Biobank-Scale Polygenic Prediction in Admixed Populations Using Local Ancestry via the Group Lasso

The paper introduces "Combine," a scalable group lasso framework that integrates local ancestry dosages with genotype data to significantly enhance polygenic risk prediction accuracy and interpretability in admixed populations, outperforming existing multi-ancestry methods while matching optimized individual-level models.

The Gist

Imagine your DNA is a massive, intricate quilt. For a long time, scientists tried to understand this quilt by looking at the whole thing as one big, uniform color. They built "prediction maps" (called Polygenic Risk Scores) based on the patterns found in European populations. But when they tried to use these maps on people with mixed heritage—whose quilts are patchworks of European, African, Asian, and Indigenous American fabrics—the maps failed. The patterns didn't match, and the predictions were often wrong.

Why? Because in a mixed-heritage person, different parts of their DNA come from different ancestors. A specific gene might sit on a "European patch" in one person and an "African patch" in another. The old maps didn't know how to read the specific patch a gene was sitting on; they just saw the gene and assumed it acted the same way everywhere.

Enter "Combine": The Smart Quilt Reader

This paper introduces a new tool called Combine. Think of Combine not just as a map reader, but as a super-smart tailor who looks at every single patch of the quilt before making a prediction.

Here is how it works, using some everyday analogies:

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

Imagine you are trying to predict how fast a car will go based on its engine size.

• The Old Way: You assume all cars with a 2.0L engine go 100 mph.
• The Reality: A 2.0L engine in a heavy truck goes 40 mph. A 2.0L engine in a race car goes 150 mph.
• The DNA Issue: In mixed-ancestry people, a specific genetic "engine" (variant) might be attached to a "heavy truck" background (African ancestry) in one person and a "race car" background (European ancestry) in another. If you ignore the background, your prediction is off.

2. The Solution: The "Local Ancestry" Lens

The authors developed Combine, which uses a technique called Local Ancestry Inference.

• The Analogy: Instead of asking, "What is this person's overall background?" (e.g., "They are 50% European, 50% African"), Combine asks, "What is the background of this specific patch of DNA right here?"
• It looks at a gene and says, "Ah, this gene is sitting on an African patch. Let's use the rules for African patches." Then it moves to the next gene and says, "This one is on a European patch. Let's use the European rules."

3. The Secret Sauce: The "Group Lasso"

How does Combine handle millions of genes without getting a headache? It uses a mathematical trick called Group Lasso.

• The Analogy: Imagine you are sorting a giant pile of mixed Lego bricks. You have red bricks, blue bricks, and green bricks.
  • Old Method: You try to sort every single brick individually. It takes forever, and you might miss the pattern.
  • Combine's Method: You group the bricks by color and by the specific shape they are attached to. You say, "If I have a Red Brick attached to a Blue Base, treat them as a team."
  • The Result: The algorithm looks at the whole team (the gene + its local ancestry background) and decides: "This team is important, let's keep it," or "This team doesn't matter, let's ignore it." This makes the process incredibly fast and efficient, even for huge datasets (like 100,000 people).

4. What Did They Find?

The researchers tested Combine on nearly 100,000 mixed-heritage people from the "All of Us" research program.

• The Score: It predicted health risks (like blood cell counts, cholesterol, and kidney disease) much better than the current best methods. For white blood cell counts, it was 144% better than the previous best tool.
• The "Aha!" Moment: Because Combine looks at the specific patches, it can explain why a prediction was made.
  • Example: It found a gene that lowers white blood cell counts, but only when that gene is sitting on an African ancestry patch. On European patches, the gene does nothing. The old tools missed this completely because they just averaged the two effects together.

5. The "External Cheat Sheet"

The authors also showed that you can give Combine a "cheat sheet" from other studies. If a gene is already known to be important for cholesterol in other big studies, Combine gives that gene a "head start" in the learning process. This made the predictions for cholesterol even more accurate without needing to throw away any data.

The Big Picture

Combine is like upgrading from a blurry, black-and-white photo of a quilt to a high-definition, color-coded map.

• It respects the complexity of mixed heritage.
• It doesn't just guess; it explains where and why a gene matters.
• It works fast enough to handle the massive databases of modern medicine.

By using this tool, doctors and scientists can finally build fair, accurate health prediction models for everyone, not just people of European descent. It's a huge step toward making genetic medicine work for the whole world.

Technical Summary

1. Problem Statement

Polygenic Risk Scores (PRS) trained on European-ancestry cohorts often fail to generalize to non-European and admixed populations. This performance gap arises from:

• Linkage Disequilibrium (LD) and Allele Frequency Differences: Correlations between variants differ across ancestries, causing "tagging" SNPs to have varying effect sizes depending on the local genetic background.
• Limitations of Current Approaches:
  • Single-Ancestry Models: Cannot handle individuals with mosaic genomes (segments of different ancestries).
  • Global Ancestry Adjustments: Averaging ancestry proportions across the whole genome ignores local heterogeneity.
  • Summary-Statistic Methods (e.g., PRS-CSx): Rely on ancestry-matched LD reference panels, which are ill-suited for admixed individuals whose genomes contain multiple ancestries simultaneously.
  • Existing Local Ancestry Methods: Often limited to univariate association testing or require pre-selected variants and summary statistics, making them non-scalable to biobank-sized individual-level data ($N \approx 100,000$).

There is a critical need for a scalable, individual-level framework that jointly models genotype and local ancestry (the specific ancestry of a chromosomal segment) to improve prediction accuracy and interpretability in admixed populations.

