2024 AIChE Annual Meeting

(595f) Towards Cancer Digital Twins: Simulating Tumor Heterogeneity and Patient Response Using Tissue Level Agent-Based Modeling of Prostate Cancer

Authors

Tao, M., University of Pennsylvania
Radhakrishnan, R., University of Pennsylvania
Cancer treatment involves various methods such as chemotherapy, radiotherapy, surgery, and targeted therapies like immunotherapy. By gathering clinical data at different scales, such as multi-omics, quantitative single-cell sequencing, and tissue-level tumor histology data, we can personalize treatments for individual patients. A robust computational framework integrating multi-omics data across scales can significantly improve patient-specific treatment selection. We hypothesize that AI-enabled agent-based modeling (ABM) can play a crucial role in leveraging patient multi-omics data and integrating cancer systems models that encode cellular signaling with digital twin models that predict the behavior of the cellular response in a customized tumor microenvironment. ABMs involve autonomous agents interacting in a shared environment, producing collective emergent outcomes. They can capture heterogeneous tumor dynamics observed across patients and even within the same tumor. ABMs flexibly incorporate multiple physical processes across scales, such as oxygen and nutrient diffusion, and seamlessly combine deterministic and data-driven stochastic processes. In
ABMs, we can integrate random mutation events into evolving tumor simulations and represent the effect of intra-tumor heterogeneity explicitly.

We developed an agent-based model to simulate the growth of a diverse population of prostate cancer cells after radical prostatectomy. The model integrates various cellular functions such as androgen receptor signaling, mitogenic signaling, and DNA damage response under androgen deprivation therapy in sensitive and resistant conditions. We customized the model to patient cohorts corresponding to non-recurrent, biochemically determined recurrent, and image-based determined recurrent tumors. The model is coupled to the ABM using cell proliferation, migration, and adhesion propensities local to signals in the microenvironment, such as mitogenic factors, drug concentrations, and cellular state. We conducted parameter scans on several parameters to gain further insight into the behavior of heterogeneous tumors and their sensitivity to changes in model parameters.

Testosterone can promote the growth of prostate cancer cells. When the cells are deprived of androgen hormones, they can develop the ability to take up or better utilize available testosterone to fuel their growth. Our study analyzed the impact of cellular testosterone uptake rate parameters on the cells. We conducted a local sensitivity analysis on this parameter. We found that the tumor recurrent (TR) cohort exhibited higher variation in the ratio of PTEN-resistant to PTEN-sensitive cells at lower uptake rates. In contrast, the biochemically recurrent (BR) cohort showed a higher mean PTEN Resistant/Sensitive ratio at higher uptake rates. This trend suggests that the tumor progression is biased towards the aggressive cell type at higher testosterone uptake rates in prostate cancer cells.

Extensive local, parametric, and global sensitivity analysis on resistant tumor cell adhesion-motility revealed clustering and micrometastatic formations in heterogeneous cell populations above distinct thresholds mimicking a sigmoid-like behavior. The TR cohort showed the highest degree of clustering (indicated by the highest plateau of the sigmoid fit). We conducted this adhesion-motility sensitivity analysis at different cellular testosterone uptake rates and observed that the degree of clustering increases with increasing testosterone uptake rates for the TR cohort.

We also developed new coupling strategies between cellular systems models and ABM enabled via ML-based algorithms to better account for intra-tumor heterogeneity and reduce the computational costs of the ABM by replacing the expensive mechanistic ODE-based models with ML-based cell fate prediction models. Specifically, we incorporate heterogeneity using machine learning to inform cancer cell growth rates based on local EGF, testosterone, phosphorylated Erk, and Akt concentrations. Conventionally, global sensitivity analysis is conducted on mechanistic models to check for robustness and identify stiff or sloppy parameters. These methods are complex to implement for hybrid models, including ODE, Boolean, and ABM constructs, with dynamic exchange of information at different time scales. In ongoing work, we are formulating methods for conducting Sobol sensitivity, robustness, and uncertainty quantification analysis on complex hybrid models.

As a more straightforward check for global sensitivity, we conducted a Shapley value analysis with the developed ML-based cell fate prediction model. We use the Shapley values to rank features in the model that impacted the observed model outputs most. We identified species/genes consistently appearing in the top features across Shapley rankings for three different ML prediction models - Neural Networks, Random Forest, and Support Vector Machines. We then plotted Kaplan Meir curves to compare the overall survival of prostate cancer patient cohorts with and without alterations in the identified species/genes. Alterations in most of these genes showed worse survival in patients, providing clinical significance to our findings that prostate cancer progression is sensitive to changes in the expression of the species identified from our multiscale hybrid tissue level modeling protocol.

Our results demonstrate the potential of multiscale ABMs and coupled data-based methods in improving our understanding of complex tumor behaviors. Such frameworks can be further extended to capture the heterogeneous responses to therapies and the development of resistance against them. The long-term goal would be building robust models informing patient-specific precision tumor therapies and enhancing decision-making.

Funding Acknowledgement: We acknowledge funding from the National Cancer Institute through the Physical Sciences Oncology Network.