This repository contains my practice implementation of the machine learning pipeline described in Kullmann et al. (2025). The goal of this project was to gain hands-on experience with applying machine learning methods to genomic-style data.

The pipeline I practiced includes three main steps:

1. Feature Selection – Elastic Net regression to reduce thousands of features down to the most informative ones.
2. Classification – Random Forest classifier trained on the selected features.
3. Evaluation – Model performance measured using:
   • Accuracy
   • Sensitivity (Recall)
   • Specificity
   • Precision
   • Confusion Matrix
   • ROC Curve & AUC

• I used simulated methylation-style data (~200 samples × 5,000 features).
• A subset of features contained a true signal, while the rest were noise.
• This setup helped me test how well the pipeline can detect relevant features under both easy (strong signal) and realistic (weaker signal + noise) conditions.

• Features selected: ~80
• Accuracy: ~0.62
• AUC: ~0.63
• Sensitivity: ~0.61
• Specificity: ~0.63
• Precision: ~0.67

Confusion Matrix (example):

Predicted / Actual    Control   Case
Control                 17       10
Case                    13       20

• ROC Curve → shows AUC ≈ 0.63 (not perfect, but better than random).
• Feature Importance Plot → highlights the most informative features (mostly from the “true signal” set).

• Perfect model results often mean unrealistic or overly clean data.
• Once noise was added, the model’s performance dropped but became more biologically realistic.
• This project helped me understand not only how to set up Elastic Net + Random Forest, but also why data realism is critical in genomics ML.

• Apply the pipeline to real methylation data instead of simulated data.
• Compare different classifiers (e.g., SVM, XGBoost) to see if performance improves.
• Experiment with larger sample sizes and cross-validation for more robust evaluation.
• Explore interpretability tools (e.g., SHAP values) to better understand model decisions.

Clone the repo and run the pipeline in R or Python (depending on your setup).
(Note: This is a practice project; code may not be fully optimized.)