Demarcating good solutions in system biology computer models using artificial neural networks

Abstract: How close a computer model comes to recreating real-world phenomena often depends on the value of its internal parameters, but investigating the outcome of the model for every point in parameter space is in practice an impossible task. Here an artificial neural network is used as a numerical predictor on two different system biology computer models. A semi-implicit solver was also implemented for one of these models in order to speed up simulations in stiff regions of parameter space. The performance of the neural networks were measured using the area under the receiver operating characteristic curve (AUC), and neural networks were used as numerical predictors for three different four-dimensional parameter regions. In the first region a training data set of 500 points were used and an auc of 1.0 was achieved. In the second region a training data set of 1000 points were used and an auc of 0.97 was obtained. In the last region training data sets of 100, 250, 1000 and 3000 points were used and the auc of the neural networks was 0.86, 0.95, 0.97 and 0.97 respectively.