Trait imputation enhances nonlinear genetic prediction for some traits

Area of Science:

• Genetics
• Bioinformatics
• Biomedical Research

Background:

• Biobanks contain vast genetic and phenotypic data, crucial for biomedical research.
• Missing phenotype data is a major limitation in utilizing biobank resources effectively.
• Developing accurate genetic prediction models, like polygenic (risk) scores (PGS), is essential for personalized medicine.

Purpose of the Study:

• To impute missing phenotypes in large genotype datasets using genome-wide association studies (GWAS) summary data.
• To integrate imputed phenotypes with existing complete datasets to build improved nonlinear polygenic (risk) score models.
• To evaluate the predictive performance of nonlinear models trained with imputed phenotypes compared to traditional methods.

Main Methods:

• Leveraged large genotype datasets (e.g., UK Biobank) lacking specific phenotypes.
• Utilized GWAS summary statistics to impute missing phenotype data for individuals.
• Trained nonlinear prediction models using both observed and imputed phenotype data.
• Developed an ensemble model to integrate predictions from multiple nonlinear models.
• Compared the R-squared values of models trained with imputed data versus observed data.

Main Results:

• Ensemble models trained with imputed phenotypes achieved higher prediction accuracy (R2) than models using only small, complete observed datasets.
• For two out of seven traits, nonlinear models trained with imputed phenotypes outperformed direct use of imputed phenotypes as polygenic (risk) scores.
• For the remaining five traits, no significant improvement was observed when using imputed traits directly as PGS.
• Demonstrated the potential of imputation and nonlinear modeling for enhancing genetic prediction accuracy.

Conclusions:

• Imputing missing phenotypes from large genotype datasets is a viable strategy to enhance biomedical research.
• Accounting for nonlinear genetic relationships through advanced modeling can improve polygenic (risk) score accuracy for specific traits.
• This methodology offers a promising approach to overcome data limitations in biobanks and advance genetic prediction.