2. Methodology: The "Combine" Framework

The authors introduce Combine, a sparse regression framework designed for biobank-scale data that integrates local ancestry inference (LAI) directly into polygenic prediction.

Core Algorithmic Design

• Group Lasso Penalty: Instead of estimating a single effect per SNP, Combine expands the feature space. For each variant, it creates a group of features comprising the SNP genotype and local ancestry dosages. A group lasso penalty is applied to these groups, allowing the model to select entire loci while shrinking uninformative groups to zero.
• Two Model Variants:
  1. Combine-R (Regular): Uses unphased allele dosages ($G$) combined with local ancestry dosages ($Y$) per variant. The feature block for variant $j$ is $[G_j, Y_j]$. This assumes a shared causal effect across ancestries but allows for ancestry-specific offsets.
  2. Combine-S (Specific): Uses phased data to estimate ancestry-specific allele dosages ($H$) alongside local ancestry dosages ($Y$). The feature block is $[H_j, Y_j]$. This allows the model to estimate distinct SNP effect sizes for each ancestry (e.g., an SNP might have a positive effect in African haplotypes and a negative effect in European haplotypes).
• Scalability & Optimization:
  • Implemented using adelie, a path-wise block-coordinate descent solver.
  • Matrix-Free Operations: The design matrix is never fully materialized in memory. Instead, it uses compressed, chunk-based sparse representations of SNP and ancestry blocks.
  • Efficiency: This allows training on millions of variants and hundreds of thousands of samples on standard high-memory nodes (completing genome-wide fits in <20 minutes per fold).
• External Priors: The framework supports weighted group lasso, where penalty weights are adjusted based on external GWAS summary statistics (e.g., GLGC for LDL). This steers sparsity toward well-supported loci without pre-filtering variants.

Data Processing

• Cohort: 99,298 admixed participants from the All of Us Research Program.
• Local Ancestry Inference (LAI): Performed using Gnomix with a multi-ancestry reference panel (6 ancestries: AFR, AMR, EAS, EUR, OCE, SAS).
• Covariates: Models adjust for age, sex, and global ancestry proportions (to control for population structure) while explicitly modeling local ancestry.

3. Key Contributions

1. Novel Framework: Introduction of Combine, the first scalable, individual-level sparse regression method that jointly models SNPs and local ancestry using group lasso, eliminating the need for ancestry-matched LD panels or summary-statistic pre-filtering.
2. Scalability: Demonstrated ability to train on biobank-scale data ($N \approx 100k$, $P \approx 1M+$) with complex ancestry models using standard hardware, achieving 3–6x faster per-feature training times compared to state-of-the-art SNP-only models (snpnet).
3. Interpretability: The model provides locus-level decomposition, distinguishing between shared allelic effects (the SNP effect itself) and ancestry-linked tagging (effects driven by LD patterns specific to an ancestry).
4. Ancestry-Dependent Discovery: The ability to detect "sign flips" (where a variant is protective in one ancestry but a risk factor in another) and differential penetrance, which are invisible to standard single-effect models.

4. Key Results

The method was benchmarked against PRS-CSx (state-of-the-art summary-statistic method) and snpnet (highly optimized individual-level SNP-only model) across nine phenotypes.

• Predictive Performance:
  • vs. PRS-CSx: Combine-R showed massive improvements, with relative gains ranging from 25.0% (CRP) to 144.2% (White Blood Cell count).
  • vs. snpnet: Combine-R matched or improved performance on 7 out of 9 phenotypes.
  • LDL Cholesterol: While snpnet initially outperformed Combine-R, incorporating external GWAS priors into Combine-R resulted in a 4.1% improvement over the snpnet baseline.
• Biological Interpretability & Validation:
  • Duffy Antigen (ACKR1): Correctly identified the strong, ancestry-specific effect of the Duffy-null variant on WBC count, which is nearly absent in non-African haplotypes.
  • CETP Region: Recovered known ancestry-divergent effects on HDL cholesterol, with effect sizes varying significantly between European and African haplotypes.
  • Sign Flips: Identified variants with opposing effects in different ancestries:
    • TENM2 (Kidney Disease): Protective in European haplotypes, risk-increasing in African haplotypes.
    • AKAP6 (Colorectal Cancer): Protective in European haplotypes, risk-increasing in African/Indigenous American haplotypes.
• Computational Efficiency: Achieved training times of under 20 minutes per fold for genome-wide data, confirming scalability without specialized HPC clusters.

5. Significance and Impact

• Equity in Genomics: Provides a practical path toward equitable polygenic prediction for admixed populations (a growing fraction of the global population), addressing the current bias where PRS models fail for non-European groups.
• Beyond "Black Box" Prediction: Unlike standard PRS, Combine offers locus-level interpretability. It can disentangle whether a prediction is driven by a causal allele shared across ancestries or by ancestry-specific LD patterns, aiding in the discovery of ancestry-dependent biology.
• Methodological Advancement: Moves the field from univariate GWAS and summary-statistic approaches toward high-dimensional, individual-level sparse regression that respects the mosaic nature of admixed genomes.
• Future Directions: The framework is designed to be extensible to multi-trait settings and can incorporate diverse external priors, setting a new standard for how local ancestry should be integrated into genomic prediction pipelines.

In summary, Combine represents a significant leap forward in polygenic risk modeling, proving that explicitly modeling local ancestry at the individual level is not only computationally feasible at biobank scales but also essential for accurate, interpretable, and equitable genetic prediction in admixed populations.