Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks

Lena Podina, Ali Ghodsi, Mohammad Kohandel

This academic paper from researchers at the University of Waterloo presents a hybrid modeling approach that combines the interpretability of biological laws with the flexible pattern-recognition of deep learning. By using Universal Physics-Informed Neural Networks (UPINNs), the authors demonstrate how to "discover" the mathematical functional forms of chemotherapy drug actions directly from experimental data, potentially accelerating the development of personalized cancer treatment schedules.

TL;DR

The researchers developed a method to identify unknown biological mechanisms—specifically how chemotherapy drugs kill cancer cells—by embedding neural networks within differential equations. This "physics-informed" approach allows the model to learn complex drug-action functions from sparse, noisy data while remaining grounded in established biological principles.

What Problem Does This Paper Address?

In pharmacology, researchers use Quantitative Systems Pharmacology (QSP) models to predict how a drug will behave in the body. These models usually rely on Ordinary Differential Equations (ODEs). However, building these equations is a "bottleneck" process: it requires manual distillation of vast literature and making simplifying assumptions about the "mechanism of action" (how the drug actually kills the cells).

While pure Machine Learning (ML) can find patterns in data automatically, it is often a "black box"—it might predict an outcome but can't explain why or through what biological mechanism. Furthermore, pure ML requires massive amounts of data, which is rarely available in clinical or in-vitro settings. This paper addresses the need for a middle ground: a model that can learn unknown biological terms automatically without losing the structural "sanity check" provided by physics and biology.

Methods

The authors utilize Universal Physics-Informed Neural Networks (UPINNs). Here is how the architecture works:

  1. The Hybrid Equation: They represent the change in cell count over time as $\frac{du}{dt} = F(u, t) + G(u, t)$.
  2. $F$ is the known component (e.g., how cells naturally grow).
  3. $G$ is the unknown component (the drug action).
  4. Dual Networks: Instead of one neural network, they use two. One acts as a "surrogate solution" to predict cell counts over time ($U_{NN}$), and the other specifically models the hidden drug-action function ($G_{NN}$).
  5. Physics-Informed Loss: The model is trained using a composite loss function. It doesn't just try to match the data points (Mean Squared Error); it also penalizes the model if the derivative of its predictions violates the structure of the underlying ODE.
  6. Symbolic Potential: Once the neural network learns the shape of the drug action, that "black box" function can be passed to symbolic regression algorithms to turn the network's weights back into human-readable math (like the Log-kill or Norton-Simon models).

Datasets

The study evaluated the UPINN framework on two types of data:

  • Synthetic Data: The authors generated "ground truth" datasets using three established mathematical models of cell-kill: Log-kill, Norton-Simon, and $E_{max}$. They tested the model's ability to recover these specific functions under various conditions, including sparse data and 3% Gaussian noise.
  • In-vitro Experimental Data: They used real-world time-series cell count data from the SUM-149 breast cancer cell line treated with doxorubicin at varying concentrations (10nM to 2500nM) and exposure durations (6h, 12h, 24h).

Key Insights and Ideas

  • Function Discovery vs. Parameter Fitting: Traditional PINNs are used to find a specific number (like a growth rate). UPINNs are "Universal" because they can discover an entire function (the shape of the curve), which is much more powerful for identifying unknown biological mechanisms.
  • Data Efficiency through "Structure": By "baking in" the known parts of the biology (like logistic growth), the neural network only has to learn the "residual" drug effect. This allows the model to work with much smaller datasets than standard deep learning.
  • Dose Interpolation: A major advantage of this approach is its ability to interpolate. By training on a few specific dosages, the model can predict the drug's effect at a dosage it has never seen before, which is critical for optimizing treatment schedules.
  • Handling Noise: The study found that while noise makes the task harder, the "physics" constraint acts as a regularizer, preventing the neural network from over-fitting to random fluctuations in the data.

Key Results

  • High Accuracy: On synthetic data, the UPINN recovered the hidden drug-action terms with a Mean Squared Error (MSE) as low as $10^{-4}$ to $10^{-5}$.
  • Robustness: Even with added noise, the model successfully identified the correct type of drug action (e.g., distinguishing between a Log-kill and an $E_{max}$ model).
  • Real-World Validation: When applied to doxorubicin data, the model successfully captured a time-dependent proliferation rate. It identified a specific peak in cell-killing activity that simpler, traditional models might miss, providing a more nuanced view of how the drug interacts with the SUM-149 cell line.

Limitations and Future Work

  • Stochasticity: Because neural networks start with random weights, the results can vary slightly between training runs. The authors suggest running multiple trials to ensure stability.
  • Identifiability: In some complex biological systems, different mathematical functions might produce the same output data. The authors note that if a function isn't "identifiable" from the data provided, the model may struggle.
  • Uncertainty Quantification: The current model provides a "best guess" but doesn't yet quantify how confident it is in its discovered function. The authors suggest integrating Bayesian methods or conformal prediction in the future.
  • Closing the Loop: The next step is to fully integrate symbolic regression to automatically output the final discovered equations in a form that clinicians and biologists can use directly.

Summary updated

Abstract

Quantitative systems pharmacology (QSP) is widely used to assess drug effects and toxicity before the drug goes to clinical trial. However, significant manual distillation of the literature is needed in order to construct a QSP model. Parameters may need to be fit, and simplifying assumptions of the model need to be made. In this work, we apply Universal Physics-Informed Neural Networks (UPINNs) to learn unknown components of various differential equations that model chemotherapy pharmacodynamics. We learn three commonly employed chemotherapeutic drug actions (log-kill, Norton-Simon, and E_max) from synthetic data. Then, we use the UPINN method to fit the parameters for several synthetic datasets simultaneously. Finally, we learn the net proliferation rate in a model of doxorubicin (a chemotherapeutic) pharmacodynamics. As these are only toy examples, we highlight the usefulness of UPINNs in learning unknown terms in pharmacodynamic and pharmacokinetic models.
BibTeX
@inproceedings{podina2025learning,
  author = {Lena Podina and Ali Ghodsi and Mohammad Kohandel},
  title = {{Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks}},
  booktitle = {Pharmaceutical},
  year = {2025},
  doi = {10.1007/s11095-025-03858-8},
}

Loading related content...

Figure