GWAS-based Machine Learning for Prediction of Age-Related Macular Degeneration Risk

Qi Yan, Yale Jiang, Heng Huang, Anand Swaroop, Emily Y. Chew, Daniel E. Weeks, Wei Chen, Ying DingmedRxiv 2019 · DOI

Abstract

Numerous independent susceptibility variants have been identified for Age-related macular degeneration (AMD) by genome-wide association studies (GWAS). Since advanced AMD is currently incurable, an accurate prediction of a person’s AMD risk using genetic information is desirable for early diagnosis and clinical management. In this study, genotype data of 32,215 Caucasian individuals with age above 50 years from the International AMD Genomics Consortium in dbGAP were used to establish and validate prediction models for AMD risk using four different machine learning approaches: neural network, lasso regression, support vector machine, and random forest. A standard logistic regression model was also considered using a genetic risk score. To identify feature SNPs for AMD prediction models, we selected the genome-wide significant SNPs from GWAS. All methods achieved good performance for predicting normal controls versus advanced AMD cases (AUC=0.81–0.82 in a separate test dataset) and normal controls versus any AMD (AUC=0.78–0.79). By applying state-of-the-art machine learning approaches on large AMD GWAS data, the predictive models established can provide an accurate estimation of an individual’s AMD risk profile across their lifespan based on comprehensive genetic information.

Introduction

Age-Related Macular Degeneration (AMD) is a progressive neurodegenerative disease and a primary cause of vision loss among the elderly. Because advanced forms of the disease—such as geographic atrophy (GA) or choroidal neovascularization (CNV)—are currently incurable, early diagnosis is essential for effective clinical management. While previous genome-wide association studies (GWAS) have identified numerous susceptibility variants, translating these genetic markers into a personalized risk profile remains a challenge. This study bridges that gap by applying machine learning (ML) to handle the complex, non-linear interactions within the human genome, moving beyond simple association to robust predictive modeling.

Methods

The researchers developed a multi-step pipeline to evaluate and compare various predictive models:

  1. Feature Selection: A GWAS was conducted on a training subset to identify the most significant Single Nucleotide Polymorphisms (SNPs), focusing on those with genome-wide significance ($P < 5 \times 10^{-8}$).
  2. Model Comparison: Five modeling approaches were tested:
  3. Neural Networks (NN): A multi-layer feedforward network.
  4. Lasso Regression: A linear model utilizing "shrinkage" to prevent overfitting.
  5. Support Vector Machines (SVM) & Random Forests (RF): Standard high-performance ML classifiers.
  6. Genetic Risk Score (GRS): A traditional approach using a weighted sum of risk alleles.
  7. Interpretability: To address the "black box" nature of neural networks, the study employed LIME (Local Interpretable Model-Agnostic Explanations) to visualize the genetic features contributing most to specific predictions.

Datasets

The study primarily utilized data from the International AMD Genomics Consortium (via dbGaP), comprising 32,215 Caucasian individuals over the age of 50. This cohort included normal controls, intermediate cases, and advanced cases.

Demographic characteristics of the study population.

In addition to the primary dataset, the models were validated using:

  • UK Biobank: An independent set of 383 Caucasian subjects to test generalizability.
  • Ancestry Testing: A mixed non-Caucasian dataset (African, Asian, and unknown ancestry) to evaluate cross-ethnic performance.

Results and Key Insights

The predictive models demonstrated high accuracy across several metrics:

  • Predictive Performance: For the primary task of differentiating Normal vs. Advanced AMD, all five methods achieved an AUC of 0.81 to 0.82. For "any AMD" (Normal vs. Intermediate/Advanced), the models achieved an AUC of 0.78 to 0.79.
  • Model Calibration: Brier scores were approximately 0.17, indicating good calibration compared to the 0.33 benchmark for random guessing.
  • Genetic Drivers: LIME heatmaps confirmed that the CFH and ARMS2/HTRA1 genes were the most significant contributors to the neural network's decisions. Interestingly, the study also identified ABHD2 on chromosome 15 as a marker for differentiating intermediate from advanced AMD.
  • The Role of Age: Despite the depth of genetic data, age remained a critical predictor. Including age was found to be essential for maintaining model accuracy.
  • Ancestry Specificity: A significant performance drop was observed when models trained on Caucasian data were applied to non-Caucasian individuals (AUC decreased to ~0.72), highlighting a "genomic gap" in current predictive medicine.

Discussion and Limitations

While the models provide a high-accuracy "snapshot" of AMD risk, they do not yet predict the time to progression—the specific point at which a patient might transition from intermediate to advanced stages. Furthermore, the limited generalizability to non-Caucasian populations underscores the need for more diverse training cohorts. Future research will focus on incorporating longitudinal data to track disease progression over time and expanding datasets to ensure equitable clinical outcomes across different ancestries.

Summary updated

Abstract

ABSTRACT Numerous independent susceptibility variants have been identified for Age-related macular degeneration (AMD) by genome-wide association studies (GWAS). Since advanced AMD is currently incurable, an accurate prediction of a person’s AMD risk using genetic information is desirable for early diagnosis and clinical management. In this study, genotype data of 32,215 Caucasian individuals with age above 50 years from the International AMD Genomics Consortium in dbGAP were used to establish and validate prediction models for AMD risk using four different machine learning approaches: neural network, lasso regression, support vector machine, and random forest. A standard logistic regression model was also considered using a genetic risk score. To identify feature SNPs for AMD prediction models, we selected the genome-wide significant SNPs from GWAS. All methods achieved good performance for predicting normal controls versus advanced AMD cases (AUC=0.81∼0.82 in a separate test dataset) and normal controls versus any AMD (AUC=0.78∼0.79). By applying the state-of-art machine learning approaches on the large AMD GWAS data, the predictive models we established can provide an accurate estimation of an individual’s AMD risk profile across the person’s lifespan based on a comprehensive genetic information.
BibTeX
@misc{yan2019gwas,
  author = {Qi Yan and Yale Jiang and Heng Huang and Anand Swaroop and Emily Y. Chew and Daniel E. Weeks and Wei Chen and Ying Ding},
  title = {{GWAS-based Machine Learning for Prediction of Age-Related Macular Degeneration Risk}},
  year = {2019},
  doi = {10.1101/19006155},
  howpublished = {medRxiv},
}
Figure