Local Ancestry Modeling for Multi-Ancestry Polygenic Risk Scores

This direction is merged into another direction: population generalizability (ancestry bias). Its last version stays readable and citable here.

Version 1 of 126 papers40 conceptsfrozen 2026-10-04

Most genetic risk models today suffer from a "portability" problem: they are trained on data from individuals of European descent and lose significant accuracy when applied to other populations. This is not just a statistical quirk; it is a clinical barrier. As Kristjan Eerik Kaseniit and colleagues demonstrated in Genetic-Ancestry Analysis on >93,000 Individuals Undergoing Expanded Carrier Screening Reveals Limitations of Ethnicity-Based Medical Guidelines, relying on self-reported ethnicity can miss up to 77% of genetic carriers because ethnicity is a poor proxy for biological risk.

To solve this, researchers are moving toward ancestry-aware modeling, which treats the genome not as a uniform block, but as a mosaic of ancestral segments.

The Mosaic Genome and Local Ancestry

The core intuition behind this research is that individuals with admixed backgrounds (such as African Americans or Latinx individuals) possess "mosaic" genomes. Over generations of recombination, their chromosomes have become patchworks where one segment might have West African origins while the adjacent segment is Northern European.

This matters because the "effect size" of a genetic variant—how much it actually increases your risk for a disease—is often dependent on its genomic neighborhood. A variant might be a strong predictor of disease when it sits on a chromosomal segment of African origin but have no effect when it sits on a European segment. This phenomenon, often driven by differences in linkage disequilibrium (how variants are inherited together), means that a single "global" ancestry score is insufficient.

As Quan Sun and others argue in Improving polygenic risk prediction in admixed populations by explicitly modeling ancestral-specific effects via GAUDI, we must account for local ancestry to achieve true precision.

Modeling Ancestry-Specific Effects with GAUDI and HAUDI

To capture this mosaic complexity, researchers developed GAUDI (and its high-dimensional successor, HAUDI). These frameworks use a "fused lasso" approach to assign different weights to the same genetic variant depending on the ancestral origin of the local chromosomal segment it inhabits.

In An Efficient Lasso Framework for Admixture-Aware Polygenic Scores, Franklin P Ockerman explains that by re-parameterizing these complex ancestral interactions into a standard LASSO problem, they can reduce computation time by over 99%. This allows the model to learn which variants are "pan-population" (working the same way across all ancestries) and which are "ancestry-specific." For traits like white blood cell count, this method improved prediction accuracy by 64% over traditional models.

Calibration and Relative Risk

While building new models from scratch is the gold standard, we often have to work with existing European-centric scores. Lars G. Fritsche and Xiang Zhou have shown in On Cross-ancestry Cancer Polygenic Risk Scores that while absolute risk scores don't translate well across populations, the relative ranking often does.

If you take a European-trained score and apply it to a South Asian cohort, the "raw" numbers might be wrong, but the people the model identifies as being in the "top 5% of risk" are often still the people at the highest risk within that specific group. By performing post-hoc calibration—adjusting the score based on an individual's specific ancestry proportions—researchers can "rescue" these models. Luciana B. Vargas demonstrated in Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults that this calibration prevents "statistical ghosts" (inflated risk estimates) and ensures that high-risk classifications actually correlate with biological reality.

The Strategy Depends on the Trait

There is no one-size-fits-all solution for multi-ancestry genetics. Brieuc Lehmann and Gil McVean found in Optimal strategies for learning multi-ancestry polygenic scores vary across traits that the best approach depends on the "genetic architecture" of the disease:

  1. Highly Polygenic Traits: Diseases influenced by thousands of tiny variants benefit most from massive, diverse datasets where information is "borrowed" across all ancestries.
  2. Population-Specific Hotspots: Traits influenced by specific, high-impact variants (like APOL1 in kidney disease, as noted by Atlas Khan in Genome-wide polygenic score with APOL1 risk genotypes predicts chronic kidney disease across major continental ancestries) require models that explicitly prioritize those ancestry-matched markers.

By moving from broad ethnic categories to high-resolution local ancestry modeling, genomic medicine is becoming more equitable, ensuring that the predictive power of the DNA revolution is available to everyone, regardless of their ancestral background.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_local_ancestry_modeling_for_multi_ancest_99e30a27,
  title = {Local Ancestry Modeling for Multi-Ancestry Polygenic Risk Scores},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-ancestry-modeling-for-multi-ancestry-polygenic-risk-scores/rdr_99e30a27764246d2ae65ec25481a50b8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_local_ancestry_modeling_for_multi_ancest_99e30a27_v1,
  title = {Local Ancestry Modeling for Multi-Ancestry Polygenic Risk Scores},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-ancestry-modeling-for-multi-ancestry-polygenic-risk-scores/rdr_99e30a27764246d2ae65ec25481a50b8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}