Ponentes
Descripción
CAR T cell therapy is a revolutionary cancer immunotherapy. Yet tumor immune dynamics demand quantitative approaches. Mechanistic models like CARRGO require accurate parameter estimation from experimental data. Machine learning offers a powerful alternative for direct curve fitting and prediction without explicit model inversion. In this work, we implement and compare multiple machine learning regression models to fit dynamic curves derived from CAR T cell experimental data (from the literature). In vitro time course assays (487 samples) were preprocessed and reduced to time (independent variable) and the measured response. Seven regression models are trained using an 80/20 split: Linear Regression (baseline), Random Forest (100 trees), SVM (RBF kernel), AdaBoost (100 estimators), K Nearest Neighbors (k=5), Decision Tree, and a Deep Neural Network (3 hidden layers: 64-128-64, ReLU, Dropout 20%, 500 epochs). Models are compared using R2, RMSE, and 5-fold cross-validation. In addition, ML model predictors are implemented.
Comparative results show that KNN achieves the best performance, with an R2 of 0.9997 and an RMSE of 0.0080 on the test set (CV R2 = 0.9987 ± 0.0005). The deep neural network achieves R2 = 0.9938, while the Decision Tree and Random Forest achieve R2 = 0.9992 and 0.9989, respectively. The Linear Regression model yields poor performance (R2 = 0.5363). Finally, the KNN-based predictor successfully estimated response values for new time inputs (e.g., at t = 10.0 hours, predicted response = 0.8814 ± 0.0080).
This work shows that simple non-parametric models (such as aKNN, Random Forest, and Decision Tree) outperform classical linear regression and even a deep neural network for curve fitting of CAR T cell kinetic data from in vitro assays. Future work will include the integration with physics-informed models.