Valid inference for machine learning-assisted genome-wide association studies.

Area of Science:

• Human genetics
• Statistical genetics
• Bioinformatics

Background:

• Machine learning (ML) is increasingly used in human genetics for complex trait analysis.
• ML-assisted genome-wide association studies (GWAS) impute phenotypes but face validity concerns.
• Existing ML-assisted GWAS methods carry risks of false-positive associations.

Purpose of the Study:

• To evaluate the validity of ML-assisted GWAS associations.
• To introduce a robust statistical framework, Post-Prediction GWAS (POP-GWAS), for analyzing ML-imputed outcomes.
• To ensure valid and powerful statistical inference in complex trait genetics.

Main Methods:

• Developed the Post-Prediction GWAS (POP-GWAS) statistical framework.
• Redesigned GWAS methodology for ML-imputed outcomes.
• Required only GWAS summary statistics as input, irrespective of ML imputation quality or algorithm.

Main Results:

• Identified pervasive risks for false-positive associations in current ML-assisted GWAS.
• Successfully employed POP-GWAS for a GWAS of bone mineral density across 14 skeletal sites.
• Discovered 89 novel genetic loci associated with bone mineral density, revealing skeletal site-specific genetic architecture.

Conclusions:

• POP-GWAS provides a statistically rigorous solution for ML-assisted GWAS.
• The framework ensures valid inference, addressing limitations of previous ML-imputed GWAS approaches.
• POP-GWAS is a robust tool for future complex trait genetics research utilizing ML.