📊 statistics

Empirical Bayes Shrinkage of Functional Effects, with Application to Analysis of Dynamic eQTLs

This paper introduces Functional Adaptive Shrinkage (FASH), an empirical Bayes framework that leverages Gaussian processes to jointly estimate and test effect functions across units, offering improved power and principled inference for dynamic data such as time-varying eQTL studies.

Original authors: Ziang Zhang, Peter Carbonetto, Matthew Stephens

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ziang Zhang, Peter Carbonetto, Matthew Stephens

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the Signal in the Noise

Imagine you are trying to listen to a specific conversation in a crowded, noisy room. You have thousands of people (genes) talking to you, but they are all whispering, shouting, or changing their voices over time. Some of them are just making random noise, while others are telling you a real story about how their behavior changes as the day goes on.

This paper introduces a new statistical tool called FASH (Functional Adaptive Shrinkage). Think of FASH as a super-smart noise-canceling headphone that doesn't just block out noise, but actually learns what "normal" conversation sounds like so it can highlight the interesting stories.

The Problem: Too Many Noisy Measurements

In the world of genetics, scientists often study how genes react to changes, like how a gene behaves as a stem cell turns into a heart cell over 16 days. They measure this for millions of gene pairs.

The Issue: Each measurement is a bit "noisy" or fuzzy. If you look at just one gene on one day, you might think it's changing wildly, but that could just be random error.
The Old Way: Previously, scientists tried to fit these changes into simple shapes, like a straight line (going up steadily) or a curve (going up then down). If the data didn't fit that specific shape, they missed it. It was like trying to fit a jagged rock into a round hole; if it didn't fit, you threw it away.

The Solution: FASH (The Smart Filter)

The authors created FASH to solve this. Here is how it works, using a few metaphors:

1. The "Group Hug" (Borrowing Information)
Imagine you are trying to guess the height of 1,000 different people, but your ruler is broken and gives you shaky measurements. If you look at them one by one, you might get it wrong. But if you look at the whole group, you realize most people are between 5 and 6 feet tall.
FASH does this. It looks at all the genes at once. It learns what the "average" behavior looks like across the whole group. Then, for any single gene that looks weirdly noisy, FASH gently pulls (or "shrinks") its estimate toward what the group suggests is normal. This makes the estimates much clearer.

2. The "Flexible Rubber Band" (Adaptive Smoothing)
The authors use a mathematical concept called a "Gaussian Process" (specifically the L-GP family). Imagine a rubber band.

Tight Rubber Band: If the data looks very messy and random, the rubber band is tight. It forces the estimate to stay close to a simple, boring line (like "nothing is changing").
Loose Rubber Band: If the data shows a clear, strong pattern, the rubber band stretches out to follow the wiggles and curves of the data.
FASH automatically decides how tight or loose the rubber band should be for each gene, based on how much evidence there is.

3. The "Safety Net" (Conservative Adjustment)
One of the biggest worries in statistics is getting too excited and thinking you found a pattern when it's just noise. The authors added a special "safety net" to their method.
They realized that sometimes the math might be slightly off, leading to false alarms. So, they built in a rule that says: "If we aren't 100% sure, assume it's just noise." This makes the results "conservative." It's better to miss a few real discoveries than to claim you found something that isn't there.

What They Found (The Heart Cell Experiment)

To test this, the authors re-analyzed data from a study where stem cells were turning into heart cells over 16 days.

More Discoveries: FASH found many more genes that changed over time compared to the old methods. The old methods were like looking for a square peg in a round hole; FASH realized the pegs could be triangles, stars, or jagged rocks.
New Patterns: They found genes that didn't just go up or down in a straight line. Some genes suddenly spiked in the middle of the 16 days, or changed direction. FASH caught these complex shapes because it didn't force them into a straight line.
The "Switch" Genes: They found a special group of genes that acted like a light switch—turning "on" (positive effect) and then "off" (negative effect) or vice versa. When they looked closely at these specific genes, they found they were heavily linked to how cells react to low oxygen (hypoxia) and a specific protein called K-Ras. This suggests these genes play a key role in how cells grow and change.

Why This Matters

It's Flexible: It doesn't force data into a box. It lets the data tell you the shape of the story.
It's Honest: It uses the "safety net" to avoid false alarms.
It's Efficient: Even though it looks at millions of data points, it runs fast enough to be practical.

The authors have made this tool available as free software (an R package called fashr) so other scientists can use it to analyze their own "noisy" data, not just in genetics, but in any field where things change over time or conditions.

Technical Summary: Empirical Bayes Shrinkage of Functional Effects (FASH)

Problem Statement
Modern genomics, particularly dynamic expression quantitative trait locus (eQTL) studies, generates massive datasets where the effect of a genetic variant on gene expression is measured across a continuous condition variable (e.g., time, oxygen levels). These studies produce noisy effect estimates for thousands to millions of genomic units (gene-variant pairs) at multiple time points. The primary challenges are:

Estimation: Accurately recovering the underlying "effect function" (how the effect changes over the condition) from noisy data.
Hypothesis Testing: Identifying which units deviate from a null or baseline behavior (e.g., no change over time, or linear change) without making restrictive parametric assumptions about the shape of the effect.
Calibration: Ensuring that Empirical Bayes (EB) inferences remain calibrated and conservative, even when the prior distribution under the alternative hypothesis is misspecified—a common issue in high-dimensional functional data.

