Data-Driven Inference of Chemical Reaction Networks via Graph-Based Variational Autoencoders

We propose a data-driven machine learning framework that automatically infers an explicit representation of a Chemical Reaction Network (CRN) together with its dynamics. The contribution is twofold: on one hand, our technique can be used to alleviate the computational burden of simulating a complex, multi-scale stochastic system; on the other hand, it can be used to extract an interpretable model from data. Our methodology is inspired by Neural Relational Inference and implements a graph-based Variational Autoencoder with the following structure: the encoder maps the observed trajectories into a representation of the CRN structure as a bipartite graph, and the decoder infers the respective reaction rates. Finally, the first two moments of the stochastic dynamics are computed with the standard linear noise approximation algorithm. Our current implementation demonstrates the applicability of the framework to single-reaction systems. Extending the framework to- wards inferring more complex CRN in a fully automated and data-driven manner involves implementation challenges related to neural network architecture and hyperparameter search, and is a work in progress.