yzhu13/poag-pgs-ml

Polygenic risk score (PGS) and machine learning framework for primary open-angle glaucoma prediction in African ancestry cohorts. # POAG PGS + ML — Analysis Code **Multimodal Prediction of Primary Open-Angle Glaucoma Using Polygenic Risk Scores and Clinical Features in a High-Risk African Ancestry Cohort** Yan Zhu, Aude Benigne Ikuzwe Sindikubwabo, Yuki Bradford, et al. *iScience*, 2026 — Manuscript ISCIENCE-D-26-03991R2 --- ## Overview Python analysis pipeline for POAG risk prediction using: - Four polygenic risk scores (PGS) matched to African ancestry - Four ML classifiers (LR, RF, SVM, MLP) - **5-fold × 20-repeat stratified cross-validation** (5×20 CV; 100 fits per configuration) - External validation in the Penn Medicine BioBank (PMBB; N = 9,817 AFR) --- ## Repository structure ``` . ├── 01_pgs_standalone.py Figure 2: standalone PGS performance (5×20 CV + PMBB) ├── 02_training_external_validation.py Figure 3: 12-feature-set training + PMBB external validation ├── 03_suspect_enrichment.py Figure 4: clinical enrichment in 1,013 suspects ├── 04_asymmetry_analysis.py Figure 5: inter-eye asymmetry (ΔIOP, ΔCDR) + PMBB ├── 05_shap_calibration.py Fig S3: SHAP feature importance + calibration curves ├── 06_learning_curves_sex_stratified.py Fig S4–S5: learning curves + sex-stratified AUC ├── 07_delta_auc_paired.py Fig S7, Tables S12: paired ΔAUC (DeLong + bootstrap + CV-fold) ├── 08_same_classifier_comparison.py Table S11: within-classifier Base vs Base+PGS AUC ├── 09_pgs_residualized_on_pc.py Table S13: PGS residualized on ancestry PCs (sensitivity) ├── requirements.txt └── data/ ├── README.md Data access instructions ├── poaagg/ Place POAAGG cohort files here └── pmbb/ Place PMBB external validation files here ``` All Excel output tables are written to `outputs/tables/` All figures (PNG + PDF) are written to `outputs/figures/` --- ## Setup ```bash pip install -r requirements.txt ``` Python 3.11 recommended. --- ## Data Raw data are not publicly available (IR …