This paper introduces RareEffect, a new computational framework designed to estimate the effect sizes of rare genetic variants. While common variants have long been the focus of genomic studies, rare variants (found in less than 1% of the population) are thought to hold the key to much of the "missing heritability" in complex diseases. RareEffect provides a scalable, two-level estimation approach—at both the gene and individual variant levels—to improve disease risk prediction and biological interpretation.
TL;DR
RareEffect is a scalable Bayesian method that estimates effect sizes for rare genetic variants by first calculating gene-level heritability and then applying a fast, bias-corrected shrinkage estimator. It outperforms standard regression models in accuracy and speed, significantly improving the performance of polygenic risk scores (PRS) by identifying high-risk individuals that common-variant models often miss.
What Problem Does This Paper Address?
In the world of genomics, "rare" is a relative term, but it usually refers to variants with a Minor Allele Frequency (MAF) below 1%. Because these variants appear so infrequently in a dataset, standard statistical methods struggle to determine their impact.
Existing tools like SKAT or SKAT-O are great at providing p-values (telling us if a gene is associated with a trait), but they don't provide effect sizes (telling us how much a specific mutation changes a trait). Conversely, "Burden tests" assume all rare variants in a gene act in the same direction and with the same magnitude—an assumption that is often biologically incorrect. The gap addressed here is the need for a method that is computationally efficient enough for biobank-scale data (hundreds of thousands of people) while providing stable, directional effect size estimates for individual rare mutations.
Methods
RareEffect operates in three distinct steps:
- Null Model Fitting: It first fits a standard mixed-effect model to account for covariates (like age or sex) and genetic relatedness, producing "clean" residuals for the next steps.
- Heritability Estimation: It models rare variant effects as random effects. To handle the massive scale of modern biobanks, it uses FaST-LMM, an algorithm that leverages the low-rank nature of genetic matrices to speed up calculations. It also uses a Method of Moments (MoM) approach to combine different functional categories (like Loss-of-Function vs. missense mutations) into a single gene-level heritability estimate.
- Effect Size Estimation: Finally, it uses an empirical Bayesian approach (specifically, Best Linear Unbiased Predictor or BLUP) to estimate the effect size of each variant.
- Fast Firth Correction: For binary traits (like "has disease" vs. "doesn't"), the authors implemented a high-speed version of the Firth bias correction. This prevents the "exploding" effect sizes often seen in imbalanced case-control data, reducing computation time from minutes to seconds.
Datasets
The authors validated RareEffect using:
- Simulated Data: Based on actual genotypes from the UK Biobank to ensure realistic linkage disequilibrium (LD) patterns.
- UK Biobank (UKB) Whole Exome Sequencing (WES): Data from ~393,247 participants.
- Phenotypes: 5 quantitative traits (HDL/LDL cholesterol, triglycerides, height, BMI) and 5 binary disease traits (breast cancer, prostate cancer, lymphoid leukemia, type 2 diabetes, and coronary atherosclerosis).
Key Insights and Ideas
- Directionality Matters: The authors highlight the gene SLC12A3. A standard burden test failed to find a significant signal because some rare variants in the gene increased the trait while others decreased it, canceling each other out. RareEffect successfully identified these opposing directions.
- Complementary Risk: Rare variants and common variants often point to different high-risk individuals. The study found that the sub-groups identified as high-risk by common-variant PRS and rare-variant PRS were "substantially distinct," meaning a combined model captures a much larger cohort of at-risk people.
- Sparsity is a Feature: By leveraging the fact that most people have a "zero" for any given rare variant, the authors optimized their Firth correction to be 100x faster than non-sparse approaches.
Key Results
- Accuracy: RareEffect consistently showed lower Root Mean Squared Error (RMSE) in estimating effect sizes compared to linear or ridge regression.
- Computational Efficiency: For the DOCK6 gene (4,114 variants), RareEffect was 90% faster than ridge regression (4.2 seconds vs 44.6 seconds).
- Firth Speedup: Their fast implementation of the Firth correction reduced analysis time for a gene with 250 variants from 708 seconds to just 2.9 seconds.
- PRS Improvement: In predicting lipid levels, combining RareEffect scores with common variant scores significantly boosted predictive power ($R^2$). For example, in the top/bottom 0.5% of the population, the combined model achieved an $R^2$ of ~0.62-0.66, compared to ~0.54-0.58 for common variants alone.
Limitations and Future Work
- Binary Trait Performance: While RareEffect excelled at quantitative traits, the predictive gains for binary disease traits were more marginal. The authors attribute this to the limited number of highly associated rare variants currently known for those specific diseases.
- Shrinkage Bias: Because the method uses BLUP (a shrinkage estimator), the estimated effect sizes are inherently biased toward zero to ensure stability. While this makes them more reliable for prediction, they may not represent the "true" unconstrained biological effect.
- Ultra-rare Variants: The current model collapses "ultra-rare" variants (those appearing in fewer than 10 people) into a single "super-variant" to maintain statistical power, which might obscure the specific impact of a single, highly potent mutation.
RareEffect is available as part of the open-source SAIGE R package (version 1.3.7 or higher).
Summary updated