Testing for gene-environment (GxE) interaction using p-value aggregation identifies many GxE loci

This paper proposes a robust gene-environment (GxE) interaction testing method using Cauchy p-value aggregation across additive, dominant, and recessive genetic models, which simulation and UK Biobank analyses demonstrate significantly outperforms standard approaches in detecting GxE loci, particularly when the underlying genetic model is non-additive.

The Gist

Imagine you are trying to find the "secret recipe" that makes some people get sick when they smoke, while others stay healthy. Scientists know that your genes (your DNA) and your environment (like smoking or sleep) work together to determine your health. This teamwork is called a Gene-Environment (GxE) interaction.

However, finding these secret recipes is incredibly hard. It's like trying to find a needle in a haystack, but the needle might be shaped like a square, a circle, or a triangle, and you don't know which one it is.

The Problem: The "One-Size-Fits-All" Mistake

For years, scientists have been looking for these interactions using a "one-size-fits-all" approach. They usually assume that genes work in a simple, additive way.

The Analogy: Imagine you are trying to open a locked door.

• The Old Way: You assume the lock is a standard keyhole. You try to turn a standard key (the Additive Model). If the lock is actually a standard keyhole, you open it easily.
• The Problem: But what if the lock is actually a digital keypad (a Dominant model) or a biometric scanner (a Recessive model)? If you keep trying to use the standard key, you will never open the door, even though the door can be opened. You just have the wrong tool for the job.

In genetics, if the true way a gene works is "Recessive" (you need two copies of a gene to see the effect), but scientists only test for "Additive" (one copy is enough), they miss the signal entirely. They lose the "needle" in the haystack because they are looking for the wrong shape.

The Solution: The "Swiss Army Knife" (GETAP)

The authors of this paper, led by Saurabh Mishra and Arunabha Majumdar, proposed a new method called GETAP (GxE Testing using Aggregated P-value).

The Analogy: Instead of carrying just one key, GETAP is like a Swiss Army Knife.

1. It tries the Standard Key (Additive model).
2. It tries the Keypad (Dominant model).
3. It tries the Biometric Scanner (Recessive model).

Instead of picking just one and hoping for the best, GETAP tries all three at once. It takes the results from all three attempts and combines them into a single, super-strong signal using a mathematical trick called Cauchy p-value aggregation.

Think of it like a choir. If one singer is slightly off-key, the whole song might sound bad. But if you have three singers, and even if one is quiet, the other two might be loud enough to carry the tune. GETAP listens to all three "singers" (genetic models) and combines their voices. If any of them hears a signal, the combined voice is loud enough for the scientists to hear it.

How They Tested It

The researchers didn't just guess; they put their Swiss Army Knife to the test in two ways:

1. The Simulation Lab: They created millions of fake people with fake genes and fake environments. They knew exactly which "lock" (genetic model) was real.

   • Result: When the real lock was a "Recessive" one, the old method (Additive key) failed miserably. GETAP, however, found the door every time. It was also faster and more powerful than other complex methods that tried to be "model-free."

2. The Real World (UK Biobank): They applied their method to real data from 500,000 people in the UK. They looked at things like:

   • Smoking vs. Blood Sugar: How does smoking affect blood sugar in people with different genes?
   • Sleep vs. Diabetes: How does sleep duration interact with genes to cause Type 2 Diabetes?

The Results:

• For Blood Sugar (HbA1c) and Smoking: The old method found 24 "hits" (locations in the DNA). GETAP found 82 hits. That's more than triple the discoveries!
• For Diabetes and Sleep: The old method found a few hundred hits. GETAP found 563 hits.

Why This Matters

This paper is a game-changer because it stops scientists from guessing which "lock" a gene uses. By using a method that covers all the bases (Additive, Dominant, and Recessive) simultaneously, they can find many more genetic interactions that were previously invisible.

In simple terms:
Before, scientists were looking for a specific type of key to open a door, and they kept missing the door because they didn't know what kind of lock it was. Now, with GETAP, they have a master key that works on almost any lock. This means we can finally understand how our lifestyle (like smoking or sleeping) interacts with our DNA to make us sick or healthy, leading to better treatments and prevention strategies in the future.

The Takeaway

The authors showed that by combining different ways of looking at the data, we don't just get a little bit more information; we get a massive amount of new discoveries. It's a smarter, more robust way to solve the puzzle of human health.

Technical Summary

1. Problem Statement

Genome-wide gene-environment (GxE) interaction studies have historically struggled with low statistical power and limited success in detecting reliable signals. A primary contributor to this "power deficit" is genetic model misspecification.

• The Issue: Standard Genome-Wide Association Studies (GWAS) and GxE scans typically assume a single genetic mode of inheritance (usually the additive model) for all Single Nucleotide Polymorphisms (SNPs).
• The Consequence: If the true underlying genetic model for a specific SNP is non-additive (e.g., dominant or recessive), assuming an additive model leads to significant power loss. Conversely, assuming a specific non-additive model when the truth is additive also results in power loss.
• Existing Alternatives & Limitations:
  • 2df Genotypic Test: A model-free approach treating genotypes as categories. While robust, it consumes an extra degree of freedom, often reducing power compared to correctly specified 1-degree-of-freedom (1df) tests.
  • MAX3 Test: Takes the maximum of test statistics from three models. It requires computationally intensive resampling (permutation) to derive valid p-values, making it impractical for genome-wide scans involving millions of variants.
  • Multiple Testing Corrections: Running separate tests for additive, dominant, and recessive models and applying Bonferroni correction is overly conservative, further reducing power.

