Inferring disease architecture and predictive ability with LDpred2-auto.

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Polygenic scores (PGSs) are crucial for predicting complex traits.
• LDpred2 is a popular Bayesian method for PGS construction.
• LDpred2-auto previously inferred SNP heritability (h²) and polygenicity (p) without validation datasets.

Purpose of the Study:

• To validate LDpred2-auto for inferring multiple genetic parameters, including a new parameter for negative selection (α).
• To assess the calibration of per-variant causal probabilities for fine-mapping.
• To introduce a formula for inferring out-of-sample PGS predictive performance (r²).

Main Methods:

• Introduced a new version of LDpred2-auto with an optional third parameter, α.
• Validated the inference of two and three genetic parameters.
• Developed a formula to calculate predictive performance (r²) directly from the Gibbs sampler.

Main Results:

• LDpred2-auto successfully infers two and three genetic parameters, including negative selection.
• Per-variant causal probabilities are well-calibrated, suitable for fine-mapping.
• The extended HapMap3 variant set (37% larger) improved heritability capture (12%) and PGS predictive performance (6% on average in UK Biobank).

Conclusions:

• The enhanced LDpred2-auto provides robust inference of genetic architecture.
• Calibrated causal probabilities enable effective genetic fine-mapping.
• Optimized variant sets and improved methods enhance polygenic score accuracy and utility.