Abstract
Genetic risk models, such as Polygenic Risk Scores (PRS), often suffer from a "diversity gap," performing poorly on non-European and admixed populations due to confounding population structures. This paper introduces DisPred, a deep-learning framework that utilizes a disentangling autoencoder to separate ancestry-related genetic patterns from disease-specific signals. By learning "ancestry-invariant" representations, DisPred significantly improves Alzheimer’s risk prediction for minority groups without requiring ancestry labels during inference.
Introduction
In genomics, Polygenic Risk Scores (PRS) are the standard tool for predicting disease likelihood based on DNA. However, a massive "diversity gap" exists because the vast majority of genomic data comes from individuals of European descent. When these models are applied to minority populations or "admixed" individuals (those with mixed ancestral backgrounds), they perform poorly. This occurs because models become "confounded" by population structure, mistaking ancestral genetic markers for disease risk markers. This hurdle significantly impacts the promise of equitable precision medicine.
Methodology
The authors introduce DisPred, a three-stage deep-learning framework designed to learn "ancestry-invariant" representations of genetic data.
- Disentangling Autoencoder (Disentgl-AE): The model compresses high-dimensional genotype data into two distinct latent vectors: $z_a$ (ancestry-specific) and $z_d$ (disease-specific).
- Supervised Contrastive Loss: To force this separation, a contrastive learning strategy ensures that $z_a$ vectors of people with the same ancestry are similar, while $z_d$ vectors are only similar if they share the same disease status. This effectively "unmixes" the ancestry signal from the disease signal.
- Prediction and Ensembling: A linear predictor is trained on the "purified" $z_d$ representation. DisPred then uses an ensemble approach, combining latent-space predictions ($p_z$) with predictions from original raw data ($p_x$) to maximize accuracy.
Crucially, while the model uses ancestry labels during training, it does not require ancestry information to make predictions on new patients.
Datasets and Evaluation
DisPred was evaluated using two major Alzheimer’s Disease (AD) cohorts:
- ADSP (Alzheimer’s Disease Sequencing Project): ~11,640 participants and ~3,892 genetic variants.
- UK Biobank (UKB): ~461,579 participants and ~4,967 variants.
The researchers used SNPWeight software to estimate precise ancestry percentages, testing the model on European (EUR), African (AFR), Asian (ASN), and Admixed (MIX) subgroups.
Key Insights and Results
The core innovation of DisPred is the mathematical enforcement of ancestry invariance. By using a deep autoencoder to create $z_d$, the model captures complex, non-linear genetic interactions while maintaining stability through a linear final predictor.
As shown in the UMAP visualizations, the ancestry-specific representation clearly clusters individuals by population, whereas the phenotype-specific representation (intended for disease prediction) shows a more mixed distribution, indicating successful disentanglement of the two signals.
Key results include:
- Superior Performance in Minority Groups: In the ADSP cohort, DisPred achieved the highest Area Under the Curve (AUC) for African American and admixed participants.
- Closing the Gap: For admixed individuals in the UK Biobank, DisPred achieved an AUC of 0.6692, nearly identical to traditional model performance on pure European samples (0.6695).
- Robustness: DisPred showed resilience to "dataset shift" when trained on ADSP and tested on UK Biobank, and consistently outperformed standard models when tested on mismatched ancestries.
Discussion and Limitations
While DisPred offers a significant step toward equitable genomics, it faces some limitations. The multi-stage deep learning architecture makes "feature importance" analysis difficult; the model can predict risk but cannot easily pinpoint specific causal genes. Future work may involve more advanced architectures or different similarity-enforcing losses. Additionally, UMAP visualizations suggest that the "Admixed" group still shows some overlap in ancestry-specific space, indicating room for improved regularization.
Summary updated