Strategies for developing prediction models from genome-wide association studies

Area of Science:

• Genetics and Bioinformatics
• Disease Risk Prediction Modeling

Background:

• Genome-wide association studies (GWASs) identify numerous single nucleotide polymorphisms (SNPs) linked to complex diseases.
• Current risk prediction models using SNPs have limited accuracy, prompting research into improved model building strategies.
• Including a large number of SNPs is hypothesized to enhance predictive performance.

Purpose of the Study:

• To investigate methods for improving the discriminatory accuracy of genetic risk prediction models, measured by the area under the receiver operating characteristic curve (AUC).
• To evaluate different aspects of prediction model building, including SNP selection procedures, data allocation, selection criteria, SNP quantity, and estimation methods.
• To compare single-phase versus two-phase procedures for SNP selection and estimation.

Main Methods:

• Utilized realistic genetic effect size, allele frequency, and linkage disequilibrium (LD) data from GWAS for Crohn's disease and prostate cancer.
• Employed theoretical calculations and simulations to estimate AUC for various model building strategies.
• Assessed SNP selection based on P-value thresholding versus ranking, and compared univariate (marginal) versus multivariate estimation.

Main Results:

• Empirical risk models with 10,000 cases and controls showed significantly lower AUC than theoretically possible.
• The single-phase procedure for SNP selection and estimation outperformed the two-phase procedure in terms of AUC.
• Univariate (marginal) estimation was superior to multivariate estimation, especially in the presence of LD.

Conclusions:

• Initial SNP selection is the most critical step in building accurate genetic risk prediction models.
• For complex diseases and sample sizes of 10,000 or fewer, limiting the number of selected SNPs to tens or hundreds is recommended.
• A single-phase approach for SNP selection and estimation is more effective than a two-phase approach for improving predictive accuracy.