Representation in genetic studies affects inference about genetic architecture

This study demonstrates that inferences about a trait's genetic architecture, particularly regarding SNP heritability and the inferred direction of allelic effects, are significantly influenced by study design and cohort representation, with the latter often driven by the skewness of the trait distribution within the specific biobank.

The Gist

Imagine you are trying to understand the "recipe" for a complex dish, like a perfect chocolate cake. In genetics, this recipe is called the genetic architecture. It tells us which ingredients (genes) are in the cake, how much of each is needed, and whether they make the cake sweeter or more bitter.

For a long time, scientists assumed that this recipe was fixed and universal. If you studied the cake in Paris, the recipe should look the same as if you studied it in New York.

However, this new paper argues that where and how you taste the cake changes the recipe you write down.

Here is the breakdown of their findings using simple analogies:

1. The Problem: Different Kitchens, Different Tastes

The researchers looked at three massive "kitchens" (biobanks) where people's genetic data is stored:

• UK Biobank: Like a general community center. It recruited random volunteers from the general population.
• All of Us (AoU): Like a diverse community outreach program. It tried to include people from many different backgrounds, often those who are usually left out of studies.
• FinnGen: Like a hospital waiting room. It recruited people specifically because they were already diagnosed with health issues.

The team asked: If we look at the same trait (like height or diabetes) in these three different kitchens, do we get the same genetic recipe?

2. Finding #1: The "Strength" of the Recipe Changes

They found that some parts of the recipe changed depending on the kitchen.

• The Analogy: Imagine trying to hear a whisper in a quiet library (UK Biobank) versus a noisy construction site (FinnGen).
• The Result: In the "hospital" style kitchens (FinnGen) or the diverse "outreach" kitchens (AoU), the genetic signal for certain traits was "quieter" (lower heritability) than in the general population kitchen. This means it's harder to find the genetic causes of a disease if your study group is already full of sick people or has messy data, even if the people are genetically similar.

3. Finding #2: The "Direction" of the Ingredients (The Big Surprise)

This is the most fascinating part. They looked at Sign Bias.

• The Analogy: Imagine you are trying to figure out if a specific spice makes the cake better or worse.
  • In the UK Biobank (general population), they found that 99% of the rare spices seemed to make the cake worse (risk-increasing).
  • In the All of Us study, they found that only 72% of those same spices seemed to make the cake worse.
  • In FinnGen, it was even lower.

Why would the same spice look different in different kitchens?

The authors discovered the culprit: Skewness (or "The Tail of the Distribution").

• The Metaphor: Imagine a room full of people.
  • In a balanced room (UK Biobank), heights are spread out evenly.
  • In a skewed room (FinnGen or AoU for certain diseases), almost everyone is very tall, with only a few short people. The "tail" of the room is stretched out.

The researchers found that when a trait is skewed (e.g., almost everyone has the disease, or almost no one does), our math gets confused.

• If a disease is rare, it's easy to spot the "bad" genes because the few people who do have the disease stand out.
• But if a disease is common (or the data is messy), it becomes hard to tell the difference between a "bad" gene and a "good" gene. The math starts to guess that everything is bad just because the room is so full of "sick" people.

The "Skewness" Effect:
The paper proves that the shape of the data (how lopsided the group is) tricks the computer into thinking genes have a specific direction (risk-increasing) when they might not. It's like looking at a funhouse mirror; the mirror (the study design) distorts the reflection (the genetic signal), making it look like the genes are pushing in one direction when they are actually balanced.

4. The Simulation: Proving the Mirror Trick

To prove this wasn't just a fluke, they built a computer simulation.

• They created a fake world where genes were perfectly balanced (50% good, 50% bad).
• Then, they "sampled" people from this world in a biased way (only picking people who were very sick or very healthy).
• Result: Even though the genes were balanced in the simulation, the "biased" sample made it look like 90% of the genes were bad.
• Conclusion: The distortion came entirely from who was in the room (the skewness), not from the genes themselves.

