Exploring the Link Between Additive Heritability and Prediction Accuracy From a Ridge Regression Perspective

Area of Science:

• Genetics and Bioinformatics
• Statistical Genomics
• Machine Learning in Biology

Background:

• Genome-Wide Association Studies (GWAS) explain limited heritability for complex human traits.
• Genomic heritability quantifies variance explained by SNPs using mixed models.
• A gap exists in understanding the relationship between heritability estimation and predictive accuracy.

Purpose of the Study:

• To explore the connection between mixed models and ridge regression from a machine learning perspective.
• To propose and validate a machine learning-based method for estimating genomic heritability.
• To derive formulae for prediction accuracy based on population size and SNP count.

Main Methods:

• Utilized ridge regression and Generalized Cross Validation (GCV) for a predictive estimation of genomic heritability.
• Established consistency between the proposed GCV method and classical mixed model estimations.
• Derived analytical formulae linking prediction accuracy to the ratio of population size (n) to SNP count (p).

Main Results:

• The proposed ridge regression and GCV approach provides a valid alternative for estimating genomic heritability.
• Derived formulae demonstrate that prediction accuracy is not solely dependent on heritability, especially when p > n.
• Validation using simulated and UK Biobank data confirmed the efficacy of the GCV estimation and prediction formulae.

Conclusions:

• Machine learning, specifically ridge regression and GCV, offers a robust framework for genomic heritability estimation.
• Prediction accuracy in genomic studies is influenced by the relationship between sample size and the number of SNPs.
• High heritability does not guarantee accurate genomic prediction when the number of SNPs exceeds the population size.