Improving genetic risk prediction across diverse population by disentangling ancestry representations

Prashnna Gyawali, Yann Le Guen, Xiaoxia Liu, Hua Tang, James Zou, Zihuai He

TL;DR

Researchers have developed DisPred, a deep-learning framework that improves genetic risk prediction for diverse populations by "disentangling" ancestry from disease-related genetic signals. By stripping away the confounding effects of ancestral background, the model achieves significantly higher accuracy in minority and admixed populations compared to traditional methods, all without requiring self-reported ancestry at test time.

What Problem Does This Paper Address?

In the world of precision medicine, Polygenic Risk Scores (PRS) are the "gold standard" for predicting a person's likelihood of developing complex diseases like Alzheimer’s based on their DNA. However, there is a massive equity gap: most genomic data comes from individuals of European descent.

When these models are applied to non-European or "admixed" individuals (people with diverse ancestral backgrounds, such as African Americans), they often fail. This happens because the models get "confused" by population structure—genetic variations that are tied to ancestry rather than the actual disease. Essentially, the ancestry signal acts as a loud background noise that drowns out the disease signal, leading to biased predictions and potential misdiagnoses in minority groups.

Methods

The authors introduce DisPred, a three-stage framework designed to isolate the "pure" disease signal:

  1. Disentangling Autoencoder (Disentgl-AE): This is the core innovation. Using a deep-learning architecture, the model compresses genomic data into a "latent space." It uses a supervised contrastive loss to force the model to separate this space into two distinct parts: an ancestry-specific representation ($z_a$) and a phenotype-specific (disease) representation ($z_d$).
  2. Phenotype Prediction: A linear model is trained specifically on the $z_d$ representation. Because $z_d$ has been stripped of ancestry info but retains non-linear genetic relationships, the predictor becomes more robust across different groups.
  3. Ensemble Modeling: Finally, the system combines the predictions from the disentangled representation with predictions from the original raw data to maximize accuracy.

Unlike previous "adversarial" approaches that are notoriously hard to train, DisPred uses a more stable contrastive learning approach to achieve this separation.

Datasets

The study evaluated DisPred using two major genomic cohorts:

  • Alzheimer’s Disease Sequencing Project (ADSP): A multi-ethnic cohort (including African American, European, and Hispanic participants) with ~11,640 participants and ~3,900 genetic variants.
  • UK Biobank (UKB): A massive dataset of ~461,579 participants and ~5,000 variants. While UKB is predominantly European, it provided a critical test for "dataset shift"—applying a model trained on one cohort to a completely different one.

Key Insights and Ideas

  • Ancestry as a Confounder, Not a Feature: Traditional models treat ancestry as a fixed variable. DisPred treats it as something to be "unlearned" or filtered out to find the universal genetic drivers of a disease.
  • The "Invariance" Win: UMAP visualizations (a way to see high-dimensional data in 2D) showed that while the ancestry latent space ($z_a$) clustered people perfectly by their background, the disease latent space ($z_d$) was a "mixed bag" where ancestry labels were scattered. This proved the model successfully created an ancestry-invariant representation.
  • No Labels Needed at Test Time: One of the most practical "aha" moments is that DisPred doesn't need to know a patient's self-reported race or ancestry to work. It figures out the relevant disease signals automatically, bypassing the messy and often inaccurate nature of self-reported labels in clinical settings.

Key Results

  • Superior Performance in Minority Groups: In African American (ADSP AFR) and admixed test groups, DisPred consistently achieved the highest Area Under the Curve (AUC), significantly outperforming standard PRS and even standard Neural Networks.
  • Robustness to "Ancestral Mismatch": When a model trained on European data was tested on African American data, DisPred maintained its predictive power while other models' performance plummeted.
  • Handling Heterogeneity: As individuals became more "admixed" (having a more complex mix of ancestral backgrounds), DisPred’s performance degraded much more slowly than traditional linear models (Lasso) or standard deep learning.

Limitations and Future Work

  • The "Black Box" Problem: Because the model uses a multi-stage deep learning process, it is difficult to perform "feature importance" analysis. We know the model is more accurate, but we don't yet know exactly which genetic variants it is prioritizing for admixed individuals.
  • Architecture Optimization: The authors used a relatively standard Multi-Layer Perceptron (MLP). They suggest that more advanced architectures or newer similarity-enforcing losses could further boost performance.
  • Broadening the Scope: This study focused specifically on Alzheimer’s Disease. Future work is needed to see if this disentangling approach works as effectively for other complex traits like diabetes or heart disease.

Summary updated

Abstract

Abstract Risk prediction models using genetic data have seen increasing traction in genomics. However, most of the polygenic risk models were developed using data from participants with similar (mostly European) ancestry. This can lead to biases in the risk predictors resulting in poor generalization when applied to minority populations and admixed individuals such as African Americans. To address this bias, largely due to the prediction models being confounded by the underlying population structure, we propose a novel deep-learning framework that leverages data from diverse population and disentangles ancestry from the phenotype-relevant information in its representation. The ancestry disentangled representation can be used to build risk predictors that perform better across minority populations. We applied the proposed method to the analysis of Alzheimer’s disease genetics. Comparing with standard linear and nonlinear risk prediction methods, the proposed method substantially improves risk prediction in minority populations, including admixed individuals, without needing self-reported ancestry information.
BibTeX
@inproceedings{gyawali2023improving,
  author = {Prashnna Gyawali and Yann Le Guen and Xiaoxia Liu and Hua Tang and James Zou and Zihuai He},
  title = {{Improving genetic risk prediction across diverse population by disentangling ancestry representations}},
  booktitle = {Communications},
  year = {2023},
  doi = {10.1038/s42003-023-05352-6},
}
Figure