2. Methodology: The GETAP Approach

The authors propose GETAP (GxE Testing using Aggregated P-value), a robust framework that aggregates evidence from multiple genetic models without requiring computationally expensive resampling.

• Core Strategy: Instead of choosing one genetic model, GETAP performs three separate GxE interaction tests for each SNP using Additive (A), Dominant (D), and Recessive (R) genotype codings.
• P-value Aggregation: The three resulting p-values ($p_A, p_D, p_R$) are combined into a single global p-value using the Aggregated Cauchy Association Test (ACAT).
  • Formula: $T_{ACAT} = \sum w_i \tan[(0.5 - p_i)\pi]$, where weights $w_i$ are uniform ($1/3$).
  • Advantage: ACAT provides a valid combined p-value even when the input tests are dependent (which they are, as they use the same data) and is computationally extremely fast.
• Alternative Method: The authors also evaluated Harmonic Mean P-value (HMP) aggregation but found ACAT to be slightly superior in power.
• Statistical Framework: The method utilizes Generalized Linear Models (GLM) for both continuous and binary phenotypes, adjusting for covariates (age, sex, principal components).

3. Key Contributions

1. Novel Framework: Introduction of GETAP, a scalable, model-agnostic GxE testing procedure that mitigates the risk of genetic model misspecification.
2. Computational Efficiency: Unlike MAX3, GETAP avoids permutation-based p-value calculation, making it feasible for large-scale biobank data (hundreds of thousands of individuals and millions of SNPs).
3. Systematic Evaluation: Comprehensive simulation studies and real-world applications in the UK Biobank comparing GETAP against standard 1df models (A, D, R) and the 2df genotypic test.
4. Discovery Scale: Application of the method to diverse phenotype-environment pairs, significantly expanding the catalog of known GxE loci.

4. Key Results

A. Simulation Studies

• Type I Error Control: GETAP (specifically ACAT) maintained adequate control of the Type I error rate (TIER) across various scenarios, including continuous and binary phenotypes, though slight inflation was observed for rare variants (MAF < 0.05) in binary outcomes, consistent with the constituent recessive tests.
• Power Analysis:
  • True Additive Model: GETAP performed nearly identically to the correctly specified additive test, with negligible power loss (1–5%).
  • True Dominant Model: GETAP performed competitively with the additive test (the second-best performer) and substantially better than the misspecified recessive model.
  • True Recessive Model: GETAP showed substantial power gains (up to 70%) over the standard additive and dominant tests, which suffer severe power loss in this scenario. It incurred only a marginal power loss (1–13%) compared to the correctly specified recessive test.
  • Comparison with 2df Test: GETAP was uniformly more powerful than the 2df test when the true model was additive or dominant. In recessive scenarios, GETAP was comparable to or slightly less powerful than the 2df test at very low MAFs but performed similarly at moderate MAFs.

B. Real-World Application (UK Biobank)

The authors analyzed 9 phenotype-environment combinations, including continuous traits (HbA1c, FEV1/FVC, BMI, CRP, Triglycerides) and binary diseases (Type 2 Diabetes, COPD).

• HbA1c & Smoking: GETAP identified 82 independent GxE loci (vs. 24 for the additive model). The top signal was at rs407423 (Chr 8).
• Type 2 Diabetes (T2D) & Sleep Duration: GETAP detected 563 independent GxE loci, a massive increase compared to the additive model (414 loci) and other methods. This represents a significant expansion of known GxE signals for T2D.
• COPD & Smoking: GETAP identified 219 independent loci, outperforming the recessive model (200 loci) and the additive model (13 loci).
• General Trend: Across all analyses, GETAP consistently identified the highest or a comparable number of significant loci compared to single-model and 2df tests. It successfully recovered signals missed by individual models (e.g., 20 unique HbA1c loci, 73 unique T2D loci).
• Genomic Inflation: Inflation factors were generally higher for methods detecting more signals (indicating polygenic architecture rather than confounding). Sample-size standardized inflation factors were close to 1, confirming robustness.

C. Functional Annotation

• Functional analysis (using FUMA/MAGMA) of T2D and HbA1c loci revealed that GxE variants are predominantly located in non-coding regions (intronic/intergenic), suggesting regulatory mechanisms.
• Pathway enrichment highlighted biological relevance, including xenobiotic metabolism (CYP enzymes) and lipid synthesis for T2D/sleep interactions, and metabolic transport for HbA1c/smoking interactions.

5. Significance

• Robustness: GETAP solves the "model uncertainty" problem in GxE studies. It ensures that researchers do not miss true interactions simply because they guessed the wrong genetic model.
• Scalability: By leveraging the speed of Cauchy p-value aggregation, GETAP makes robust GxE scanning feasible for modern biobank-scale datasets where computational cost is a bottleneck.
• Discovery: The study demonstrates that the "true" genetic architecture of GxE interactions is heterogeneous. Aggregating models reveals a much larger landscape of GxE loci (e.g., 563 for T2D) than previously thought, providing new targets for understanding disease etiology and precision medicine.
• Practical Utility: The method can be easily implemented as a post-processing step on top of standard regression pipelines (e.g., PLINK), lowering the barrier for adoption in the broader genetics community.

In conclusion, the paper establishes GETAP as a superior, robust, and computationally efficient standard for genome-wide GxE interaction screening, particularly in scenarios where the underlying genetic model is unknown or potentially non-additive.