Combining model-based and data-driven models: An application to synthetic biology resource competition.

Area of Science:

• Computational Biology
• Systems Biology
• Machine Learning in Science

Background:

• Mechanistic models (MM) offer interpretability but can be complex to build.
• Machine learning (ML) excels at data-driven modeling but may lack interpretability.
• Integrating ML and MM is desirable for robust predictions and deeper system insights.

Purpose of the Study:

• To develop a hybrid modeling framework combining ML and MM.
• To ensure ML components respect mechanistic constraints, avoiding overfitting and maintaining interpretability.
• To address challenges in modeling complex biological systems, such as synthetic genetic circuits.

Main Methods:

• Introduction of Partially Uncertain Model Structures (PUMS) to guide ML components.
• Development of embedded Physics-Informed Neural Networks (ePINNs) with shared loss functions.
• Application of the ePINNs framework to a gene network model with resource competition.

Main Results:

• Demonstrated effectiveness of ePINNs in capturing complex system interactions.
• Showcased the ability of the hybrid approach to maintain physical consistency.
• Validated the framework's utility in synthetic biology applications.

Conclusions:

• The ePINNs framework offers a powerful method for integrating ML and MM.
• This hybrid approach improves model robustness, interpretability, and predictive accuracy.
• ePINNs are particularly valuable for modeling systems with inherent mechanistic principles, like engineered biological systems.