Methodology: Functional Adaptive Shrinkage (FASH)
The authors introduce FASH, an Empirical Bayes framework designed for the joint analysis of observation units where each unit estimates an effect function.

Statistical Framework:
- Data Model: For each unit $j$ , observed effect estimates $\hat{\beta}_j(t)$ at discrete time points are modeled as normally distributed around the true underlying function $\beta_j(t)$ with known standard errors.
- Prior Specification (L-GP Family): The method utilizes a family of Gaussian Processes (GPs) defined by linear differential operators, termed the L-GP family. A baseline model $S_0$ is defined by a linear differential operator $L$ (e.g., $L=D_1$ for constant functions, $L=D_2$ for linear functions). The prior assumes $\beta \sim \text{L-GP}(\sigma)$ , where $\sigma$ quantifies the deviation from the baseline space $S_0$ .
- Adaptive Mixture Priors: To allow data-driven shrinkage, FASH employs a mixture of L-GP priors: $g_\beta = \sum \pi_k \text{L-GP}(\sigma_k)$ . The mixture weights $\pi_k$ are learned via maximum likelihood estimation across all units, effectively pooling information to adapt the amount of shrinkage for each unit based on how well its data fits the baseline model.
- Posterior Inference: Due to conjugacy, the posterior distribution for each unit is also a mixture of GPs. This allows for the computation of posterior means (smoothed effect functions) and credible intervals.
Hypothesis Testing:
- FASH computes local false discovery rates (lfdr) and local false sign rates (lfsr).
- Testing $H_0: \beta_j \in S_0$ corresponds to estimating the posterior probability that the unit belongs to the null component of the mixture ( $k=0$ ).
- The framework supports testing arbitrary functionals $F(\beta_j)$ (e.g., "does the effect switch sign?" or "is the effect strongest in the middle period?"), not just simple deviations from a baseline.
Prior Adjustment for Conservatism:
- Recognizing that misspecification of the alternative prior can lead to anti-conservative inferences (inflated false discovery rates), the authors propose a Bayes Factor (BF)-based adjustment.
- Leveraging the property that the expected Bayes factor under the null is 1, the algorithm adjusts the estimated null proportion $\hat{\pi}_0$ to ensure it is conservative (i.e., $\hat{\pi}_0 \geq \pi_0$ ). This guarantees that the resulting inferences are conservative even if the alternative prior family is misspecified, provided the null model is correctly specified.

Key Results: Application to Dynamic eQTLs
The authors applied FASH to reanalyze a dynamic eQTL dataset from Strober et al. (2019), involving 19 cell lines undergoing cardiomyocyte differentiation over 16 days (approx. 1 million gene-variant pairs).

Discovery Power: Using a constant baseline model ( $L=D_1$ ), FASH identified 9,205 dynamic eQTLs (1,177 genes) at an FDR of 0.05. This is a substantial increase compared to the original study's parametric interaction models (linear and quadratic), which identified 5,404 and 6,824 pairs respectively.
Flexibility: FASH successfully captured diverse temporal patterns that parametric models missed, including abrupt changes in effect strength and non-monotonic trajectories.
Characterization: Using a linear baseline model ( $L=D_2$ ), FASH identified 44 nonlinear dynamic eQTLs (9 genes).
Novel Biological Insights: The authors categorized dynamic eQTLs into temporal patterns (early, middle, late, and "switch" where the effect direction changes). Gene set enrichment analysis revealed that "switch" dynamic eQTLs were significantly enriched for genes upregulated by hypoxia and K-Ras activation, pathways known to influence stem cell proliferation and differentiation.
Software: The method is implemented in the R package fashr, which operates on summary statistics, making it applicable even when individual-level data is unavailable.

Significance and Claims
The paper claims that FASH offers a flexible, principled statistical framework for joint analysis of functional data that extends beyond genomics. Its primary contributions are:

Adaptive Smoothing: It enables adaptive smoothing and information borrowing across units without requiring pre-specified parametric forms for the effect functions.
Robust Hypothesis Testing: It provides a unified framework for testing complex hypotheses about functional shapes (e.g., nonlinearity, sign-switching) with calibrated error rates.
Conservative Inference: The proposed BF-based adjustment addresses a known limitation of Empirical Bayes methods regarding prior misspecification, ensuring conservative inference in high-dimensional settings.
Practical Utility: By operating on summary statistics and providing an accessible R package, FASH facilitates the reanalysis of existing dynamic eQTL studies and the analysis of new datasets where individual-level data may be restricted.

The authors note that while FASH is demonstrated on dynamic eQTLs, the methodology is general and applicable to other functional data settings, including spatial data or single-cell RNA sequencing trajectories.