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

The human leukocyte antigen (HLA) region is the strongest genetic contributor to many immune-mediated diseases. These range from type 1 diabetes to rheumatoid arthritis. These genes encode proteins that present peptide antigens (small protein fragments) to T cells. This process teaches the immune system what belongs to the body and what is a pathogen. Because this region is incredibly diverse, it acts as a critical shield against infection. However, it also creates a massive challenge for genomic medicine. We must determine if the genetic drivers of disease are universal or specific to certain ancestral lineages.

For decades, our understanding of HLA-mediated risk has been skewed toward individuals of European ancestry. This leaves a fundamental question unanswered. Is the "architecture" of immune disease—the specific way certain alleles (different versions of a gene) drive illness—truly different across global populations? Or have we simply failed to see shared patterns because our data is lopsided? This paper uses the diverse All of Us Research Program to argue that much of what looks like population-specific biology is actually an artifact of how we collect and analyze data.

The trap of apparent ancestry-specific variation

The central difficulty in studying the HLA region is its extreme polymorphism (the tendency of a gene to have many versions). It also has complex linkage disequilibrium (LD). LD describes the tendency of certain genetic variants to be inherited together as a block, or haplotype, rather than being shuffled. In the HLA region, these blocks are so tightly packed that it is difficult to distinguish which specific allele is actually causing a disease. Often, another allele is merely "tagging along" because it is nearby.

Historically, large-scale studies relied on SNP-based imputation (using known markers to guess missing DNA). This approach often fails to capture the full complexity of the HLA region. It especially struggles in non-European populations. Consequently, researchers often observe "ancestry-private" alleles. These are variants that appear to exist only in one group. This leads to the assumption that disease risk is biologically unique to certain ancestries. However, the authors note this can be a mirage caused by unequal discovery depth. If you search a small forest, you will find fewer trees than in a vast one. This does not mean the small forest contains a different species.

Resolving the HLA landscape through graph-based inference

To overcome these limitations, the researchers implemented a high-resolution workflow. Their approach relied on three primary pillars:

1. High-resolution genotyping: Instead of imputation, the authors used whole-genome sequencing (WGS) and Kourami (v0.9.6). Kourami is a graph-guided assembly tool. This allowed them to identify HLA alleles at "G-group" resolution. This resolution groups alleles that share identical protein sequences in the critical antigen-recognition domain. This ensures the functional impact of the variant is captured.
2. Massive-scale PheWAS: They performed phenome-wide association studies (PheWAS). This method scans a single genetic variant against thousands of different clinical traits. They tested 3,430 different phecodes (standardized clinical descriptors). This allowed them to see if an allele correlates with skin conditions, neurological issues, or respiratory infections.
3. Stepwise conditional modeling: To untangle the mess of linkage disequilibrium, the authors used stepwise selection. They started with all significant alleles for a disease. They then "conditioned" on them. This mathematically removes their influence to see if other independent signals remain. This process collapses hundreds of redundant signals into a few independent biological drivers.

Evidence of a shared biological architecture

The scale of the study is immense. It involves 390,823 participants and identifies 4,780 distinct HLA alleles across 20 loci. The most striking result is the consistency of the disease signals.

The authors report that among associations evaluated in at least two ancestry groups, 80.3% demonstrated concordant effect directions [Figure 4c]. This means that if an allele increases disease risk in Europeans, it almost always increases risk in African or Admixed American populations too. Even if the statistical significance varies, the direction stays the same. To prove that "private" alleles were a matter of sampling, the researchers performed a downsampling experiment. They repeatedly sampled subsets of the larger European cohort. They found that the number of "European-private" alleles dropped significantly [Figure 2b]. This brought the numbers closer to levels seen in other groups.

Furthermore, conditional modeling revealed that the apparent complexity of HLA-disease links is often overstated. For common complex traits like type 1 diabetes or multiple sclerosis, dozens of correlated alleles collapse into just five to seven independent signals .

This demonstrates that a compact set of gene-distinct effects drives much of the observed risk across the MHC (Major Histocompatibility Complex) region.

Limitations in the high-resolution view

Several technical boundaries remain. First, the use of short-read whole-genome sequencing still struggles with complex structures. It cannot fully resolve structural variation or long-range phasing (determining which alleles sit on which inherited chromosome).

Second, the "G-group" resolution focuses on the antigen-recognition domain. While this is functionally relevant, it may overlook subtle regulatory variations located just outside these domains. These could influence how strongly a disease manifests. Finally, the authors note uneven representation in the cohort. The large European sample provides more statistical power than the smaller Middle Eastern or South Asian groups. This imbalance might still mask some subtle, population-specific nuances.

The verdict: a unified framework for immune risk

The study provides a definitive answer regarding whether HLA-mediated disease architecture is shared across ancestries. Yes, it is largely shared. Apparent differences are primarily a function of allele frequency, linkage disequilibrium, and statistical power.

For researchers, the takeaway is clear. We should not treat HLA-based risk as a series of disconnected, ancestry-specific puzzles. Instead, we should view them as a single, complex system. We are simply viewing them through different lenses of frequency and visibility. The developers created the interactive HLA PheWAS Explorer (https://manticore.niehs.nih.gov/AoU_HLA_PheWAS_Explorer) to help the community. This tool moves from simple scans to a more nuanced understanding of human immunity. Moving forward, polygenic risk scores (PRS) must be updated. They should include these high-resolution WGS-inferred alleles. This will ensure that precision medicine works for everyone, not just the populations that were easiest to sequence.

Figures from the paper