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

Abstract

Determining accurate kinetic models for new biochemical systems is time-intensive, requiring experimental data collection, model construction, validation, and discrimination. Traditional black-box machine learning-based transfer learning methods leverage prior knowledge but lack interpretability and physical insights. To address this, we propose a novel model structural transfer learning approach that combines symbolic regression with artificial neural network feature attribution. The method enables automatic structural modification of an inaccurate or low-fidelity mechanistic model developed for one system when being applied to another system. Through a comprehensive in silico case study, our framework successfully adapted a kinetic model from one biochemical system to a different but related one, improving predictive accuracy. Moreover, the framework can significantly accelerate model identification when being integrated with model-based design of experiments. By comparing the old and new model structures, physical insight can be obtained, altogether highlighting the framework's potential for advancing automated knowledge discovery and facilitating high-fidelity predictive digital twin design for novel biochemical processes.