🧬 biology

Shared trans-ancestry architecture of HLA-mediated disease risk in the All of Us Research Program

This study analyzes high-resolution HLA variation across 390,823 diverse participants in the All of Us Research Program to demonstrate that while many HLA-disease associations appear ancestry-specific due to differences in allele frequency and statistical power, the underlying biological effects are largely shared across genetic ancestries.

Original authors: Alison Motsinger-Reif, Kwangmi Ahn, John House, Adam Burkholder, Tam Tran, Joseph Breeyear, Cristina Justice, Jacqueline Durney, Alyssa Jones, Parker Reyes, Matthew Bailey, Mary Davis, Anthony Vicenti

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Alison Motsinger-Reif, Kwangmi Ahn, John House, Adam Burkholder, Tam Tran, Joseph Breeyear, Cristina Justice, Jacqueline Durney, Alyssa Jones, Parker Reyes, Matthew Bailey, Mary Davis, Anthony Vicenti, Jason Karnes, Jill Hollenbach, David Fargo, Geoffrey Ginsburg, Richard Woychik, Joshua Denny

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human immune system as a massive, high-tech security team guarding a city. The most important part of this team is the HLA region, a specific neighborhood in our DNA (chromosome 6) that acts like the ID badge scanner. These scanners read "badges" (alleles) to decide if a visitor is a friendly cell, a dangerous virus, or a rogue cancer cell. For a long time, scientists thought these scanners worked differently depending on which "ancestry" group you belonged to, like having different rulebooks for different neighborhoods.

But a giant new study using data from the All of Us Research Program has taken a fresh look at this. They didn't just peek at a few people; they analyzed the DNA of 390,823 participants from six different ancestry groups (African, Admixed American, East Asian, European, Middle Eastern, and South Asian). They even linked this genetic data to the medical records of 262,915 of those people to see who got sick with what.

Here is what they found, broken down into simple stories:

The "Missing" Scanners Were Just Hidden in the Crowd

One of the biggest questions was: "Do people from different ancestry groups have totally different sets of HLA scanners?"

The researchers found 4,780 distinct HLA alleles (different versions of the scanner). At first glance, it looked like many scanners were "private" to specific groups. For example, the European group seemed to have a huge list of unique scanners that no one else had.

The Twist: The paper argues that these "private" scanners weren't actually unique to that group's biology. Instead, it was a sampling issue. Think of it like a concert. If you have a stadium with 100,000 European fans but only 5,000 African fans, you are much more likely to spot a rare, unique t-shirt in the big crowd just because there are more people wearing them.

When the researchers "down-sampled" the European group (pretending they only had 50,000 people instead of the full crowd), the number of "unique" European scanners dropped dramatically. This suggests that the architecture of HLA disease risk is largely shared across all ancestries. The differences we see are often just because some groups were studied more deeply than others, not because their biology is fundamentally different.

The Great HLA Detective Hunt (PheWAS)

The team then played a massive game of "connect the dots." They checked 3,430 different clinical phenotypes (basically, thousands of different diseases and health conditions) against 363 common HLA alleles.

They found 1,461 significant connections.

The Good News: Even when a connection was only statistically "loud" enough to be heard in one ancestry group, the direction of the effect was almost always the same. If an allele made Europeans more likely to get a disease, it tended to make African or Admixed American people more likely to get it too, even if the signal was too quiet to detect in the smaller groups.
The Surprise: While we knew HLA was linked to autoimmune diseases (like Type 1 diabetes), this study found it was also linked to things you might not expect, like congenital conditions, cardiovascular issues, and even some psychiatric disorders.

Untangling the Knot: The "Independent Signals"

The HLA region is messy. It's like a tangled ball of yarn where many different threads (alleles) are stuck together. When you pull one, the whole ball moves. This made it look like dozens of different alleles were causing a single disease.

To fix this, the researchers used a "conditional modeling" technique. Imagine you are trying to figure out which specific person in a crowded room is shouting. You ask everyone to sit down one by one until only the loudest voice remains.

