TL;DR
This paper introduces LDpred-funct, a novel method that significantly improves the accuracy of polygenic risk prediction by incorporating functional genomic information. By leveraging trait-specific "priors" about which genetic variants are more likely to be causal, LDpred-funct achieves better prediction of complex traits compared to existing state-of-the-art methods, paving the way for more precise genetic risk assessments.
What Problem Does This Paper Address?
Polygenic risk prediction (PRP) aims to estimate an individual's genetic predisposition to complex traits (like height, disease susceptibility, etc.) by aggregating the effects of many genetic variants. While PRP has shown promise for clinical applications, its accuracy is limited. A key observation is that genetic variants located in "functional" regions of the genome (e.g., those involved in gene regulation or protein coding) tend to contribute more to heritability than variants in non-functional regions. However, existing PRP methods either don't fully leverage this information or do so in ways that are suboptimal (e.g., by dichotomizing variants into functional/non-functional). This paper addresses the challenge of effectively integrating this functional genomic knowledge into PRP models to boost prediction accuracy, especially for complex traits where the genetic architecture is often sparse (meaning only a subset of variants are truly causal).
Methods
The core innovation is the LDpred-funct method. It builds upon the LDpred framework, which accounts for linkage disequilibrium (LD) – the non-random association of alleles at different loci. LDpred-funct introduces two key enhancements:
- Trait-Specific Functional Priors: It uses the baseline-LD model to estimate trait-specific enrichments of functional annotations (coding, conserved, regulatory, etc.). These enrichments act as "priors," informing the model about the relative likelihood of variants in different functional categories being causal.
- Regularization via Cross-Validation: After analytically estimating posterior mean causal effect sizes (incorporating the functional priors and LD), LDpred-funct uses cross-validation to regularize these estimates. This is particularly important for sparse architectures, where only a few variants are truly causal. By grouping SNPs into bins based on their estimated effect sizes and optimizing weights for these bins, the method effectively shrinks less likely causal effects towards zero, improving prediction accuracy.
The paper also compares LDpred-funct to several other methods, including:
- P+T (Pruning + Thresholding): A basic method that selects SNPs based on p-values and LD.
- LDpred: A method that models LD but not functional priors.
- SBayesR: A Bayesian method that accounts for sparsity but not functional priors.
- P+T-funct-LASSO: A method that uses functional priors to define "high-prior" and "low-prior" SNP sets and applies LASSO shrinkage.
- AnnoPred: A Bayesian method that incorporates functional annotations.
- LDpred-funct-inf: A version of LDpred-funct that incorporates functional priors but does not perform the cross-validation-based regularization step.
Datasets
The study utilizes several large-scale datasets:
- UK Biobank: Used for both simulations (using an interim release) and analyses of 21 complex traits. The full UK Biobank dataset comprises ~459K European-ancestry samples and ~20 million imputed SNPs. Training samples were restricted to ~373K British-ancestry individuals, and validation samples consisted of ~22K non-British European ancestry individuals.
- 23andMe: Data from 23andMe customers was used for a meta-analysis of height prediction, combining ~698K training samples with UK Biobank data to reach a total of ~1.1 million training samples.
- 1000 Genomes Project: Used as a reference panel for estimating LD.
- UK10K: Used in secondary analyses for LD estimation.
The paper does not introduce new benchmark datasets but rather applies its method to existing, well-characterized datasets.
Key Insights and Ideas
- Functional regions matter, and trait-specifically: The core idea that functional regions are enriched for heritability is well-established. However, LDpred-funct's innovation lies in leveraging trait-specific functional enrichments, acknowledging that the functional architecture can vary across different traits.
- Sparsity is key, and regularization helps: For many complex traits, the genetic architecture is sparse, meaning only a small fraction of variants are truly causal. LDpred-funct's regularization step, informed by cross-validation, is crucial for handling this sparsity and improving prediction accuracy, especially when compared to methods that assume an infinitesimal model.
- Functional priors are more than just a filter: Unlike some earlier methods that might simply filter SNPs based on functional status, LDpred-funct integrates these priors into a probabilistic framework, allowing for a more nuanced weighting of SNP effects.
- Computational efficiency is a win: While AnnoPred achieved comparable accuracy in some settings, LDpred-funct was significantly faster, making it more practical for large-scale genomic analyses.
Key Results
- Improved Prediction Accuracy: LDpred-funct achieved a +4.6% relative improvement in average prediction R² across 21 UK Biobank traits compared to SBayesR (the best non-functional method). The average prediction R² was 0.144, with a peak of 0.413 for height.
- Height Prediction Boost: In a meta-analysis of height using UK Biobank and 23andMe data (N=1.1M), LDpred-funct achieved an R² of 0.431.
- Simulation Performance: Simulations showed that LDpred-funct consistently outperformed methods that did not incorporate functional information, with improvements ranging from +4.7% to +7.3% depending on the architecture and sample size.
- Comparison to Other Functional Methods: LDpred-funct generally outperformed P+T-funct-LASSO and AnnoPred, particularly in simulations and for average prediction accuracy across traits.
- Calibration: LDpred-funct demonstrated good calibration, meaning its predictions were generally well-aligned with observed phenotypes.
- Computational Speed: LDpred-funct was significantly faster than AnnoPred, taking 71 minutes compared to 5249 minutes for estimating effect sizes.
Limitations and Future Work
- Raw Genotype Data: LDpred-funct primarily uses summary statistics, which are publicly available. However, methods that use raw genotype and phenotype data can potentially achieve higher accuracy. Future work could explore integrating functional enrichment into such methods.
- Heuristic Regularization: The regularization step in LDpred-funct relies on heuristic cross-validation rather than a fully Bayesian approach for sparse functional models. This choice was made due to the complexity of defining appropriate sparse functional models and potential convergence issues with MCMC. The effectiveness of this step might depend on the number of validation samples.
- Single-Trait Focus: The current study focuses on single-trait analyses. Future work could explore leveraging genetic correlations among traits for improved prediction.
- Population Diversity: While the study evaluated performance in diverse populations, further investigation into applying LDpred-funct to diverse populations, potentially with non-European training data, is warranted.
- Functional Annotation Models: The study used the baseline-LD model. Future research could explore how different functional annotation models (e.g., incorporating tissue-specific enrichments, MAF-dependent architectures) might further improve prediction accuracy.
- Related Individuals: The current framework does not explicitly consider polygenic prediction in related individuals.
Summary updated