The Bottom Line

This paper is a warning label for genetic research.

The Takeaway:
When scientists say, "Gene X causes Disease Y," they are actually saying, "Gene X causes Disease Y in this specific group of people we studied."

If you change the group (e.g., from a general population to a hospital clinic), the "recipe" changes. The direction of the genes can flip or shift simply because of how the data was collected.

Why does this matter?
It means we need to be careful when we try to use genetic data to predict health risks for everyone. If we only study people from one type of hospital or one specific country, our "genetic map" might be distorted. To get the true picture of human biology, we need to look at many different "kitchens" and understand how the cooking method changes the taste.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in understanding how study design and participant representation influence inferences about the genetic architecture of complex traits. Genetic architecture is defined as the joint distribution of allele frequencies and the direction/magnitude of their effects.

While these architectures are often assumed to be intrinsic properties of a trait, previous studies have shown variability in estimates across different cohorts. This variability is frequently attributed to differences in genetic ancestry. However, the authors argue that recruitment strategies (e.g., population-based volunteers vs. clinical registry-enriched cohorts) and phenotyping heterogeneity introduce systematic biases that are often overlooked. Specifically, the paper investigates whether the "representation" of a cohort (how participants are recruited and how traits are measured) distorts key summaries of genetic architecture, such as heritability, polygenicity, and the direction of allelic effects.

2. Methodology

The authors conducted a comparative analysis using Genome-Wide Association Study (GWAS) data from three major biobanks with distinct recruitment profiles:

• UK Biobank (UKB): A population-based volunteer cohort (broad representation).
• All of Us (AoU): A cohort emphasizing under-represented groups and diversity, largely derived from electronic health records (EHR).
• FinnGen: A diagnosis-enriched cohort drawn from clinical registries.

Data Processing:

• Traits: 14 traits were analyzed, including morphological (height, weight, BMI), hematological (blood cell counts, percentages), and disease endpoints (Type 1/2 Diabetes, Schizophrenia, Alzheimer's, Asthma).
• Ancestry Matching: To isolate recruitment effects from ancestry effects, the authors created ancestry-matched subsets of UKB and AoU (focusing on "White British" ancestry in both).
• Phenotyping: Traits were analyzed in their raw units without inverse-rank normalization or standardization to preserve the natural distribution (skewness) of the data.

Statistical Approaches:

1. Genetic Architecture Summaries:
   • SNP Heritability ($h^2_{SNP}$): Estimated using LD Score Regression (LDSC).
   • Effective Polygenicity: Estimated using S-LD4M to measure how heritability is distributed across variants.
   • Genetic Correlation ($r_g$): Estimated via bivariate LDSC to compare effect sizes between biobanks.
2. Sign Bias Analysis (Core Focus):
   • Defined as the mean direction (sign) of allelic effects for a set of variants.
   • Estimated using Adaptive Shrinkage (ash), an empirical Bayes method that accounts for uncertainty in effect sizes.
   • Calculated for rare minor alleles (MAF $\le$ 0.1%) across independent Linkage Disequilibrium (LD) blocks.
3. Simulation Studies:
   • Two simulation schemes were designed to test the hypothesis that trait skewness drives sign bias.
   • Scheme A: A unimodal population with equal trait-increasing and decreasing effects, sampled non-randomly above a threshold to induce skewness.
   • Scheme B: A tri-modal population where cohorts were sampled from specific modes to vary skewness while maintaining a symmetric underlying population.
   • Goal: To demonstrate that even with zero true biological sign bias, skewed sampling induces a biased estimation of effect directions.

