Causal Inference in Transcriptome-Wide Association Studies with Invalid Instruments and GWAS Summary Data.

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Transcriptome-wide association studies (TWAS) integrate genome-wide association study (GWAS) and gene expression (eQTL) data to identify causal genes.
• Existing TWAS methods, often based on two-stage least squares (2SLS), can be unreliable due to invalid genetic instruments (SNPs) and pleiotropy.
• Identifying causal genes for low-density lipoprotein cholesterol (LDL) is critical for treating hyperlipidemia and cardiovascular diseases.

Purpose of the Study:

• To develop a robust and efficient statistical method for identifying causal genes using GWAS and eQTL data.
• To address limitations of standard TWAS, specifically accounting for invalid instrumental variables and hidden confounding.
• To apply the novel method to identify causal genes for LDL cholesterol.

Main Methods:

• Proposed a novel two-stage constrained maximum likelihood (2ScML) method, extending 2SLS.
• Developed the method for both individual-level and summary-level GWAS data, applicable to two-sample TWAS designs.
• Utilized sparse regression techniques to enhance robustness against invalid instrumental variables.

Main Results:

• The 2ScML method provides asymptotically valid statistical inference for causal effects.
• Demonstrated superior finite-sample performance compared to standard 2SLS/TWAS and Mendelian randomization (MR) methods.
• Successfully applied the method to identify putative causal genes for LDL cholesterol using large-scale lipid GWAS and eQTL data.

Conclusions:

• The proposed 2ScML method offers a more reliable approach for causal gene discovery in TWAS.
• The method is broadly applicable, especially for summary-level data where other robust methods are not suitable.
• This work advances the identification of genes influencing complex traits like LDL cholesterol, aiding therapeutic development.