Understanding Multistationarity of Fully Open Reaction Networks.

Area of Science:

• * Chemical reaction network theory
• * Computational chemistry
• * Systems biology

Background:

• * Multistationarity, the ability of a system to exist in multiple steady states, is a key property of complex reaction networks.
• * Existing methods primarily focus on generating multistationary networks, with limited tools for non-multistationary cases.
• * Understanding multistationarity is crucial for modeling biological systems and chemical processes.

Purpose of the Study:

• * To introduce a novel deterministic operation for generating new non-multistationary reaction networks.
• * To establish a new graph representation for reaction networks.
• * To apply machine learning for predicting the multistationarity of reaction networks.

Main Methods:

• * Development of a deterministic operation to construct non-multistationary networks.
• * Creation of a novel graph representation for reaction networks, invariant to species name permutations.
• * Training a graph attention neural network (GNN) model using reaction network data.

Main Results:

• * A new method for generating non-multistationary networks, complementing existing multistationary network generation techniques.
• * A unique graph representation enabling consistent analysis of reaction networks.
• * Successful application of a GNN to predict multistationarity, demonstrating the potential of machine learning in this field.

Conclusions:

• * The study provides essential tools for constructing diverse reaction network examples.
• * The novel graph representation and GNN application open new avenues for data-driven analysis in reaction network theory.
• * This work represents the first use of machine learning for classifying reaction networks based on their multistationarity properties.