Poster T5 - Towards Cancer Digital Twins: Simulating Tumor Heterogeneity and Patient Response using Tissue level Agent-Based Modeling of Prostate Cancer

Introduction: Tumors exhibit substantial heterogeneity, manifesting as distinct morphological and molecular variations, which frequently undermines the efficacy of conventional oncological treatments.[1,2] This intra-tumoral diversity necessitates the development of precision oncology therapies. To address this imperative, our research initiative proposes establishing a computational framework to facilitate personalized therapeutic strategies that are finely tuned to the unique characteristics of each tumor. Central to our methodology is the implementation of AI-enabled agent-based modeling. We hypothesize that AI-enabled agent-based modeling 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.

Agent-based modeling (ABM) operates on a fundamental principle where each agent (i.e., a cell in a tumor) is programmed to follow a set of rules defined by the user. These rules dictate how agents interact with each other and respond to their environment, capturing the complex, dynamic behaviors observed in biological systems. [3] Agent-based modeling frameworks are useful for biological systems because they can incorporate specific phenotypes of cells (agents) based on specific intra-cellular molecular pathways, tumor-host interactions, and distribution of soluble substrates (e.g., growth factors, nutrients, drugs). ABMs can meticulously capture cell-cell and cell-micro environment communications. ABMs also offer an optimal framework for combining physics-driven, data-driven, and rules-driven processes.

References have been added in acknowledgements.

Materials and

Methods: We use PhysiCell [4] to develop an agent-based model simulating the growth of prostate cancer tissue after radical prostatectomy. Our methodological workflow is pictorially summarized in Figure 1.

We have established a cellular systems biology model that predicts cell states, which are averaged across runs to calculate net cell growth (NCG) probability. This 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 [5].

This cellular systems model is coupled to the tissue level ABM through key cell phenotypic variables such as cell proliferation, migration, and adhesion propensities, which are determined from the systems model as well as diffusible signals local to the microenvironment, such as mitogenic factors, drug concentrations. We conducted scans on key parameters to quantify their effect on intra-tumor heterogeneity and tumor amplification.

We also use machine learning approaches to reduce the computational expense of ABMs [6] as well as to conduct feature ranking and global sensitivity analysis leveraging new data-driven methodologies. We used the cellular systems biology model to construct an ML-based cell fate prediction model to reduce the computational expense of the ABM. As a straightforward check for the global sensitivity of the effect of parametric variation on key phenotypic properties, we conduct a Shapley value analysis [7] with this ML-based cell fate prediction model and identify top sensitive parameters.

References have been added in acknowledgements.

Results, Conclusions, and Discussions: The cellular systems model was customized to two types of cells - resistant (with PTEN gene deletion) and sensitive. The model was simulated under varying testosterone and EGF conditions, revealing that net cell growth (NCG) probability followed a hill function as testosterone levels increased. Resistant cells exhibited higher average NCG than sensitive cells. This cellular model was then customized to patient cohorts corresponding to non-recurrent, biochemically determined recurrent (BR), and image-based determined recurrent tumors (TR) by projecting cohort-specific genomic signatures onto the nodes of our systems model. The TR cohort showed the highest NCG predictions.

This cellular systems model informs agent growth in the ABM, which is seeded with a mixed population (sensitive and resistant cell types). Beyond the intrinsic genetic heterogeneity of PTEN deletion levels, we studied other contributors to tissue-level heterogeneity. Parameter scans on extrinsic factors, such as tissue packing densities (crowding), revealed statistically significant differences in tumor composition due to crowding. We also scanned cell-intrinsic properties like testosterone uptake rates, cell adhesion and motility. We found that tumor progression in the BR cohort is biased towards the aggressive cell type at higher cellular testosterone uptake rates. Local sensitivity on cell adhesion-motility revealed clustering and micrometastatic formations in the tissue above distinct thresholds mimicking a sigmoid-like behavior, with the TR cohort showing the highest degree of clustering.

Conventionally, global sensitivity analysis is conducted to identify sensitive model parameters. As a straightforward check for global parametric sensitivity, we conducted Shapley analysis and identified PTEN, MDM4, and AR genes as the most sensitive parameters. We established the significance of these genes identified from our methodology by conducting Kaplan-Meier survival analysis. In ongoing work, we are building formal global sensitivity methods for conducting Sobol sensitivity and uncertainty quantification on complex hybrid models.

Our results demonstrate the potential of multiscale ABMs and coupled data-driven methods in explaining complex tumor behaviors. Such frameworks can be extended to capture heterogeneous responses to therapies and development of resistance against them. Our goal is to create cancer digital twin models that can inform patient-specific precision tumor therapies and enhance clinical decision-making.

Acknowledgements (Optional): We acknowledge funding from NCI under the Physical Sciences in Oncology Network.

References
1. Zhu, L. et al. A narrative review of tumor heterogeneity and challenges to tumor drug therapy. Ann. Transl. Med. 9, 1351 (2021).
2. Dagogo-Jack, I. & Shaw, A. T. Tumour heterogeneity and resistance to cancer therapies. Nat. Rev. Clin. Oncol. 15, 81–94 (2018).
3. Bonabeau, E. Agent-based modeling: Methods and techniques for simulating human systems. Proc. Natl. Acad. Sci. 99, 7280–7287 (2002).
4. Ghaffarizadeh, A., Heiland, R., Friedman, S. H., Mumenthaler, S. M. & Macklin, P. PhysiCell: An open source physics-based cell simulator for 3-D multicellular systems. PLOS Comput. Biol. 14, e1005991 (2018).
5. Kolokotroni, E. et al. A Multidisciplinary Hyper-Modeling Scheme in Personalized In Silico Oncology: Coupling Cell Kinetics with Metabolism, Signaling Networks, and Biomechanics as Plug-In Component Models of a Cancer Digital Twin. J. Pers. Med. 14, 475 (2024).
6. Cess, C. G. & Finley, S. D. Multi-scale modeling of macrophage—T cell interactions within the tumor microenvironment. PLOS Comput. Biol. 16, e1008519 (2020).
7. Rodríguez-Pérez, R. & Bajorath, J. Interpretation of machine learning models using shapley values: application to compound potency and multi-target activity predictions. J. Comput. Aided Mol. Des. 34, 1013–1026 (2020).