3. Key Results

A. Variation in Standard Genetic Architecture Summaries

• SNP Heritability: AoU estimates were consistently lower than UKB estimates (average ~17% lower), even in ancestry-matched samples. This aligns with findings that disease-enriched or EHR-based cohorts often yield lower heritability, likely due to increased environmental variance or measurement noise.
• Polygenicity: Effective polygenicity varied little between biobanks, suggesting this metric is more robust to cohort differences.
• Genetic Correlation: For 6 out of 13 traits, the genetic correlation between UKB and AoU was significantly different from 1 (e.g., basophil percentage $r_g \approx 0.45$), indicating that effect sizes are not perfectly transferable across cohorts.

B. Sign Bias and Trait Skewness (The Primary Finding)

• Discrepancy in Sign Bias: The inferred direction of rare minor alleles varied drastically across biobanks.
  • Example: For Type 2 Diabetes, ~99% of rare minor alleles were inferred as risk-increasing in UKB, but only ~72% in AoU and ~57% in FinnGen.
  • Example: For Mean Corpuscular Hemoglobin, the sign bias was positive in UKB but negative in AoU.
• The Skewness Hypothesis: The authors found a remarkably strong correlation between the skewness of the trait distribution in the cohort and the inferred sign bias.
  • Traits with high positive skew (e.g., rare diseases with few cases, or skewed quantitative traits like white blood cell count in AoU) showed a strong bias toward inferring "risk-increasing" or "trait-increasing" effects for minor alleles.
  • Statistical Fit: A quadratic logit model explained 82% of the variance in sign bias for trait-associated SNPs and 97% for random SNPs solely based on trait skewness. Adding biobank-specific fixed effects provided negligible improvement.
• Simulation Validation: Simulations confirmed that when a population with no true sign bias is sampled in a skewed manner (e.g., enriching for high trait values), the resulting GWAS estimates exhibit a strong, monotonic increase in inferred sign bias.

C. Mechanism

The paper posits a statistical "coupling problem": In a skewed distribution, it is statistically easier to detect a strong positive correlation between a rare allele and a trait if the allele increases the trait (enriching the right tail) than to detect a negative correlation (depleting the right tail). This asymmetry in statistical power and uncertainty leads to a systematic overestimation of "increasing" effects in skewed cohorts.

4. Key Contributions

1. Decoupling Ancestry from Recruitment: The study demonstrates that differences in genetic architecture summaries can arise purely from recruitment and phenotyping strategies, even when genetic ancestry is matched.
2. Identification of Sign Bias as a Recruitment Artifact: The paper identifies "sign bias" not as a biological signal of natural selection, but largely as a statistical artifact driven by the skewness of the trait distribution within the study cohort.
3. Quantitative Explanation: It provides a mathematical framework (via simulations and regression) showing that trait skewness alone can explain the vast majority of cross-biobank variation in inferred allelic effect directions.
4. Critique of Data Transformation: The authors highlight that while inverse-rank normalization (common in GWAS) removes skewness and thus sign bias, it also removes biologically relevant information about the trait's natural distribution and effect magnitudes.

5. Significance and Implications

• Interpretation of Genetic Architecture: Inferences about the "map" between genetics and traits are cohort-dependent, not intrinsic to the trait. Researchers must be cautious when generalizing findings from one biobank to another.
• Study Design: The findings suggest that disease-enriched cohorts (like FinnGen) or EHR-based cohorts (like AoU) may suffer from reduced power to detect protective variants or may systematically misestimate the direction of rare variant effects due to distribution skewness.
• Future Directions: The paper calls for broader access to diverse recruitment profiles to ensure generalizability. It suggests that when comparing genetic architectures across populations, researchers must account for differences in recruitment and trait distribution, rather than attributing all discrepancies to biological or evolutionary differences.
• Methodological Caution: The results challenge the assumption that sign bias reflects natural selection (e.g., purifying selection). Instead, it suggests that observed biases may be artifacts of how participants are selected into the study.

In conclusion, the paper provides strong evidence that representation matters: the way a study is designed and who participates fundamentally alters the inferred genetic architecture, particularly the direction of allelic effects for rare variants.