Improved genetic prediction of complex traits from individual-level data or summary statistics

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Existing genetic prediction models often assume equal genetic variant contribution to phenotypes.
• This assumption is suboptimal for accurately modeling heritability distribution across the genome.

Purpose of the Study:

• To develop and evaluate novel genetic prediction tools that incorporate user-defined heritability models.
• To compare the performance of these new tools against established methods using large-scale genetic datasets.

Main Methods:

• Developed new prediction tools: LDAK-Bolt-Predict (individual-level data) and LDAK-BayesR-SS (summary statistics).
• Compared tools using 14 UK Biobank phenotypes for individual-level data and 225 phenotypes for summary statistics.
• Evaluated performance based on the proportion of phenotypic variance explained.

Main Results:

• LDAK-Bolt-Predict outperformed Lasso, BLUP, Bolt-LMM, and BayesR for all 14 tested phenotypes.
• LDAK-BayesR-SS outperformed lassosum, sBLUP, LDpred, and SBayesR for 223 out of 225 phenotypes.
• Improving the heritability model increased the proportion of phenotypic variance explained by an average of 14%.

Conclusions:

• User-specified heritability models significantly improve genetic prediction accuracy.
• The developed LDAK tools offer superior performance compared to existing methods.
• Enhanced heritability modeling provides a substantial gain in predictive power, equivalent to a significant increase in sample size.