Comparative modeling of SNP selection and phenotype prediction using sparse linear models (L₀-regularization) and non-linear models (Random Forests) on the QTLMAS 2010 GWAS dataset.
This project investigates and compares sparse linear and non-linear modeling strategies to identify significant SNPs associated with a quantitative trait in GWAS data. The objective is to benchmark model sparsity, generalization, and biological interpretability in high-dimensional genomic data.
- Apply and evaluate L₀-regularized regression and Random Forest models for SNP selection and phenotype prediction.
- Explore hybrid modeling workflows combining sparse feature selection with non-linear model ranking.
- Analyze the biological relevance of selected SNPs and their mapped genes.
- Dataset: QTLMAS 2010 GWAS dataset (3,226 individuals, 9,723 SNPs).
- Approaches:
- L₀-Regularization: Iterative Hard Thresholding (IHT), Greedy Forward Selection, L0Learn R package (via Python
rpy2). - Random Forest: Hyperparameter tuning, importance filtering, cumulative feature selection.
- L₀-Regularization: Iterative Hard Thresholding (IHT), Greedy Forward Selection, L0Learn R package (via Python
- Tools & Libraries: Python, R, L0Learn, Scikit-learn, Random Forest, Iterative Hard Thresholding, Pandas, Seaborn, Matplotlib.
- L₀-regularized models: Achieved >85% SNP selection stability, identified <80 SNPs, and reduced Test MSE to 0.77.
- Random Forest: Overfitted on the dataset (Test R² ≈ 0.30), revealing the additive nature of the trait with few dominant SNPs.
- Biological Findings: Key SNPs mapped to genes like PRKAB2, RHOA, and SPTBN1, linked to metabolic and neuronal pathways.
This was a collaborative project by:
- Sanyukta Chapagain (LinkedIn)
- Divyansh Rana
- Vinamrata Sharma
- Pranay Bandaru
- Virija Nandamudi
We would like to thank our mentors and peers for their guidance and support throughout this project.
This project is shared for academic and learning purposes.