Adversarial Removal of Population Bias in Genomics Phenotype Prediction
We propose a general-purpose adversarial framework that learns phenotype-relevant features while explicitly removing population-related information from the latent representation.
🔑 Key Contributions
- Introduces an adversarial debiasing framework for genomic phenotype prediction
- Ensures latent genomic representation is population-agnostic
📄 Paper
Adversarial Removal of Population Bias in Genomics Phenotype Prediction 2022 IEEE International Conference on Data Mining Workshops (ICDMW)
👉 View on IEEE Xplore
💻 Method
The framework consists of:
- Encoder: learns latent SNP representation
- Regressor: predicts phenotype (main task)
- Adversarial classifier: predicts population label
- Gradient Reversal Layer (GRL): forces encoder to remove population information
The encoder is trained to:
- Minimize phenotype prediction loss
- Maximize adversarial classification error
This ensures the learned representation contains phenotype information but not population bias.
Performance Comparison
| Method | MAE ↓ | MSE ↓ | PCC ↑ |
|---|---|---|---|
| Random Forest | 0.0719 | 0.0084 | 72.00% |
| Dual-CNN | 0.0742 | 0.0090 | 72.47% |
| RF + Linear Regression | 0.0706 | 0.0081 | 72.64% |
| Ours (Adversarial + MLP) | 0.0689 | 0.0077 | 74.17% |
Our adversarial framework improves:
- MAE by 0.0036
- MSE by 0.0009
- PCC by +2.64%
🌍 Generalization
Population stratification inflates associations and reduces generalization in:
- Human disease risk prediction
- Crop yield modeling
- Livestock genomic selection
- Epidemiological modeling