Exploration of hepatitis B virus infection dynamics through physics-informed deep learning approach

Area of Science:

• Computational biology
• Virology
• Machine learning

Background:

• Accurate viral disease forecasting is vital for public health.
• Physics-Informed Neural Networks (PINNs) show promise but struggle with parameter estimation in inverse problems.
• Disease-Informed Neural Networks (DINNs) offer a robust framework for parameter estimation.

Purpose of the Study:

• To apply DINNs to a hepatitis B virus (HBV) infection model for predicting transmission.
• To assess the impact of various factors on DINN performance.
• To compare DINN parameter estimation with traditional methods.

Main Methods:

• Structural identifiability analysis of the HBV model.
• Application of DINNs to estimate parameters from experimental data.
• Investigation of parameter ranges, noise, sample size, network architecture, and learning rate effects.
• Global sensitivity analysis.
• Comparison with nonlinear least-squares fitting and maximum likelihood estimation.

Main Results:

• DINNs successfully estimated unknown parameters and captured HBV infection dynamics.
• The method predicted disease progression even with incomplete compartment data.
• DINNs identified significant parameters varying across chimpanzees, consistent with sensitivity analysis.
• DINNs demonstrated superior performance compared to nonlinear least-squares fitting and maximum likelihood estimation.

Conclusions:

• DINNs provide a powerful tool for parameter estimation in viral infection models.
• The approach is effective even with sparse or noisy experimental data.
• DINNs enhance the prediction of viral disease dynamics and progression.