From Years to Hours: Accelerating Model Refinement

Network-based dynamic modeling is useful for studying how complex biomolecular systems respond to environmental changes and internal perturbations. The main challenge in constructing a dynamic model is integrating evidence from perturbation (e.g. gene knockout) experiments, because their results arise from the collective function of the regulatory network. For a model to capture these non-local effects, its construction, validation, and refinement necessarily involve trial and error, constrained by domain knowledge. We propose and implement a genetic algorithm-based workflow to streamline model refinement. This workflow applies to any biological system for which an interaction network and enough perturbation experiments exist. The algorithm we introduce adjusts the functions of the model to enhance agreement with a corpus of curated experimental results and leverages existing mechanistic knowledge to automatically limit the search space to biologically plausible models. To account for the interdependence of experimental results, we develop a hierarchical scoring technique for assessing model performance. We implement our workflow for Boolean networks, which are a popular and successful tool for modeling biological systems, but the workflow is readily adaptable to multi-level discrete models. Our implementation is available as the open-source Python library boolmore . We demonstrate boolmore ’s effectiveness in a series of published plant signaling models that exemplify the challenges of manual model construction and refinement. These models describe how plant stomata close in response to the drought hormone abscisic acid. After several hours of automatic refinement on a personal computer, the fittest models recapture and surpass the accuracy gain achieved over 10 years of manual revision. The refined models yield new, testable predictions, such as explanations for the role of reactive oxygen species in drought response. By automating the laborious task of model validation and refinement, this workflow is a step towards fast, fully automated, and reliable model construction. Author summary Biomolecular networks are quintessential complex systems, wherein the interactions of proteins and molecules give rise to cellular phenotypes. Modeling these systems requires making choices about the rules governing individual genes and proteins, but often experiments only constrain their effect on the system-level behavior. This contrast presents a challenge to updating an existing model to align with new experiments. The traditional approach to revising a baseline model is essentially trial-and-error. We present a method, implemented as the open source Python library boolmore , that leverages recent advances in the computational analysis of discrete dynamical systems to automate this process, reducing a task that often takes years to a matter of several hours on a personal computer. We showcase the power of this method on a model describing how plant leaf pores respond to the drought hormone abscisic acid. This model was first published in 2006 and has been updated several times, by hand, to incorporate new experimental data or to improve model performance. Boolmore not only recaptures these refinements, but produces models that better explain experimental results and uncover new insights into the regulatory mechanisms of drought response. ### Competing Interest Statement The authors have declared no competing interest.