Machine Learning Predictor Captures Most Heritable Height Variance Using 20,000 SNPs
A team from Michigan State University and collaborating institutions developed genomic predictors for height, heel bone density, and educational attainment using machine learning applied to UK Biobank data covering almost 500,000 individuals. The height predictor, built on approximately 20,000 activated SNPs, explained around 40 percent of total variance in validation data not used for training. Predicted heights correlated 0.65 with actual heights, placing most individuals within a few centimeters of their true measurement. This proportion of explained variance matches the estimated common SNP heritability from genome-wide complex trait analysis and appears close to the asymptotic value as sample size increases, indicating the team captured most of the heritability attributable to SNPs. The predictor thereby closes the gap between prediction R-squared and common SNP heritability. Bone density and educational attainment predictors explained 20 percent and 9 percent of variance, respectively. The researchers validated their findings using additional datasets and SNPs identified in earlier genome-wide association studies.
The results demonstrate that common variant genetic architecture for height can be accurately characterized, with prediction performance reaching the theoretical ceiling of SNP heritability rather than leaving unexplained variance on the table.
Written by the Genomes desk from the primary source linked above and checked against it. Research use only; not medical advice. Corrections: [email protected].