Programming and training rate-independent chemical reaction networks

Area of Science:

• Biochemistry
• Synthetic Biology
• Computational Science

Background:

• Biochemical systems offer potential for computation beyond traditional electronics.
• Chemical Reaction Networks (CRNs) model natural biochemical systems and serve as a specification for synthetic chemical computation.

Purpose of the Study:

• To identify a class of CRNs robust to reaction rates for reliable chemical computation.
• To develop a method for programming these robust CRNs using deep learning models.

Main Methods:

• Identification of noncompetitive (NC) CRNs, whose behavior depends solely on stoichiometric structure.
• Development of a translation procedure from rectified linear unit (ReLU) neural networks to NC-CRNs.
• Numerical simulations of translated CRNs for machine learning tasks and biological applications.

Main Results:

• NC-CRNs exhibit equilibria robust to reaction rates and kinetic laws.
• A compact translation from binary weight ReLU networks to NC-CRNs was achieved, with a one-to-one mapping between ReLU nodes and bimolecular reactions.
• Simulations demonstrated the feasibility of using translated neural networks for tasks like virus detection and spatial pattern formation.

Conclusions:

• Noncompetitive CRNs provide a foundation for rate-independent chemical computation.
• Neural networks are a suitable paradigm for programming these robust chemical systems.
• This approach opens possibilities for advanced synthetic biology and medical applications.