They did this for five major diseases: Type 1 diabetes, celiac disease, hypothyroidism, multiple sclerosis, and rheumatoid arthritis.

Before: They saw dozens of alleles linked to each disease.
After: They found that most of those signals collapsed into just 5 to 7 independent "shouts" (signals).
The Result: For example, Type 1 diabetes, which looked like it had 47 different culprits, was actually driven by just 7 independent signals. This proves that the complexity was an illusion caused by the DNA being so tightly packed together (linkage disequilibrium), not because there are dozens of separate biological mechanisms.

New Discoveries and Tools

The study didn't just re-explain old news; it found 42 candidate novel associations that hadn't been clearly linked before.

One example: HLA-DOA (a less famous part of the HLA family) was linked to Streptococcus pneumoniae infection.
Another: HLA-C was linked to drug-induced psychotic disorders.

To help everyone else see these patterns, the team built a HLA PheWAS Explorer. Think of this as a giant, interactive video game dashboard where you can zoom in on any gene, any disease, or any ancestry group to see the connections for yourself.

What This Means (and What It Doesn't)

The paper suggests that the "rules" of how HLA affects disease are shared across humanity, but our ability to see those rules depends on how many people we study.

What the paper rules out: It argues against the idea that HLA disease architecture is fundamentally different or "private" to specific ancestry groups. The differences are mostly due to sampling depth (how many people were studied) and statistical power, not unique biology.

What the paper admits is still tricky:

The study used short-read sequencing, which is great but can't perfectly resolve the most complex, knotted parts of the DNA (structural variations).
The data came from electronic health records, which rely on doctors' codes and might miss some details.
The study suggests that these findings could help improve risk prediction models, but it doesn't claim to have solved the problem of predicting disease for everyone yet.

In short, the HLA region is a shared, complex, and highly polymorphic neighborhood. While some parts look different depending on who you ask, the underlying blueprint is the same for everyone; we just needed a bigger, more diverse crowd to finally see the whole picture clearly.

Technical Summary: Shared Trans-Ancestry Architecture of HLA-Mediated Disease Risk in the All of Us Research Program

Problem Statement
The human leukocyte antigen (HLA) region is the most significant genetic contributor to numerous immune-mediated diseases, yet the extent to which HLA-mediated disease architecture is shared across diverse genetic ancestries remains unclear. Historically, large-scale genetic studies have focused predominantly on individuals of European ancestry, potentially mischaracterizing the global landscape of HLA risk. A critical challenge in multi-ancestry studies is distinguishing true biological heterogeneity from artifacts caused by disparities in sample size, allele frequency distributions, and linkage disequilibrium (LD) structures. Furthermore, the high polymorphism of the HLA region complicates the resolution of whether observed population-specific associations reflect distinct biology or simply differences in statistical power and sampling depth.

Methodology
This study leveraged the All of Us Research Program, a large, longitudinal cohort designed to include historically underrepresented populations. The analysis utilized high-resolution whole-genome sequencing (WGS) data from 390,823 unrelated participants across six genetic ancestry groups: African (AFR), Admixed American (AMR), East Asian (EAS), European (EUR), Middle Eastern (MID), and South Asian (SAS).

