MOSTWAS: Multi-Omic Strategies for Transcriptome-Wide Association Studies

Abstract

Traditional predictive models for transcriptome-wide association studies (TWAS) consider only single nucleotide polymorphisms (SNPs) local to genes of interest and perform parameter shrinkage with a regularization process. These approaches ignore the effect of distal-SNPs or possible effects underlying the SNP-gene association. Here, we outline multi-omic strategies for transcriptome imputation from germline genetics for testing gene-trait associations by prioritizing distal-SNPs to the gene of interest. In one extension, we identify mediating biomarkers (CpG sites, microRNAs, and transcription factors) highly associated with gene expression and train predictive models for these mediators using their local SNPs. Imputed values for mediators are then incorporated into the final model as fixed effects with local SNPs to the gene included as regularized effects. In the second extension, we assess distal-eSNPs (SNPs in eQTLs) for their mediation effect through mediators local to these distal-eSNPs. Highly mediated distal-eSNPs are then included in the eventual transcriptomic prediction model. We show considerable gains in percent variance explained of gene expression and TWAS power to detect gene-trait associations using simulation analysis and real data applications with TCGA breast cancer data and in ROS/MAP brain tissue data. This integrative approach to transcriptome-wide imputation and association studies aids in understanding the complex interactions underlying genetic regulation within a tissue and identifying important risk genes for various traits and disorders.the matrix X Mj be the local-genotype dosages in a 500 kilobase window around mediator j, 1 ≤ j ≤ m G . Furthermore, let M j be the intensity of mediator j (methylation M -value if j is a CpG site, expression if j is a miRNA or a gene, etc). Prior to any modeling, we scale Y G and all M j , 1 ≤ j ≤ m G to zero mean and unit variance. We also residualize M j , 1 ≤ j ≤ m G and Y G with the covariate matrix X C to account for population stratification using principal components of the global genotype matrix and relevant clinical covariates to obtainM j , 1 ≤ j ≤ m G andỸ G .Transcriptome prediction in MeTWAS draws from two-step regression, as summarized in Figure 1.First, in the training set for a given training-test split, for 1 ≤ j ≤ m G , we model the residualized intensitỹ M j of training-set specific mediator j with the following additive model:where w j is the effect-sizes of the SNPs in X † Mj onM j in the training set. As in traditional transcriptomic imputation models [3,4], we findŵ j using one of the two following methods with the largest predicted adjusted R 2 : (1) elastic net regression with mixing parameter α = 0.5 and λ tuned over 5-fold cross validation using glmnet [19], or (2) linear mixed modeling assuming random effects for X Mj using rrBLUP [20]. Only significantly heritable (default P < 0.05 for the likelihood ratio test) [21] and well-cross validated (default R 2 ≥ 0.01) expression models are considered.For all j, using these opti...

Genetic control of fetal placental genomics contributes to development of health and disease

“…MOSTWAS also allows for hypothesis generation for regulation of TWAS-detected genes, through distal mediating biomarkers, like transcription factors, miRNAs, or products downstream of CpG methylation islands 34 . Our computational results prioritized 89 GTAs with strong distal associations.…”

“…Similarly, of the 6 childhood BMI-associated genes identified by Peng et al, only 1 had a significant model in ELGAN and showed no association with the trait; there were no overlaps with childhood obesityassociated genes from Peng et al 10 . We hypothesize that minimal overlap with susceptibility genes identified by Peng et al is due to differing eQTL architectures in the datasets and different inclusion criteria for significant gene expression models 10,34,44,45 .…”

“…We leveraged gene expression, CpG methylation, and miRNA expression data from fetal-side placenta tissue from the Extremely Low Gestational Age Newborn (ELGAN) Cohort Study 33 . We trained predictive models of gene expression, enriched for distal SNPs using MOSTWAS, a recent TWAS extension that integrates multi-omic data 34 . Re-analyzing 40 GWAS of European-ancestry subjects from large consortia [35][36][37][38][39] , we performed a series of TWAS for non-communicable health traits and disorders that may be influenced by the placenta to identify GTAs and functional hypotheses for regulation (Figure 2).…”

Evidence for the placenta-brain axis: multi-omic kernel aggregation predicts intellectual and social impairment in children born extremely preterm

“…Recent work in transcriptome-wide association studies (TWAS) are a promising tool that aggregates genetics and transcriptomics to identify candidate trait-associated genes [110,111]. Incorporating information from regulatory biomarkers, like transcription factors and miRNAs, into TWAS increases study power to generate hypotheses about regulation [112,113]. Given our observations in this analysis and the number of the integrated molecular datasets, we believe that the ELGAN study can be used to train predictive models for placental transcriptomics from genetics, enriched for regulatory elements [113].…”

“…Incorporating information from regulatory biomarkers, like transcription factors and miRNAs, into TWAS increases study power to generate hypotheses about regulation [112,113]. Given our observations in this analysis and the number of the integrated molecular datasets, we believe that the ELGAN study can be used to train predictive models for placental transcriptomics from genetics, enriched for regulatory elements [113]. These transcriptomic models can then be applied to genome-wide association study cohorts to study the regulation of gene-trait associations in the placenta.…”