Modeling Adoptive Cell Therapy in Bladder Cancer from Sparse Biological Data using PINNs

Area of Science:

• Computational oncology
• Biomedical modeling
• Machine learning in medicine

Background:

• Physics-informed neural networks (PINNs) integrate differential equations into neural network loss functions.
• Oncology research often faces sparse experimental data, particularly for tumor volume over time.
• Understanding time-varying interactions in the tumor microenvironment is crucial for combination therapy efficacy.

Purpose of the Study:

• To adapt and apply a Physics-informed neural network (PINN) framework to oncology.
• To learn time-varying interactions within a tumor microenvironment during combination therapy.
• To enhance PINN capabilities for handling sparse biological data using prior information and constraints.

Main Methods:

• Extended PINN framework by incorporating inductive biases from dynamical systems.
• Utilized observed biological constraints as regularization agents within the modified PINN algorithm.
• Applied the algorithm to an ordinary differential equation (ODE) model simulating intermittent combination therapy.

Main Results:

• The modified PINN algorithm successfully learned the dynamics of intermittent combination therapy.
• The approach yielded accurate solutions to the ODE model.
• Time-varying forms of ODE model parameters were successfully identified.
• Strong convergence was demonstrated using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).

Conclusions:

• The adapted PINN framework demonstrates robust performance in learning complex biological dynamics from sparse data.
• Incorporating prior knowledge and biological constraints significantly improves PINN generalization and solution accuracy in oncology.
• This method offers a powerful tool for modeling cancer treatment responses and discovering underlying mechanisms.