Interpretable-AI-Based Model Structural Transfer Learning to Accelerate Bioprocess Model Construction.

Area of Science:

• Biochemical Engineering
• Systems Biology
• Computational Chemistry

Background:

• Developing accurate kinetic models for biochemical systems is complex and time-consuming.
• Existing transfer learning methods lack interpretability and physical insights.

Purpose of the Study:

• To develop a novel model structural transfer learning approach for adapting kinetic models.
• To improve the accuracy and interpretability of kinetic models for new biochemical systems.

Main Methods:

• Combining symbolic regression with artificial neural network feature attribution.
• Automatic structural modification of mechanistic models using transfer learning.
• In silico case study for model adaptation between biochemical systems.

Main Results:

• Successfully adapted a kinetic model from one system to a related one, enhancing predictive accuracy.
• Framework accelerates model identification when integrated with model-based design of experiments.
• Comparison of model structures provides valuable physical insights.

Conclusions:

• The proposed framework facilitates automated knowledge discovery in biochemical systems.
• Enables high-fidelity predictive digital twin design for novel biochemical processes.
• Offers a more interpretable and efficient alternative to traditional black-box methods.