Biobank-scale methods and projections for sparse polygenic prediction from machine learning.

Area of Science:

• Genetics
• Machine Learning
• Bioinformatics

Background:

• Polygenic scores are crucial for predicting complex traits.
• Sparse machine learning algorithms are widely used for model training.
• Understanding model performance across diverse genetic ancestries is essential.

Purpose of the Study:

• To characterize the performance of linear models trained with sparse machine learning algorithms.
• To evaluate the impact of training set size, genetic ancestry, and training method on polygenic score performance.
• To develop a novel method for projecting predictive performance based on data size.

Main Methods:

• Building polygenic scores using sparse machine learning algorithms (e.g., LASSO).
• Examining performance metrics as a function of training data size and genetic ancestry.
• Investigating cross-ancestry predictor performance.
• Developing a projection method for AUC and correlation based on data size.

Main Results:

• Predictor performance is most dependent on training data size, with marginal gains from algorithmic improvements.
• LASSO demonstrates performance comparable to other leading methods.
• Predictors trained on one ancestry group show varying performance when applied to others.
• A novel projection method accurately predicts performance limits and aligns with theoretical predictions.

Conclusions:

• Training data quantity is the primary driver of polygenic score performance.
• LASSO is a robust and effective method for building polygenic scores.
• Future polygenic scores, like those from the Taiwan Precision Medicine Initiative, are expected to achieve high predictive accuracy across diverse populations.