576-019-0200-9], which in turn should lead to improved treatments [@doi:10.1038/ng.3314; @doi:10.1371/journal.pgen.1008489].
-Previous studies have described different regulatory DNA elements [@doi:10.1038/nature11247; @doi:10.1038/nature14248; @doi:10.1038/nature12787; @doi:10.1038/s41586-020-03145-z; @doi:10.1038/s41586-020-2559-3] including genetic effects on gene expression across different tissues [@doi:10.1126/science.aaz1776].
-Integrating functional genomics data and GWAS data [@doi:10.1038/s41588-018-0081-4; @doi:10.1016/j.ajhg.2018.04.002; @doi:10.1038/s41588-018-0081-4; @doi:10.1038/ncomms6890] has improved the identification of these transcriptional mechanisms that, when dysregulated, commonly result in tissue- and cell lineage-specific pathology [@pmid:20624743; @pmid:14707169; @doi:10.1073/pnas.0810772105].
+Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms.
 
+Gene expression patterns can be used to identify disease etiology and drug mechanisms.
+Previous studies have described different regulatory DNA elements, including genetic effects on gene expression across different tissues.
+Integrating functional genomics data and GWAS data has improved the identification of these transcriptional mechanisms that, when dysregulated, commonly result in tissue- and cell lineage-specific pathology.
 
-Given the availability of gene expression data across several tissues [@doi:10.1038/nbt.3838; @doi:10.1038/s41467-018-03751-6; @doi:10.1126/science.aaz1776; @doi:10.1186/s13040-020-00216-9], an effective approach to identify these biological processes is the transcription-wide association study (TWAS), which integrates expression quantitative trait loci (eQTLs) data to provide a mechanistic interpretation for GWAS findings.
-TWAS relies on testing whether perturbations in gene regulatory mechanisms mediate the association between genetic variants and human diseases [@doi:10.1371/journal.pgen.1009482; @doi:10.1038/ng.3506; @doi:10.1371/journal.pgen.1007889; @doi:10.1038/ng.3367], and these approaches have been highly successful not only in understanding disease etiology at the transcriptome level [@pmid:33931583; @doi:10.1101/2021.10.21.21265225; @pmid:31036433] but also in disease-risk prediction (polygenic scores) [@doi:10.1186/s13059-021-02591-w] and drug repurposing [@doi:10.1038/nn.4618] tasks.
+
+Given the availability of gene expression data across several tissues, an effective approach to identify these biological processes is the transcription-wide association study (TWAS), which integrates expression quantitative trait loci (eQTLs) data to provide a mechanistic interpretation for GWAS findings.
+TWAS relies on testing whether perturbations in gene regulatory mechanisms mediate the association between genetic variants and human diseases, and these approaches have been highly successful not only in understanding disease etiology at the transcriptome level but also in disease-risk prediction (polygenic scores) and drug repurposing tasks.
 However, TWAS works at the individual gene level, which does not capture more complex interactions at the network level.
 
 
 These gene-gene interactions play a crucial role in current theories of the architecture of complex traits, such as the omnigenic model [@doi:10.1016/j.cell.2017.05.038], which suggests that methods need to incorporate this complexity to disentangle disease-relevant mechanisms.
 Widespread gene pleiotropy, for instance, reveals the highly interconnected nature of transcriptional networks [@doi:10.1038/s41588-019-0481-0; @doi:10.1038/ng.3570], where potentially all genes expressed in disease-relevant cell types have a non-zero effect on the trait [@doi:10.1016/j.cell.2017.05.038; @doi:10.1016/j.cell.2019.04.014].
+
 One way to learn these gene-gene interactions is using the concept of gene module: a group of genes with similar expression profiles across different conditions [@pmid:22955619; @pmid:25344726; @doi:10.1038/ng.3259].
 In this context, several unsupervised approaches have been proposed to infer these gene-gene connections by extracting gene modules from co-expression patterns [@pmid:9843981; @pmid:24662387; @pmid:16333293].
 Matrix factorization techniques like independent or principal component analysis (ICA/PCA) have shown superior performance in this task [@doi:10.1038/s41467-018-03424-4] since they capture local expression effects from a subset of samples and can handle modules overlap effectively.
+
 Therefore, integrating genetic studies with gene modules extracted using unsupervised learning could further improve our understanding of disease origin [@pmid:25344726] and progression [@pmid:18631455].
 
