Understanding glioblastoma invasion using physically-guided neural networks with internal variables

Area of Science:

• Computational Biology
• Oncology
• Bioengineering

Background:

• Glioblastoma is a deadly brain tumor with complex invasion dynamics.
• Microfluidic devices enable controlled cell culture for studying tumor evolution.
• Data Science and Machine Learning offer new tools for simulating and understanding cancer progression.

Purpose of the Study:

• To apply Physically-Guided Neural Networks with Internal Variables (PGNNIV) to study Glioblastoma invasion.
• To integrate microfluidic data with physical principles for tumor modeling.
• To develop a more predictive and explanatory model for Glioblastoma evolution.

Main Methods:

• Utilized microfluidic devices for cell culture and monitoring.
• Developed a PGNNIV model incorporating a nonlinear advection-diffusion-reaction partial differential equation.
• Employed multilayer perceptrons and nodal deconvolution to learn Glioblastoma's metabolic behavior.
• Trained the PGNNIV model using synthetic data from in silico tests under varying oxygenation conditions.

Main Results:

• PGNNIV successfully integrated microfluidic data with physical laws to model Glioblastoma invasion.
• The model demonstrated the capacity to discover complex metabolic processes non-parametrically.
• PGNNIV surpassed the predictive power of traditional parametric approaches.
• The network provided explanatory capacity, moving beyond mere prediction.

Conclusions:

• PGNNIV offers a powerful, data-driven, and physics-informed approach for studying Glioblastoma.
• This method enhances understanding of tumor evolution and invasion mechanisms.
• The PGNNIV framework facilitates the development of virtual therapies and personalized medicine strategies for Glioblastoma.