Deep learning approach to parameter optimization for physiological models.

Area of Science:

• Computational Biology
• Systems Biology
• Biophysics

Background:

• Biological data modeling faces challenges in inferring nonlinear dynamics and parameters due to difficulties in constraining conventional optimization methods to physiological ranges.
• Existing methodologies often rely on hypothetical mechanisms, complicating parameter inference.

Purpose of the Study:

• To propose and evaluate a novel method using neural networks for biological modeling, parametrization, and parameter inference.
• To address limitations in conventional parameter optimization for biological systems.

Main Methods:

• Developed a deep learning framework employing convolutional neural networks (CNNs) for parameter inference.
• Utilized simulated data from physiological lipolysis models (glucose, insulin, free fatty acids) to train the CNN.
• Trained the CNN to predict model parameters from time-course data of glucose, insulin, and free fatty acids (FFA).

Main Results:

• The CNN achieved accurate parameter inference and trajectory reconstruction on testing datasets and optimized model-fitting curves.
• The methodology demonstrated high R-squared values and low p-values across various feature engineering strategies and training dataset sizes.
• Assessed the impact of feature engineering and training data size, showing improved accuracy with appropriate feature transformations and activation functions.

Conclusions:

• Established a robust deep learning framework for parameter inference in mathematical models of physiological systems.
• The proposed method offers a powerful approach to overcome challenges in biological data modeling and parameter estimation.
• The framework is adaptable to diverse physiological systems, promising broader applications in computational biology.