Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Quantitative systems pharmacology (QSP) models are crucial for predicting drug efficacy and toxicity.
• Current QSP model development requires extensive manual literature review and parameter fitting.
• Unknown biological mechanisms and parameters pose challenges in accurate QSP modeling.

Purpose of the Study:

• To apply Universal Physics-Informed Neural Networks (UPINNs) for learning unknown components in differential equations modeling chemotherapy pharmacodynamics.
• To automate the identification of parameters and model terms in pharmacodynamic and pharmacokinetic models.

Main Methods:

• Utilized UPINNs to learn three common chemotherapeutic drug actions (log-kill, Norton-Simon, and $$) from synthetic data.
• Employed UPINN for simultaneous fitting of parameters across multiple synthetic datasets.
• Applied UPINN to determine the net proliferation rate in a doxorubicin pharmacodynamics model.

Main Results:

• Demonstrated UPINN's capability to successfully learn hidden terms and unknown parameters in diverse differential equations.
• Showcased UPINN's effectiveness across models with varying time and variable scales for chemotherapeutic effects.
• Validated UPINN's performance in learning complex biological parameters.

Conclusions:

• UPINNs offer a powerful tool for discovering unknown terms in pharmacodynamic and pharmacokinetic models.
• This methodology can aid in understanding novel chemotherapeutics and their underlying biological interactions.
• UPINNs facilitate the analysis of learned terms, advancing the comprehension of drug mechanisms.