Physics-Informed Neural Networks for Modeling Postprandial Plasma Amino Acids Kinetics in Pigs.

Area of Science:

• Animal Physiology
• Computational Biology
• Machine Learning

Background:

• Postprandial plasma amino acid (AA) kinetics are crucial for assessing swine digestive efficiency and metabolic health.
• Conventional kinetic analysis using Non-Linear Least Squares (NLS) requires frequent blood sampling and accurate initial parameter estimates.

Purpose of the Study:

• To develop and evaluate a Physics-Informed Neural Network (PINN) framework for modeling swine AA kinetics.
• To compare the performance of PINN against traditional NLS methods, particularly under data-limited conditions.

Main Methods:

• Developed a PINN framework integrating mechanistic Ordinary Differential Equations (ODEs) into the deep learning loss function.
• Retrospectively analyzed benchmark data, simulating dense and sparse blood sampling strategies.
• Compared PINN performance with NLS in terms of robustness, accuracy, and parameter identification stability.

Main Results:

• PINN demonstrated high fidelity under dense sampling, comparable to NLS.
• Under sparse sampling, PINN showed superior robustness and predictive accuracy, significantly reducing Root Mean Square Error (RMSE) for Methionine and Lysine.
• PINN exhibited enhanced parameter identification stability and predictive consistency compared to NLS, overcoming sensitivity to initial guesses.

Conclusions:

• The PINN framework offers a reliable and consistent alternative for modeling AA dynamics in pigs.
• PINN's ability to integrate physical laws as regularization enables robust inverse problem-solving, even with limited data.
• Future applications may involve reconstructing accurate physiological trajectories using optimized sparse sampling with PINNs.