Designing Machine Learning Tools to Characterize Multistationarity of Fully Open Reaction Networks

Abstract

We present the first use of machine learning tools to predict multistationarity of reaction networks. Chemical Reaction Networks (CRNs) are the mathematical formulation of how the quantities associated to a set of species (molecules, proteins, cells, or animals) vary as time passes with respect to their interactions with each other. Their mathematics does not describe just chemical reactions but many other areas of the life sciences such as ecology, epidemiology, and population dynamics. We say a CRN is at a steady state when the concentration (or number) of species do not vary anymore. Some CRNs do not attain a steady state while some others may have more than one possible steady state. The CRNs in the later group are called multistationary. Multistationarity is an important property, e.g. switch-like behaviour in cells needs multistationarity to occur. Existing algorithms to detect whether a CRN is multistationary or not are either extremely expensive or restricted in the type of CRNs they can be used on, motivating a new machine learning approach. We address the problem of representing variable-length CRN data to machine learning models by developing a new graph representation of CRNs for use with graph learning algorithms. We contribute a large dataset of labelled fully open CRNs whose production necessitated the development of new CRN theory. Then we present experimental results on the training and testing of a graph attention network model on this dataset, showing excellent levels of performance. We finish by testing the model predictions on validation data produced independently, demonstrating generalisability of the model to different types of CRN.