Predictive Modeling of CFTR Modulator Efficiency: A Machine Learning Approach

Abstract

Cystic Fibrosis (CF) is a severe inherited disease caused by mutations in the CFTR gene, resulting in thick mucus buildup that damages the lungs and pancreas. Chronic infections and progressive lung disease reduce respiratory function, while patient responses to CFTR modulator therapies vary considerably. This variability makes it difficult for clinicians to determine which treatment will provide the greatest benefit for each patient. This project investigated whether machine learning models could predict changes in lung function, measured by forced expiratory volume in one second (FEV1%), after starting CFTR modulator therapy.

A comprehensive dataset was created by combining real CFTR genetic mutation data with carefully simulated clinical information, including age, sex, baseline lung function, comorbidities, treatment type, and therapy response. Additional genetic features, including mutation functional classes and genotype groups, were engineered to better represent the biological complexity of CF. Exploratory data analysis identified important trends and relationships before model development.

Supervised machine learning models, including Random Forest and Gradient Boosting Regressors, were trained to predict FEV1% improvement. Performance was optimized using hyperparameter tuning and cross-validation to improve reliability and generalizability. SHAP (SHapley Additive exPlanations) analysis was applied to interpret model predictions and identify the patient characteristics most strongly influencing treatment outcomes.

The Random Forest Regressor achieved the best performance, with an R² score of 0.4801 on previously unseen test data, explaining nearly half of the variation in FEV1% improvement. Feature importance analysis showed that treatment group, baseline lung function, and genotype were the strongest predictors of therapeutic response. Simulated patients receiving Trikafta consistently demonstrated greater predicted improvements than those receiving other modulators. These findings demonstrate that machine learning can predict individual responses to CFTR modulator therapy and highlight its potential to support personalized treatment decisions and advance precision medicine for patients with cystic fibrosis.