Unsupervised learning of chemical reaction networks from time-series data

Abstract

We consider time series data coming from multiple initial states generated by an hidden model. Using a clustering-based method, we develop an algorithm for the inference of biological reaction network. The output is a set of reactions providing us with biological understanding, and which can be used to generate new traces. A model selection method is derived from these newly generated traces. We evaluate the performance of this algorithm on a range of models from Biomodels and discuss its benefits and limits, specifically in the case of subsampled data.