Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations

Shuang Song, Lijun Wang, Lin Hou, Jun S. LiuNature Communications 2024 · DOI

TL;DR

This paper introduces MTWAS, a novel statistical framework that improves the prediction of gene expression from genetic data by explicitly modeling both shared (cross-tissue) and unique (tissue-specific) genetic effects. By doing so, MTWAS achieves significantly higher prediction accuracy across diverse tissues and cell types, leading to more robust identification of genes associated with complex traits and diseases.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) have identified numerous genetic loci associated with complex diseases, but pinpointing the exact genes and mechanisms responsible remains a significant challenge, especially for variants in non-coding regions. Transcriptome-wide association studies (TWAS) offer a powerful approach by linking genetic variants to gene expression and then to disease traits. However, existing TWAS methods often struggle with limited sample sizes in some tissues and tend to either treat each tissue independently or assume that genetic effects are shared across all tissues. This overlooks the reality that regulatory effects can be highly tissue-specific. The paper aims to bridge this gap by developing a method that can effectively leverage information from multiple tissues, distinguishing between shared and tissue-specific genetic influences to improve gene expression prediction and subsequent gene-trait association discovery.

Methods

MTWAS employs a three-step prediction stage:

  1. Non-parametric Imputation: To address missing gene expression data in certain tissues, MTWAS uses a non-parametric imputation method (MissForest) to leverage information from other correlated tissues. This step does not use genotype data.
  2. Cross-Tissue eQTL (ct-eQTL) Detection: Principal Component Analysis (PCA) is applied to the imputed expression matrix to capture overall cross-tissue expression patterns. These principal components are then regressed against cis-SNP genotypes to identify genetic variants that influence gene expression across multiple tissues. The selection of these ct-eQTLs is guided by the extended Bayesian Information Criterion (EBIC).
  3. Tissue-Specific eQTL (ts-eQTL) Detection: After accounting for the effects of ct-eQTLs, a stepwise sparse linear regression method (SODA) is used to identify genetic variants that specifically affect gene expression within individual tissues. EBIC is also used here for variable selection.

The estimated effects of both ct-eQTLs and ts-eQTLs are then combined to predict gene expression. For the association stage, MTWAS derives an explicit form for its test statistic using these estimated effects, GWAS summary statistics, and a reference LD matrix.

Datasets

  • GTEx v8: A comprehensive dataset of gene expression and genotype data across 47 human tissues. Used extensively for training and evaluating MTWAS against other methods.
  • DICE: A dataset of bulk RNA sequencing data from 13 types of immune cells under two activation conditions. Used to evaluate MTWAS on immune cell types.
  • OneK1K: A single-cell RNA sequencing dataset of peripheral blood mononuclear cells from 982 donors, comprising 14 immune cell types. Used to evaluate MTWAS on single-cell data.
  • UK Biobank (UKBB): GWAS summary statistics for 84 self-reported cancer and non-cancer illness phenotypes. Used to apply MTWAS to identify gene-trait associations for complex diseases.
  • GEUVADIS: Lymphoblastoid cell lines from 373 European individuals, used for an independent replication study.

Key Insights and Ideas

  • Tissue-specific regulation matters: The core idea is that genetic effects on gene expression are not uniform across all tissues. MTWAS's ability to distinguish between shared and tissue-specific effects is crucial for accurate prediction and association.
  • Imputation enhances data utility: By imputing missing expression data, MTWAS effectively increases the usable sample size and leverages cross-tissue correlations, particularly beneficial for tissues with limited sample sizes.
  • EBIC for robust variable selection: The use of the extended Bayesian Information Criterion (EBIC) is highlighted as a superior criterion for variable selection in high-dimensional settings, contributing to MTWAS's improved performance.
  • ct-eQTLs vs. ts-eQTLs provide biological insights: Partitioning eQTLs into cross-tissue and tissue-specific categories allows for a deeper understanding of regulatory mechanisms, with ct-eQTLs potentially linked to more fundamental cellular processes and ts-eQTLs to cell-type-specific functions.

Key Results

  • Improved Prediction Accuracy: MTWAS consistently outperformed existing methods (PrediXcan, TIGAR, UTMOST) in predicting gene expression across all 47 GTEx tissues, achieving an average improvement in prediction R² of 47.4% over PrediXcan. This improvement was more pronounced in tissues with smaller sample sizes.
  • More Predictable Genes: MTWAS identified a significantly higher number of predictable genes compared to other methods, both under a common criterion (R² > 0.01) and a stringent criterion (FDR < 0.05).
  • Replication Success: The improvements observed with MTWAS were replicated on independent datasets (DICE, OneK1K, GEUVADIS), including single-cell RNA-seq data.
  • Application to GWAS: Applying MTWAS to 84 UK Biobank GWAS studies revealed novel gene-trait associations, including unique findings for conditions like heart valve problems and irritable bowel syndrome, where other methods did not identify significant genes.
  • Computational Efficiency: MTWAS was found to be computationally efficient, taking approximately half the time of UTMOST for a comparable analysis on a chromosome.

Limitations and Future Work

  • Causality: The paper acknowledges that TWAS, including MTWAS, identifies potential functional links but does not definitively prove causality. Further experimental validation is needed.
  • Fine-mapping Integration: The authors suggest integrating fine-mapping methods (like SuSiE or DAP-K) into the MTWAS framework to prioritize causal genes, which could be a promising future direction.
  • scRNA-seq Tailoring: While MTWAS works with pseudo-bulk single-cell data, the authors suggest developing prediction models specifically tailored for scRNA-seq data for future work.
  • Interaction Terms: The current implementation of ts-eQTL identification does not explicitly consider interaction terms, though the imputed data might implicitly capture some non-linear information.

Summary updated

Abstract

TWAS have shown great promise in extending GWAS loci to a functional understanding of disease mechanisms. In an effort to fully unleash the TWAS and GWAS information, we propose MTWAS, a statistical framework that partitions and aggregates cross-tissue and tissue-specific genetic effects in identifying gene-trait associations. We introduce a non-parametric imputation strategy to augment the inaccessible tissues, accommodating complex interactions and non-linear expression data structures across various tissues. We further classify eQTLs into cross-tissue eQTLs and tissue-specific eQTLs via a stepwise procedure based on the extended Bayesian information criterion, which is consistent under high-dimensional settings. We show that MTWAS significantly improves the prediction accuracy across all 47 tissues of the GTEx dataset, compared with other single-tissue and multi-tissue methods, such as PrediXcan, TIGAR, and UTMOST. Applying MTWAS to the DICE and OneK1K datasets with bulk and single-cell RNA sequencing data on immune cell types showcases consistent improvements in prediction accuracy. MTWAS also identifies more predictable genes, and the improvement can be replicated with independent studies. We apply MTWAS to 84 UK Biobank GWAS studies, which provides insights into disease etiology.
BibTeX
@inproceedings{song2024partitioning,
  author = {Shuang Song and Lijun Wang and Lin Hou and Jun S. Liu},
  title = {{Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-49924-4},
}
Figure