Reproducible, Robust, and Reliable Biochemical Reaction Network Models for Systems Biology

Abstract

Reproducibility, robustness, and reliability are features desired for a biochemical reaction network model. Scientific research is reproducible when the findings can be independently verified and reproducibility is crucial for the integrity of science. Unfortunately, however, scientific studies, including computational studies, are often not reproducible. It is hard to achieve robustness and reliability due to technical difficulties with experiments producing high quantity, high-quality data and inherent unidentifiability arising from multiparametric nature of biological processes. Robustness and reliability can be achieved by obtaining more data and by implementing better computational algorithms. To improve the reproducibility of biochemical reaction network models, software tools are necessary. We need novel algorithms to increase the robustness and reliability of models. Most of all, a scalable and extensible computing environment is necessary for incorporating tools and algorithms. Therefore, we build a Python-based modeling and simulation environment called Tellurium to ensure reproducibility of studies while supporting a wide array of tools to help design robust and reliable models. Tellurium is specifically designed for high-throughput studies which are necessary to deploy novel modeling algorithms. Next, software tools to improve the reproducibility of computational studies have been built and integrated into Tellurium. In particular, Python support for standards related to describing simulation experiments has been improved. Lastly, two algorithms to help construct robust and reliable mechanistic models have been designed. The algorithms actively explore the concepts of ensemble modeling and specifically utilize the data from perturbation studies. It is demonstrated that a model ensemble can provide reasonable predictions on the system of interest. The idea of the model ensemble directing future experiments toward maximal reduction of the potential topology of models in the ensemble is also explored. Once deployed, ensemble-based experiment selection is expected to close the cycle between modeling and experimental endeavors, bridging the disparity between data-driven modeling and modeling-driven data collection.