SumVg: Total heritability explained by all variants in genome-wide association studies based on summary statistics with standard error estimates

Hon‐Cheong So, Xue Xiao, Pak C. Sham

This paper introduces an updated framework for SumVg, an R package designed to estimate the total heritability of complex traits using only summary statistics from Genome-Wide Association Studies (GWAS). While the authors previously established the core method, this work fills a critical gap by providing robust ways to calculate Standard Error (SE), allowing researchers to finally put "error bars" on their heritability estimates.

TL;DR

SumVg offers a flexible alternative to standard heritability tools like LD Score Regression (LDSC) by ditching rigid assumptions about how genetic variants are distributed. This paper introduces new resampling-based methods—specifically parametric bootstraps—that provide accurate standard error estimates, even for traits where traditional tools fail.

What Problem Does This Paper Address?

In genetics, "heritability" measures how much of the variation in a trait (like height or disease risk) is due to DNA. Most researchers use a tool called LDSC to estimate this from GWAS summary data. However, LDSC relies on the "infinitesimal model," which assumes that every single genetic variant contributes a tiny, equal amount to the trait.

If a trait has a different "genetic architecture"—for example, if only a few variants have large effects—LDSC can produce inaccurate or even impossible (negative) results. The authors' previous tool, SumVg, used an Empirical Bayes approach to avoid these assumptions but lacked a way to measure the precision of its estimates. Without a Standard Error (SE), it was impossible to know if a heritability estimate was statistically significant or just noise.

Methods

The authors developed and tested five different resampling-based approaches to estimate SE:

  1. Standard Jackknife (Delete-1): Systematically leaves out one observation at a time to see how much the estimate changes.
  2. Delete-d Jackknife: Leaves out larger "chunks" of data ($d = n/5$) to improve stability for non-smooth parameters.
  3. Parametric Bootstrap: Simulates new "z-statistics" (measures of effect size) based on the corrected distribution of the original data.
  4. Modified Parametric Bootstraps (fdrboot1 & fdrboot2): These incorporate the "local false discovery rate," accounting for the probability that a specific variant actually has zero effect on the trait.

The core SumVg algorithm uses a formula established by Brown and Efron to "de-noise" observed GWAS results and recover the "true" underlying genetic signals.

Datasets

The team validated their methods using two primary sources:

  • Simulated Data: 100,000 independent SNPs across various sample sizes (5,000 to 200,000) and three levels of known heritability ($V_g = 0.101, 0.191, 0.295$).
  • Real-World Immune Traits: Summary statistics for 12 immune-related traits (cytokines and growth factors) from a published GWAS of ~8,000 individuals. These specific traits were chosen because LDSC previously failed to provide reasonable heritability estimates for them.

Key Insights and Ideas

  • The "Smoothness" Problem: The standard delete-one jackknife performed poorly, consistently overestimating the error. The authors attribute this to the fact that the "sum of heritability" is not a "smooth" parameter, making it hypersensitive to single-point changes.
  • Bootstrap Wins: Parametric bootstrap methods—especially those accounting for the local false discovery rate—yielded the lowest Root-Mean-Squared-Error (RMSE). They were the most reliable at mirroring the "true" SE found in simulations.
  • Rescuing "Un-heritable" Traits: For the 12 immune traits where LDSC showed near-zero or negative heritability, SumVg revealed stable estimates of 10–20%. This suggests that many "missing" heritability links might simply be artifacts of the rigid models used by older tools.
  • LD Independence: Unlike LDSC, which requires a perfectly matched reference panel to model Linkage Disequilibrium (LD), SumVg relies on simple LD pruning. This makes it more robust when a perfectly matched genetic reference for a specific population isn't available.

Key Results

  • Accuracy: In simulations, the parametric bootstrap methods achieved the lowest absolute bias and the best overall performance across all sample sizes.
  • Immune Architecture: The study found that IL-4 and IL-17 (key signaling proteins in the immune system) showed the highest heritability among the traits studied.
  • Software Availability: The methods are implemented in the SumVg R package, which is open-source and available on GitHub.

Limitations and Future Work

  • Conservative Estimates: Because SumVg requires pruning SNPs to ensure they are independent, it may accidentally discard some causal variants, leading to heritability estimates that are likely conservative (lower-bound).
  • Computational Intensity: Resampling methods (bootstrapping) are computationally heavier than simple regressions. The authors suggest that future iterations could benefit from parallel processing to speed up calculations.
  • Confidence Intervals: While the paper focuses on SE, the authors note that fully evaluating the best way to build Confidence Intervals (CIs)—perhaps by moving beyond simple normality assumptions—is a task for future research.

Summary updated

Abstract

Genome-wide association studies (GWAS) are commonly employed to study the genetic basis of complex traits and diseases, and a key question is how much heritability could be explained by all variants in GWAS. One widely used approach that relies on summary statistics only is LD score regression (LDSC), however the approach requires certain assumptions on the SNP effects (all SNPs contribute to heritability and each SNP contributes equal variance). More flexible modeling methods may be useful. We previously developed an approach recovering the true z-statistics from a set of observed z-statistics with an empirical Bayes approach, using only summary statistics. However, methods for standard error (SE) estimation are not available yet, limiting the interpretation of results and applicability of the approach. In this study we developed several resampling-based approaches to estimate the SE of SNP-based heritability, including two jackknife and three parametric bootstrap methods. Simulations showed that delete-d-jackknife and parametric bootstrap approaches provide good estimates of the SE. Particularly, the parametric bootstrap approaches yield the lowest root-mean-squared-error (RMSE) of the true SE. In addition, we applied our method to estimate SNP-based heritability of 12 immune-related traits (levels of cytokines and growth factors) to shed light on their genetic architecture. We also implemented the methods to compute the sum of heritability explained and the corresponding SE in an R package SumVg, available at https://github.com/lab-hcso/Estimating-SE-of-total-heritability/ . In conclusion, SumVg may provide a useful alternative tool for SNP heritability and SE estimates, which does not rely on distributional assumptions of SNP effects.
BibTeX
@article{so2024sumvg,
  author = {Hon‐Cheong So and Xue Xiao and Pak C. Sham},
  title = {{SumVg: Total heritability explained by all variants in genome-wide association studies based on summary statistics with standard error estimates}},
  journal = {International Journal of Molecular Sciences},
  year = {2024},
  doi = {10.3390/ijms25021347},
}
Figure