Learning Model Parameter Dynamics in a Combination Therapy for Bladder Cancer from Sparse Biological Data

AI summary

Overview

Research area: Computational oncology and scientific machine learning — specifically physics-informed neural networks (PINNs) applied to ordinary differential equation (ODE) models of tumor–immune interactions under combination therapy.

Technical level: Intermediate. The paper assumes familiarity with ODE models, neural network training, and the PINN concept, but it is written accessibly for a machine-learning/health time-series audience (it was presented at the workshop "Learning from Time Series for Health").

Scope: The paper develops a PINN framework that learns both hidden subpopulation trajectories and a time-varying interaction parameter in a bladder cancer combination-therapy ODE model, using only six sparse ultrasound tumor-volume measurements plus flow cytometry and digital histology constraints.

What This Paper Is About

ODE models of tumor–immune interactions normally assume that model parameters are fixed, but chemotherapy and immunotherapy change how cells interact over time, so fixed-parameter models cannot capture the evolving dynamics. The problem is worse in oncology because data are sparse — often just a few tumor-volume time points, collected at irregular intervals, with measurement noise, and with only some interacting cell types observable. The goal is to learn the trajectories of the unobserved subpopulations and the time-varying interaction parameter at time points where no data exist, using a PINN that is regularized by biological knowledge.

Key Contributions

1. A PINN framework for joint inference of trajectories and time-varying parameters. The method learns neural surrogates for both the state variables (subpopulations) and the time-varying parameter simultaneously, so hidden subpopulation trajectories can be predicted at time points where no measurements exist.
2. A data-augmentation strategy for sparse measurements. Spline-based interpolation extends the few observed total tumor-volume points into additional interior points, which are fed to the network as the input time domain.
3. Incorporation of biological knowledge as regularization. Flow cytometry data (initial conditions at day 6) and digital histology data (day 17 and day 23) are encoded as loss terms, restricting the space of admissible solutions.
4. Uncertainty bounds from repeated training. The PINN is run 10 times to produce an uncertainty bound around each learned solution.

Main Findings

• Learned trajectories respect the biological constraints. The PINN learns dynamics for the subpopulations cancer cells (C), OT-1 T cells (T), and MDSCs (M) that align with the flow cytometry point (shown as a black cross), the early histology time point (black circle dot), and the late histology time point (black plus).
• Treatment effects appear at the expected times. At the GEM injection (marked by a black vertical dashed line), the learned dynamics show local depletion of MDSCs and some killing of tumor cells; at the OT-1 injection (purple vertical dashed line), the learned dynamics show some killing of tumor cells.
• The time-varying parameter is learned without supervision. The MDSC suppression of T cells, s_MT, is inferred as a time-varying quantity in an unsupervised way, serving as a surrogate for the interaction changes caused by the external interventions.
• The learned subpopulation sum matches the measurements. The sum C + T + M matches the experimental tumor-volume data at the time points where measurements were collected (days 6, 9, 13, 16, 20, and 23).
• Drug clearance is captured. The learned clearance of gemcitabine (G) is shown alongside the framework schematic.
• Uncertainty is estimated by repeated runs. Ten independent PINN runs produce an uncertainty bound for each solution.
• Data sparsity is explicit. The experimental data, denoted u_GEM-OT1, consist of tumor volumes from ultrasound images recorded on days 6, 9, 13, 16, 20, and 23 — six measurement days in total. The paper does not report absolute tumor-volume values, and it does not report a numerical error metric (such as RMSE) for the fit.

Methodology in Plain English

The researchers start from a published mouse experiment: 1×10⁵ MB49-OVA bladder cancer cells were instilled orthotopically into the mouse bladder; on day 10, 500 μg of gemcitabine (GEM) was injected intravesically to deplete immunosuppressive cells such as myeloid-derived suppressor cells (MDSCs); and on day 14, 5×10⁶ OT-1 T cells — engineered to recognize the MB49-OVA tumor and a pre-clinical analogue of adoptive cell therapy with tumor-infiltrating lymphocytes (ACT-TIL) — were administered intravesically. Tumor volume was monitored by ultrasound on days 6, 9, 13, 16, 20, and 23.

They write a system of four ODEs for cancer cells (C), OT-1 T cells (T), MDSCs (M), and GEM (G), in which cancer cells grow and are killed by T cells and GEM, T cells are injected and grow but are suppressed by cancer cells, MDSCs, and GEM, MDSCs are recruited by cancer cells and killed by GEM, and GEM is injected and cleared. They single out the MDSC suppression of T cells, s_MT, as the parameter that varies over time because of the interventions, while the remaining parameters are held constant.