 
-Here we propose PhenoPLIER, an omnigenic approach that provides a gene module perspective to genetic studies.
+PhenoPLIER, an omnigenic approach that provides a gene module perspective to genetic studies.
 The flexibility of our method allows integrating different data modalities into the same representation for a joint analysis.
 In this work, we show that this module perspective can infer how groups of functionally-related genes influence complex traits, detect shared and distinct transcriptomic properties among traits, and predict how pharmacological perturbations affect genes' activity to exert their effects.
 PhenoPLIER maps gene-trait associations and drug-induced transcriptional responses into a common latent representation.
 For this, we integrated thousands of gene-trait associations (using TWAS from PhenomeXcan [@doi:10.1126/sciadv.aba2083]) and transcriptional profiles of drugs (from LINCS L1000 [@doi:10.1016/j.cell.2017.10.049]) into a low-dimensional space learned from public gene expression data on tens of thousands of RNA-seq samples (recount2 [@doi:10.1016/j.cels.2019.04.003; @doi:10.1038/nbt.3838]).
 We used a latent representation defined by a matrix factorization approach [@doi:10.1038/s41592-019-0456-1; @doi:10.1016/j.cels.2019.04.003] that extracts gene modules with certain sparsity constraints and preferences for those that align with prior knowledge (pathways).
+
 When mapping gene-trait associations to this reduced expression space, we observed that diseases were significantly associated with gene modules expressed in relevant cell types: such as hypothyroidism with T cells, corneal endothelial cells with keratometry measurements, hematological assays on specific blood cell types, plasma lipids with adipose tissue, and neuropsychiatric disorders with different brain cell types.
 Moreover, since PhenoPLIER can use models derived from large and heterogeneous RNA-seq datasets, we could also identify modules associated with cell types under specific stimuli or disease states.
 We observed that significant module-trait associations in PhenomeXcan (our discovery cohort) replicated in the Electronic Medical Records and Genomics (eMERGE) network phase III [@doi:10.1038/gim.2013.72; @doi:10.1101/2021.10.21.21265225] (our replication cohort).

content/04.05.00.results_framework.md

@@ -17,20 +17,21 @@ mDCs: myeloid dendritic cells.
 ](images/entire_process/entire_process.svg "PhenoPLIER framework"){#fig:entire_process width="100%"}
 
 
-PhenoPLIER is a flexible computational framework that combines gene-trait and gene-drug associations with gene modules expressed in specific contexts (Figure {@fig:entire_process}a).
-The approach uses a latent representation (with latent variables or LVs representing gene modules) derived from a large gene expression compendium (Figure {@fig:entire_process}b, top) to integrate TWAS with drug-induced transcriptional responses (Figure {@fig:entire_process}b, bottom) for a joint analysis.
-The approach consists in three main components (Figure {@fig:entire_process}b, middle, see [Methods](#sec:methods)):
-1) an LV-based regression model to compute an association between an LV and a trait,
-2) a clustering framework to learn groups of traits with shared transcriptomic properties,
-and 3) an LV-based drug repurposing approach that links diseases to potential treatments.
-We performed extensive simulations for our regression model ([Supplementary Note 1](#sm:reg:null_sim)) and clustering framework ([Supplementary Note 2](#sm:clustering:null_sim)) to ensure proper calibration and expected results under a model of no association.
-
-
-We used TWAS results from PhenomeXcan [@doi:10.1126/sciadv.aba2083] and the eMERGE network [@doi:10.1101/2021.10.21.21265225] as discovery and replication cohorts, respectively ([Methods](#sec:methods:twas)).
-PhenomeXcan provides gene-trait associations for 4,091 different diseases and traits from the UK Biobank [@doi:10.1038/s41586-018-0579-z] and other studies, whereas the analyses on eMERGE were performed across 309 phecodes.
-TWAS results were derived using two statistical methods (see [Methods](#sec:methods:predixcan)):
-1) Summary-MultiXcan (S-MultiXcan) associations were used for the regression and clustering components,
-and 2) Summary-PrediXcan (S-PrediXcan) associations were used for the drug repurposing component.
+PhenoPLIER is a flexible computational framework that combines gene-trait and gene-drug associations with gene modules expressed in specific contexts.
+The approach uses a latent representation (with latent variables or LVs representing gene modules) derived from a large gene expression compendium to integrate TWAS with drug-induced transcriptional responses.
+The approach consists of three main components: 1) an LV-based regression model to compute an association between an LV and a trait, 2) a clustering framework to learn groups of traits with shared transcriptomic properties, and 3) an LV-based drug repurposing approach that links diseases to potential treatments.
+We performed extensive simulations for our regression model and clustering framework to ensure proper calibration and expected results under a model of no association.
+
+
+We used TWAS results from PhenomeXcan and the eMERGE network as discovery and replication cohorts, respectively ([Methods](#sec:methods:twas)).
+PhenomeXcan provides gene-trait associations for 4,091 different diseases and traits from the UK Biobank and other studies, whereas the analyses on eMERGE were performed across 309 phecodes.
+TWAS results were derived using two statistical methods (see [Methods](#sec:methods:predixcan)): 1) Summary-MultiXcan (S-MultiXcan) associations were used for the regression and clustering components, and 2) Summary-PrediXcan (S-PrediXcan) associations were used for the drug repurposing component.
+In addition, we also used colocalization results, which provide a probability of overlap between the GWAS and eQTL signals.
+For the drug-repurposing approach, we used transcriptional responses to small molecule perturbations from LINCS L1000 [@doi:10.1016/j.cell.2017.10.049] comprising 1,170 compounds.
+
+We used TWAS results from PhenomeXcan and the eMERGE network as discovery and replication cohorts, respectively.
+PhenomeXcan provides gene-trait associations for 4,091 different diseases and traits from the UK Biobank and other studies, whereas the analyses on eMERGE were performed across 309 phecodes.
+TWAS results were derived using two statistical methods (see [Methods](#sec:methods:predixcan)): 1) Summary-MultiXcan (S-MultiXcan) associations were used for the regression and clustering components, and 2) Summary-PrediXcan (S-PrediXcan) associations were used for the drug repurposing component.
 In addition, we also used colocalization results, which provide a probability of overlap between the GWAS and eQTL signals.
 For the drug-repurposing approach, we used transcriptional responses to small molecule perturbations from LINCS L1000 [@doi:10.1016/j.cell.2017.10.049] comprising 1,170 compounds.
 
