Genomic Disease Prediction using Polygenic Risk Scores (PRS)

The project mainly focuses on evaluating machine learning models and predicting specific genetically inherited diseases using Polygenic Risk Scores (PRS), leveraging Genomic data and machine learning techniques. The work was carried out as a part of an internship at IIT Guwahati .

🔬 Project Overview

The goal of this project is to assess the predicitive capability of common SNPs (Single Nucleotide Polymorphism) using PRS, applied across multiple diseases:

• Rheumatoid Arthritis
• Alzheimer's Disease
• Type 2 Diabetes
• Hyperthyroidism

📁 Project Structure

Internship_Project/
├─ Final Datasets/
│  ├─ ALZ_final/
│  │  ├─ ALZ_aligned_genotypes_final.csv
│  │  ├─ ALZ_aligned_phenotypes_final_extended.csv
│  │  └─ ALZ_aligned_phenotypes_final.csv
│  ├─ HYP_final/
│  │  ├─ HYP_aligned_genotypes_final.csv
│  │  ├─ HYP_aligned_phenotypes_final_extended.csv
│  │  └─ HYP_aligned_phenotypes_final.csv
│  ├─ RA_final/
│  │  ├─ RA_aligned_genotypes_final.csv
│  │  ├─ RA_aligned_phenotypes_final_extended.csv
│  │  └─ RA_aligned_phenotypes_final.csv
│  └─ T2D_final/
│     ├─ T2D_aligned_genotypes_final.csv
│     ├─ T2D_aligned_phenotypes_final_extended.csv
│     └─ T2D_aligned_phenotypes_final.csv
├─ Models/
│  ├─ Alzheimers.ipynb
│  ├─ Hyperthyroidism.ipynb
│  ├─ Rheumatoid_Arhtiritis.ipynb
│  └─ Type_2_diabetes.ipynb
├─ Outputs/
│  ├─ ALZ_results/
│  │  ├─ ALZ_confusion_matrix.png
│  │  ├─ ALZ_PRS_distribution_fixed.png
│  │  ├─ ALZ_prs_scores.csv
│  │  ├─ ALZ_selectKbest_logreg_model_AgeSex.joblib
│  │  ├─ ALZ_selectKbest_logreg_model.joblib
│  │  ├─ ALZ_selectKbest_model_performance_AgeSex.png
│  │  ├─ ALZ_selectKbest_model_performance.png
│  │  ├─ logreg_selectKbest_coefficients_AgeSex.csv
│  │  └─ logreg_selectKbest_coefficients.csv
│  ├─ HYP_results/
│  │  ├─ HYP_confusion_matrix.png
│  │  ├─ HYP_PRS_distribution_fixed.png
│  │  ├─ HYP_prs_scores.csv
│  │  ├─ HYP_selectKbest_logreg_model_AgeSex.joblib
│  │  ├─ HYP_selectKbest_logreg_model.joblib
│  │  ├─ HYP_selectKbest_model_performance_AgeSex.png
│  │  ├─ HYP_selectKbest_model_performance.png
│  │  ├─ logreg_selectKbest_coefficients_AgeSex.csv
│  │  └─ logreg_selectKbest_coefficients.csv
│  ├─ RA_results/
│  │  ├─ logreg_selectKbest_coefficients_AgeSex.csv
│  │  ├─ logreg_selectKbest_coefficients.csv
│  │  ├─ RA_confusion_matrix.png
│  │  ├─ RA_PRS_distribution_fixed.png
│  │  ├─ RA_prs_scores.csv
│  │  ├─ RA_selectKbest_logreg_model_AgeSex.joblib
│  │  ├─ RA_selectKbest_logreg_model.joblib
│  │  ├─ RA_selectKbest_model_performance_ageSex.png
│  │  └─ RA_selectKbest_model_performance.png
│  ├─ T2D_results/
│  │  ├─ logreg_selectKbest_coefficients_AgeSex.csv
│  │  ├─ logreg_selectKbest_coefficients.csv
│  │  ├─ T2D_confusion_matrix.png
│  │  ├─ T2D_PRS_distribution_fixed.png
│  │  ├─ T2D_prs_scores.csv
│  │  ├─ T2D_selectKbest_logreg_model_AgeSex.joblib
│  │  ├─ T2D_selectKbest_logreg_model.joblib
│  │  ├─ T2D_selectKbest_model_performance_AgeSex.png
│  │  └─ T2D_selectKbest_model_performance.png
│  └─ Table.ipynb
├─ Raw_Datasets/
│  ├─ ALZ_data/
│  │  ├─ ALZ_chr_dosage_matrix.csv
│  │  ├─ ALZ_chr_genotypes_raw.txt
│  │  ├─ ALZ_chr_genotypes_with_header.tsv
│  │  ├─ ALZ_top2000_cleaned.csv
│  │  ├─ phenotype_ALZ_aligned.csv
│  │  ├─ phenotype_ALZ_simulated_50_50.tsv
│  │  └─ sample_ids(1).txt
│  ├─ HYP_data/
│  │  ├─ HYP_chr_dosage_matrix.csv
│  │  ├─ HYP_chr_genotypes_raw.txt
│  │  ├─ HYP_chr_genotypes_with_header.tsv
│  │  ├─ HyperTH_top2000_cleaned.csv
│  │  ├─ phenotype_HYP_aligned.csv
│  │  ├─ phenotype_HYP_simulated_50_50.tsv
│  │  └─ sample_ids(3).txt
│  ├─ RA_data/
│  │  ├─ phenotype_RA_aligned.csv
│  │  ├─ phenotype_RA_simulated_50_50.tsv
│  │  ├─ RA_aligned_genotypes_final.csv
│  │  ├─ RA_aligned_phenotypes_final.csv
│  │  ├─ RA_chr1_6_dosage_matrix.csv
│  │  ├─ RA_chr1_6_genotypes_raw.txt
│  │  ├─ RA_chr1_6_genotypes_with_header.tsv
│  │  ├─ RA_top10000_cleaned.csv
│  │  └─ sample_ids.txt
│  ├─ T2D_data/
│  │  ├─ phenotype_T2D_aligned.csv
│  │  ├─ phenotype_T2D_simulated_50_50.tsv
│  │  ├─ sample_ids(2).txt
│  │  ├─ T2D_chr_dosage_matrix.csv
│  │  ├─ T2D_chr_genotypes_raw.txt
│  │  ├─ T2D_chr_genotypes_with_header.tsv
│  │  └─ T2D_top2000_cleaned.csv
│  ├─ Dataset_processing_ALZ.ipynb
│  ├─ Dataset_processing_HYP.ipynb
│  ├─ Dataset_processing_RA.ipynb
│  └─ Dataset_processing_T2D.ipynb
├─ .gitignore
├─ environment.yaml
├─ LICENSE
└─ README.md

