Neural CRNs: A Natural Implementation of Learning in Chemical Reaction Networks

Area of Science:

• Synthetic biology
• Bioengineering
• Computational neuroscience

Background:

• Existing chemical neural computing relies on discrete-layered architectures and steady-state kinetics.
• This approach presents limitations in practicality and complexity for advanced computations.

Purpose of the Study:

• To propose an alternative framework for molecular learning circuits using continuous-time chemical kinetics.
• To demonstrate the viability and advantages of this analog approach for neural computation.

Main Methods:

• Modeling neural computations through the continuous-time evolution of molecular concentrations.
• Implementing supervised learning pipelines with minimal phases.
• Utilizing unimolecular and bimolecular reactions for linear and nonlinear circuit implementation.
• Incorporating first-order gradient approximations for scalable nonlinear models.

Main Results:

• Successfully assembled an end-to-end supervised learning pipeline using only two phases.
• Demonstrated implementation of both linear and nonlinear modeling circuits using simple reaction orders.
• Showcased linear scaling of nonlinear models with input dimensionality via gradient approximations.
• Validated circuit constructions through simulations for regression and classification tasks.

Conclusions:

• The proposed continuous-time framework offers a practical and viable pathway for embedding learning behaviors in synthetic biochemical systems.
• This approach simplifies circuit design by avoiding higher-order chemistries and enabling efficient scaling.