Genotyping: HLA alleles were inferred at G-group resolution (grouping alleles with identical protein sequences in the antigen recognition domain) for 20 HLA loci using Kourami (v0.9.6) and the IPD-IMGT/HLA database (v3.47.0). This approach identified 4,780 distinct alleles.
Diversity and LD Analysis: Allele frequencies and diversity metrics (including expected heterozygosity and the number of alleles required to capture 90% of cumulative frequency) were calculated. To assess the impact of sample size on apparent ancestry-specific variation, the authors implemented a downsampling framework, generating 20 independent subsets of 50,000 European individuals to compare against other groups. Linkage disequilibrium (LD) was evaluated at the allele level using squared Pearson correlation coefficients ( $r^2$ ).
Phenome-Wide Association Studies (PheWAS): Analyses were restricted to 262,915 participants with linked electronic health records (EHR). Associations were tested between 363 HLA alleles (frequency > 0.001) and 3,430 clinical phenotypes (phecodes) across the three largest ancestry groups (EUR, AFR, AMR). Logistic regression models were adjusted for age, sex, EHR follow-up duration, and the first 15 genetic principal components.
Meta-Analysis and Conditional Modeling: Ancestry-specific summary statistics were combined using fixed-effects inverse-variance-weighted meta-analysis. Heterogeneity was assessed using Cochran's Q and $I^2$ metrics. To resolve independent signals, forward stepwise conditional analyses were performed for five pre-specified complex traits (type 1 diabetes, celiac disease, hypothyroidism, multiple sclerosis, and rheumatoid arthritis).
Novelty Classification: A structured screening strategy was employed to classify associations as novel, gene-known/allele-unresolved, or allele-level novel by comparing results against curated catalogs (e.g., NHGRI-EBI GWAS Catalog) and applying LD-aware collapsing procedures.

Key Results

HLA Diversity: The study identified 4,780 distinct G-group alleles. While African (AFR) and European (EUR) groups exhibited the largest absolute numbers of distinct alleles, downsampling analysis revealed that a substantial fraction of "ancestry-private" alleles in the larger EUR group reflected sampling depth rather than true population specificity. When EUR was downsampled to 50,000 individuals, the number of private alleles dropped significantly, approaching levels observed in AFR and AMR.
Cross-Ancestry Architecture: Meta-analysis of 3,430 phenotypes identified 1,461 significant HLA-phenotype associations (FDR < 0.05). While many associations reached statistical significance in only one ancestry group, 80.3% of associations evaluated across multiple groups demonstrated concordant effect directions. This suggests that differences in significance are often driven by allele frequency, LD structure, and statistical power rather than fundamentally distinct biological mechanisms.
Conditional Fine-Mapping: Stepwise conditional modeling demonstrated that the apparent complexity of HLA associations could be resolved into a compact set of independent signals. For example, type 1 diabetes, initially associated with 47 distinct alleles, resolved into seven independent signals; celiac disease resolved into six; and multiple sclerosis into five. These independent signals were primarily localized to class II loci (e.g., HLA-DRB1, HLA-DQB1) with additional contributions from class I loci.
Pleiotropy and Novelty: Extensive pleiotropy was observed, with 70.7% of significant alleles associated with two or more distinct phecodes. After rigorous filtering for novelty and LD, 42 independent high-confidence signals were identified as novel at either the gene or allele level. These included associations with non-classical loci (e.g., HLA-DOA, HLA-DOB) and phenotypes beyond traditional immune domains, such as infectious diseases, psychiatric disorders, and congenital traits.
Resource Release: The study introduced the "HLA PheWAS Explorer," an interactive dashboard allowing users to query results by allele, trait, and ancestry, facilitating the exploration of high-dimensional association data.

Significance and Claims
The authors claim that this study provides a framework for distinguishing true biological heterogeneity from sampling-driven detectability differences in HLA-mediated disease risk. By utilizing high-resolution WGS data across a multi-ancestry cohort, the work demonstrates that HLA disease architecture is largely shared across populations, with apparent differences often attributable to statistical power and LD patterns rather than unique ancestry-specific biology.

The paper emphasizes that dense allele-level association signals can be collapsed into a smaller number of independent gene-distinct effects, reinforcing the utility of conditional modeling in interpreting HLA data. Furthermore, the identification of novel associations in non-classical loci and diverse disease domains highlights the value of phenome-wide approaches in uncovering previously underexplored manifestations of HLA variation. The authors conclude that multi-ancestry, sequence-based analysis offers benefits beyond mere discovery expansion, providing a more precise understanding of how HLA variation shapes human disease and suggesting that polygenic risk scores should be updated to explicitly include high-resolution HLA alleles inferred from WGS to improve calibration and transferability.