Human genotype-to-phenotype predictions: Boosting accuracy with nonlinear models.

Area of Science:

• Human Genetics
• Computational Biology
• Statistical Genomics

Background:

• Genotype-to-phenotype prediction is a key challenge in human genetics.
• Large genomic datasets and machine learning tools enable complex phenotype prediction models.

Purpose of the Study:

• To evaluate nonlinear predictive models against linear models for phenotype prediction.
• To develop a more efficient SNP selection procedure than conventional GWAS.
• To determine if ensembling models enhances prediction accuracy.

Main Methods:

• Comparison of boosted decision trees (nonlinear) with linear models.
• Development of a sequential feature importance estimation with decision trees for SNP selection.
• Ensembling of individual linear and nonlinear models.

Main Results:

• Nonlinear models show improved phenotype prediction with rich covariates.
• The proposed SNP selection method yields more compact and informative SNP sets than GWAS.
• Ensembling models achieves the highest prediction accuracy for phenotypes like asthma and hypothyroidism.

Conclusions:

• State-of-the-art nonlinear models and ensemble methods advance genotype-to-phenotype prediction.
• The novel SNP selection approach offers a more efficient alternative to GWAS.
• Results establish a new state-of-the-art for predicting certain complex human phenotypes.