Improving Boolean models of biological systems using a genetic algorithm

The model-checking and improvement process is one of the bottlenecks of model construction. We present a genetic algorithm-based tool that provides a systematic method of comparing a model to various experimental data and creating improved versions. The algorithm starts with a pre-existing model consisting of an interaction graph and a Boolean function for each node. Many offspring models are created that are consistent with the starting interaction graph and optional constraints. Each model is scored based on the agreement of experimental perturbation results (expressed in Boolean form) with the model’s attractors under the same perturbations. The models with the highest score are kept and more offspring are generated from them. Repeated iteration generates models that agree far better with the experimental input than the starting model.

For a benchmark, artificial experiments were generated from models uploaded on The Cell Collective. The algorithm was able to reach a similar score as each original model when started from randomly scrambled models.

As a case study, the algorithm was used to improve a previously published model of abscisic acid induced stomatal closure in plants, and showed a good performance. Notably, the algorithm was able to recapitulate the manual improvements made to the model in two follow-up publications.