Because only total tumor volume is measured (not each subpopulation), they assume the measured volume is the sum C + T + M. They use spline-based interpolation to create additional points between the six measurement days, then train a feedforward neural network surrogate u_NN for the state trajectories and a second surrogate Λ_NN for the time-varying parameter. Both take the interpolated time domain as input.

Training details: SiLU activations in the hidden layers and Softplus on the output layers to enforce non-negativity; normalization of the interpolated data before training; the Adam optimizer for 20,000 epochs; u_NN with three hidden layers of 100 neurons each; Λ_NN with two hidden layers of 200 neurons each; implemented in PyTorch.

The loss is a weighted sum of four terms: an ODE residual loss that enforces the governing equations (with derivatives taken by automatic differentiation), a data loss that matches the observed total volume, an initial-condition loss, and a biological-constraint loss. The weights are adaptively adjusted using multi-task learning. The biological constraints come from the data: at day 6 (t₀ = 6), flow cytometry gives proportions of q₀ = 0.99887 for C, q₁ = 0 for T, and q₂ = 1 − 0.99887 for M; at day 17 (histology), the proportions are q₀₁ = 0.95755 for C, q₁₁ = 0.01818 for T, and q₂₁ = 0.02427 for M; and at day 23 (histology), q₀₂ = 0.95665 for C, q₁₂ = 0.00078 for T, and q₂₂ = 0.04256 for M. In the digital histology, tumors (C) are outlined in black, MDSCs (M) are green dots, and T cells (T) are blue dots.

Why This Matters

Impact on research. Most differential-equation models in oncology assume constant parameters, which is a poor fit when treatments continually reshape the tumor microenvironment. This work shows that time-varying parameters can be recovered from very few measurements while still producing biologically plausible hidden trajectories, and it demonstrates a general recipe for injecting heterogeneous biological data (imaging, flow cytometry, histology) as soft constraints in a machine-learning loss. The authors note that the approach applies to many other differential-equation models in biology where interacting organisms change their behavior in response to external interventions.

Real-world applications.

• Designing and scheduling combination regimens — for example, deciding the timing and dose of a microenvironment-depleting chemotherapy followed by an immune-cell therapy.
• Supporting adoptive cell therapy analogous to ACT-TIL, where a patient's own T cells are expanded ex vivo and reinfused.
• Building patient-specific "digital twin" models that reconstruct unmeasured cell populations from routine imaging volumes.
• Making better use of sparse clinical and preclinical datasets, where only a handful of imaging time points are available per patient or animal.

Industry relevance. The method targets pharmaceutical and immuno-oncology development, where combination-therapy decisions are expensive and preclinical datasets are small. It offers a way to extract more signal from already-collected imaging plus histology/flow cytometry data, and it is relevant to organizations building simulation-based trial design, digital-twin platforms, or model-informed drug development for bladder-sparing and other immunotherapies.

Future Directions

1. Extend time-varying treatment beyond a single parameter. Only s_MT is modeled as time-varying here; the authors explicitly motivate time-varying parameters as surrogates for unobserved or unmodeled effects, so learning several simultaneously varying parameters is a natural next step.
2. Quantify reliability more rigorously. The paper reports uncertainty via 10 repeated PINN runs but reports no numerical accuracy metric; establishing error metrics and calibration against held-out data would strengthen confidence in the predicted hidden trajectories.
3. Test generality across models and datasets. The framework is claimed to be applicable to other biological differential-equation systems and other external interventions, but demonstration here is limited to one bladder cancer ODE model and one mouse experiment.
4. Translate toward clinically available data. Moving from the mouse MB49-OVA/GEM/OT-1 setting and flow cytometry/histology inputs toward human bladder-sparing treatment data would test whether the same constraints and observation model hold in patients.

Target Audience

Researchers and practitioners at the intersection of scientific machine learning and oncology: PINN and physics-informed modeling developers, mathematical oncologists building tumor–immune ODE models, computational biologists working with sparse time-series biological data, and clinical/translational scientists interested in combination chemo-immunotherapy for bladder cancer. The paper is most useful to readers who already understand ODEs and neural network training, since it assumes the PINN formulation as background and focuses on its biological application.

Authors’ abstract

In a mathematical model of interacting biological organisms, where external interventions may alter behavior over time, traditional models that assume fixed parameters usually do not capture the evolving dynamics. In oncology, this is further exacerbated by the fact that experimental data are often sparse and sometimes are composed of a few time points of tumor volume. In this paper, we propose to learn time-varying interactions between cells, such as those of bladder cancer tumors and immune cells, and their response to a combination of anticancer treatments in a limited data scenario. We employ the physics-informed neural network (PINN) approach to predict possible subpopulation trajectories at time points where no observed data are available. We demonstrate that our approach is consistent with the biological explanation of subpopulation trajectories. Our method provides a framework for learning evolving interactions among biological organisms when external interventions are applied to their environment.