Improved methods for empirical Bayes multivariate multiple testing and effect size estimation

Yunqi Yang, Peter Carbonetto, David Gerard, Matthew StephensarXiv 2024 · DOI · arXiv

TL;DR

The authors introduce "Ultimate Deconvolution," a new empirical Bayes framework that significantly speeds up and improves the accuracy of multivariate analyses in genomics. By combining a novel "truncated eigenvalue decomposition" (TED) algorithm with adaptive regularization, the method provides better estimates of how genetic effects are shared across dozens of tissues or conditions compared to the widely used MASH pipeline.

What Problem Does This Paper Address?

In modern genomics, researchers often want to know how a single genetic variant affects multiple outcomes—for example, gene expression across 49 different human tissues (eQTL analysis). The simplest approach is to analyze each tissue one by one, but this is inefficient. It ignores the fact that tissues are biologically related; a variant affecting the heart is likely to have a similar effect in the artery.

To "borrow information" across conditions, researchers use the Empirical Bayes Multivariate Normal Means (EBMNM) model. However, as the number of conditions ($R$) grows, estimating the underlying patterns of sharing (covariance matrices) becomes a nightmare. Existing tools like "Extreme Deconvolution" (ED) are often painfully slow, sensitive to how they are started, and prone to overfitting when the sample size is small relative to the number of conditions.

Methods

The paper introduces two major technical upgrades to the EBMNM framework:

  1. Truncated Eigenvalue Decomposition (TED): Instead of relying solely on iterative EM algorithms (like ED), the authors developed TED. In the "homoskedastic" case (where measurement noise is consistent), TED can solve the covariance estimation problem exactly and much faster. It effectively "rescues" the optimization process from the local optima that plague older methods.
  2. Adaptive Regularization: To prevent the model from hallucinating patterns in high-dimensional data, the authors introduced penalty functions—specifically Inverse Wishart (IW) and Nuclear Norm (NN) penalties. Crucially, they reformulated these penalties to be "scale-invariant," meaning the results won't change just because you changed the units of your measurements.
  3. The udr Package: These methods are implemented in a new R package called udr (Ultimate Deconvolution in R), designed to be a drop-in improvement for the popular mashr pipeline.

Datasets

The authors validated their approach using:

  • Extensive Simulations: They tested "Hybrid" scenarios (a mix of shared and independent effects) and "Rank-1" scenarios (where effects are perfectly correlated across conditions), varying the ratio of samples ($n$) to conditions ($R$).
  • GTEx Project Data: They analyzed expression quantitative trait loci (eQTLs) across 49 human tissues using a dataset of over 15,000 genes. This is a standard benchmark for multi-tissue genomic analysis.

Key Insights and Ideas

  • The "Subspace-Preserving" Trap: The authors discovered that the older ED algorithm is "subspace-preserving." If you start it with a simple, low-rank guess, it can never "learn" its way out of that simple structure to find more complex, realistic patterns. TED breaks this limitation.
  • Rank-1 Constraints are Risky: While low-rank (Rank-1) matrices are easier to interpret, the paper demonstrates they lead to poorly calibrated significance tests. Specifically, they force the model to be equally confident about the sign of an effect across all tissues, which is rarely true in biology.
  • Regularization as Robustness: Adding a penalty doesn't just improve accuracy; it makes the model robust to the "number of components" ($K$) problem. In the new pipeline, you can specify a large number of potential sharing patterns, and the regularization will automatically "shrink" the irrelevant ones to zero without overfitting.

Key Results

  • Speed: In the GTEx analysis, the new TED-based pipeline reached convergence in significantly fewer iterations than the previous ED-based approach (e.g., 412 vs. 1,083 iterations in one test case).
  • Accuracy: Penalized TED consistently achieved higher out-of-sample log-likelihoods than unpenalized methods, especially in "small $n$, large $R$" settings.
  • Biological Discovery: When applied to real GTEx data, the new pipeline identified a much wider variety of tissue-specific sharing patterns. For instance, it successfully captured distinct sharing among brain-specific tissues that the previous pipeline had partially obscured.

Limitations and Future Work

  • Homoskedasticity: The TED algorithm is mathematically optimized for cases where measurement error is the same across observations. While the authors use data transformation to handle more complex cases, a direct "heteroskedastic" version of TED remains an open challenge.
  • Anti-Conservatism: Like many Empirical Bayes methods, these tools can sometimes be "anti-conservative," meaning they might slightly overstate confidence in an effect's sign. The authors suggest that users should use more stringent significance thresholds (e.g., an $lfsr$ of 0.01 instead of 0.05) to compensate.
  • Future Directions: The authors point toward improving the calibration of multivariate tests and exploring "sparse" estimates of sharing patterns to further improve interpretability.

Summary updated

Abstract

Estimating the sharing of genetic effects across different conditions is important to many statistical analyses of genomic data. The patterns of sharing arising from these data are often highly heterogeneous. To flexibly model these heterogeneous sharing patterns, Urbut et al. (2019) proposed the multivariate adaptive shrinkage (MASH) method to jointly analyze genetic effects across multiple conditions. However, multivariate analyses using MASH (as well as other multivariate analyses) require good estimates of the sharing patterns, and estimating these patterns efficiently and accurately remains challenging. Here we describe new empirical Bayes methods that provide improvements in speed and accuracy over existing methods. The two key ideas are: (1) adaptive regularization to improve accuracy in settings with many conditions; (2) improving the speed of the model fitting algorithms by exploiting analytical results on covariance estimation. In simulations, we show that the new methods provide better model fits, better out-of-sample performance, and improved power and accuracy in detecting the true underlying signals. In an analysis of eQTLs in 49 human tissues, our new analysis pipeline achieves better model fits and better out-of-sample performance than the existing MASH analysis pipeline. We have implemented the new methods, which we call ``Ultimate Deconvolution'', in an R package, udr, available on GitHub.
BibTeX
@misc{yang2024improved,
  author = {Yunqi Yang and Peter Carbonetto and David Gerard and Matthew Stephens},
  title = {{Improved methods for empirical Bayes multivariate multiple testing and effect size estimation}},
  year = {2024},
  eprint = {2406.08784},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.08784},
}
Figure