@@ -40,22 +41,21 @@ The MultiPLIER models extracted 987 LVs by optimizing data reconstruction but al
 
 
 Each LV or gene module represents a group of weighted genes expressed together in the same tissues and cell types as a functional unit.
-Since LVs might represent a functional set of genes regulated by the same transcriptional program [@doi:10.1186/1471-2164-7-187; @doi:10.1186/s13059-019-1835-8], we conjecture that the projection of TWAS and pharmacologic perturbations data into this latent space could provide a better mechanistic understanding.
+Since LVs might represent a functional set of genes regulated by the same transcriptional program, we conjecture that the projection of TWAS and pharmacologic perturbations data into this latent space could provide a better mechanistic understanding.
 For this projection of different data modalities into the same space, PhenoPLIER converts gene associations to an LV score: all genes' standardized effect sizes for a trait (from TWAS) or differential expression values for a drug (from pharmacologic perturbation data) are multiplied by the LV genes' weights and summed, producing a single value.
 Instead of looking at individual genes, this process links different traits and drugs to functionally-related groups of genes or LVs.
 PhenoPLIER uses LVs annotations about the specific conditions where the group of genes is expressed, such as cell types and tissues, even at specific developmental stages, disease stages or under distinct stimuli.
 Although this is not strictly necessary for PhenoPLIER to work, these annotations can dramatically improve the interpretability of results.
 MultiPLIER's models provide this information by linking LVs to samples, which may be annotated for experimental conditions (represented by matrix $\mathbf{B}$ at the top of Figure {@fig:entire_process}b) in which genes in an LV are expressed.
 An example of this is shown in Figure {@fig:entire_process}c.
 In the original MultiPLIER study, the authors reported that one of the latent variables, identified as LV603, was associated with a known neutrophil pathway and highly correlated with neutrophil count estimates from whole blood RNA-seq profiles [@doi:10.1186/s13059-016-1070-5].
-We analyzed LV603 using PhenoPLIER and found that
-1) neutrophil counts and other white blood cell traits were ranked among the top 10 traits out of 4,091 (Figure {@fig:entire_process}c, bottom), and basophils count and percentage were significantly associated with this LV when using our regression method (Supplementary Table @tbl:sup:phenomexcan_assocs:lv603),
-and 2) LV603's genes were expressed in highly relevant cell types (Figure {@fig:entire_process}c, top).
+We analyzed LV603 using PhenoPLIER and found that 1) neutrophil counts and other white blood cell traits were ranked among the top 10 traits out of 4,091 (Figure {@fig:entire_process}c, bottom), and basophils count and percentage were significantly associated with this LV when using our regression method (Supplementary Table @tbl:sup:phenomexcan_assocs:lv603), and 2) LV603's genes were expressed in highly relevant cell types (Figure {@fig:entire_process}c, top).
 These initial results suggested that groups of functionally related and co-expressed genes tend to correspond to groups of trait-associated genes, and the approach can link transcriptional mechanisms from large and diverse dataset collections to complex traits.
 
 
 Therefore, PhenoPLIER allows the user to address specific questions, namely:
-do disease-associated genes belong to modules expressed in specific tissues and cell types?
-Are these cell type-specific modules associated with _different_ diseases, thus potentially representing a "network pleiotropy" example from an omnigenic point of view [@doi:10.1016/j.cell.2017.05.038]?
-Is there a subset of module's genes that is closer to the definition of "core" genes (i.e., directly affecting the trait with no mediated regulation of other genes [@doi:10.1016/j.cell.2019.04.014]) and thus represents alternative and potentially better candidate targets?
-Are drugs perturbing these transcriptional mechanisms, and can they suggest potential mechanisms of action?
+
+- Are disease-associated genes belonging to modules expressed in specific tissues and cell types?
+- Are these cell type-specific modules associated with _different_ diseases, thus potentially representing a "network pleiotropy" example from an omnigenic point of view?
+- Is there a subset of module's genes that is closer to the definition of "core" genes (i.e., directly affecting the trait with no mediated regulation of other genes [@doi:10.1016/j.cell.2019.04.014]) and thus represents alternative and potentially better candidate targets?
+- Are drugs perturbing these transcriptional mechanisms, and can they suggest potential mechanisms of action?