Unpacking the role of complexity in multi-class models of the tumor microenvironment

Abstract

Cancer is an emergent phenotype generally resulting from the disregulation of cell deci- sion processes as well as the disregulation of cell interactions with their microenvironment. However, the tumor microenvironment (TME) is difficult to probe and experimentally in- terrogate. This relative inaccessibility motivates the construction of computational models to generate hypothesies and design high impact experiments. Efforts to study this system holistically neccessitate considerations of the resulting complexity in the compuational mod- els describing it. Leveraging diverse modeling paradigms can generate in silico approaches towards biological parity, but the compounding complexity resulting from this integration must be properly managed.Agent-based models (ABM) are a popular approach to study emergence in multi-scale systems with complex interactions. ABM modularity provides a means of incorporating mutiple classes of models that can be regulated at different scales in an intuitive manner. However, robust analysis methods remain an open challenge. The multi-class nature of many ABMs leads to limited application of traditional parameter assessment methods such as sensitivity analysis or optimization, which can present challenges to their validation. Thus, many techniques to analyze ABMs are analagous to those found in high-resolution experimental methods, where ABMs are treated as in silico test beds. I discuss many of the challenges and approaches to studying biological phenomena with complex ABMs, especially in cases with the dynamic coupling between agents and their environment. I highlight a review article discussing both the growing popularity and the resulting challenges of combining modeling paradigms. I then present work leveraging ma- chine learning techniques to emulate the simulation outputs from an ABM towards priori- tizing data acquisition and resolution in the TME. We identify a gap between the between spatio-temporal emergent phenomena and information used to build robust emulation mod- els. Counter-intuitively, spatial information confers far less benefit than leveraging temporal information to predict tumor aggression metrics. I also present results towards leveraging ABMs as an platform for translating in vitro derived models to in situ contexts. I integrate a mechanistic model of hypoxia-induced factors (HIFs), an ubiquitous feature in cancer progression, into a TME ABM. This platform enables predictions of additional regulation requirements for key growth factors. I conclude with a perspective on the current state of code commonly found in academic journals and methods to improve code and reproducibility, through analogies from experi- mental biology. I then highlight many possible directions of continuing and extending the research outlined in this dissertation specifically in the context of emulating agent-based models and characterizing angiognesis in the TME.