Predicting The Resource Needs And Outcomes of Computationally Intensive Biological Simulations

Accurately simulating viral dynamics within a human body is often computationally intensive, thus requiring dedicated computing infrastructures. This limits the use of simulations in personalized medicine since patients may not all have access to such infrastructures or the possibility of waiting weeks for predictions. Lowering the computational burden by simplifying the models is a challenge since simulation results must remain accurate in order to safely support the decision-making activities of patients. In this work, we use machine learning to predict the results of a simulation (i.e. surrogate model) and its computational resources (i.e. performance model). This allows patients to use machine learning as a computationally light proxy to simulations and to know how long an accurate simulation would take. We demonstrate that machine learning can build such surrogate and performance models with low errors for five previously developed models of HIV, while outperforming interpolation methods commonly used for surrogate modeling.