🧬 Datasets Used

• 1000 Genomes Project (genotype data)
• Phenotype labels were simulated to follow disease-specific prevalance are used for binary classificaton.
• Two splits were tested:
  • 50/50 Case-Control
  • 10/90 Case-Control (produced more realistic and better-performing results)

🧠 Methodology

• Dataset Collection:

  • Used GWAS Catalog for SNP collection for the four diseases and 1000 Genomes Project for chromosomes related to each particular trait
  • Used pandas to remove the NULL values and taking only the necessary columns (rs_id, chrom, pos, p_value, beta)

• Genotype Preprocessing:

  • Used bcftools to filter SNPs based on:
    • MAF > 0.001
    • Missing Rate < 3%
    • Hardy-Weinberg Equillibrium (HWE) > 1e-6
  • SNPs encoded as dosage values (0, 1, 2)

• Feature Selection:

  • GWAS p-value filtering (p < 1e-5)
  • SelectKBest using chi-square or ANOVA F-score

• Modelling:

  • Logistic Regression (Lasso)
  • Scoring Metrics: AUC, sensitivity, specificity
  • PRS calculated as a weighted sum:
    • PRSᵢ = Σⱼ=1ᵏ (βⱼ × dosageᵢⱼ)

• Evaluation:

  • ROC AUC range for SNP-only models (50/50 split):
    • RA: ~0.55
    • ALZ: ~0.60
    • T2D: ~ 0.53
    • HYP: ~0.53
  • ROC AUC range for SNP-only models (10/90 split):
    • RA: ~0.60
    • ALZ: ~0.64
    • T2D: ~ 0.62
    • HYP: ~0.52

⚒️ Tools and Libraries

• Python (pandas, scikit-learn, seaborn, matplotlib, numpy)
• bcftools
• Jupyter Notebook

🔗 Linked Resources

• 1000 Genomes Project
• GWAS Catalog
• bcftools

🙌 Acknowledgments

• This project was completed under the guidance of Prof. M.K. Bhuyan, Department of EEE, IIT Guwahati.
• Thanks to Pragyan Thapa for collaborating on model development and experiments.