Genetic association studies using disease liabilities from deep neural networks

Area of Science:

• Genetics
• Computational Biology
• Biostatistics

Background:

• Case-control studies are common for binary trait genetics but struggle with prospective data.
• Long-term cohort studies face challenges with absent or evolving health outcomes.
• Deep patient phenotyping offers a rich data source for genetic analysis.

Purpose of the Study:

• To propose novel genome-wide association study (GWAS) methods leveraging disease liabilities from deep phenotyping.
• To compare these new methods against traditional case-control approaches.
• To assess the utility of liability-based polygenic risk scores for predicting future disease onset.

Main Methods:

• Developed two new GWAS methods: liability and meta, utilizing deep patient phenotyping.
• Analyzed 38 common traits in approximately 300,000 UK Biobank participants.
• Validated findings in external genome-wide association studies (GWAS) and assessed robustness with imputed phenotype accuracy.

Main Results:

• Identified a greater number of genetic loci compared to conventional case-control methods.
• Achieved high replication rates in larger external genome-wide association studies (GWAS).
• Demonstrated that polygenic risk scores based on disease liabilities effectively predicted newly diagnosed cases.

Conclusions:

• Integrating high-dimensional phenotypic data enhances genetic association studies.
• The proposed liability-based methods capture disease-relevant genetic architecture effectively.
• Deep neural networks combined with deep phenotyping improve genetic discovery and prediction.