🤖 AI Summary
This work addresses the challenge of realizing efficient, autonomous, and general-purpose learning in mass-action chemical reaction systems. We propose a fully analog neural chemical reaction network (CRN) paradigm that directly encodes neural computation into continuous molecular concentration dynamics—bypassing discrete-state mappings entirely. Methodologically, we introduce a clock-synchronized supervised learning scheme requiring only a two-phase clock and design compact CRN circuits; our minimal linear regression CRN comprises just 13 reactions and 15 molecular species. Experiments demonstrate feasibility across diverse linear/nonlinear regression and classification tasks, while confirming robustness to intrinsic noise and parameter perturbations. To our knowledge, this is the first demonstration of native concentration-based dynamical learning—eliminating reliance on digital abstractions. The approach markedly enhances biological realizability and circuit compactness, establishing a deployable foundation for chemical intelligence in synthetic biology and adaptive biomedical applications.
📝 Abstract
We present Neural CRNs, an efficient, autonomous, and general-purpose implementation of learning within mass action chemical reaction systems. Unlike prior works, which transliterate discrete neural networks into chemical systems, Neural CRNs are a purely analog chemical system, which encodes neural computations in the concentration dynamics of its chemical species. Consequently, the chemical reactions in this system stay true to their nature, behaving as atomic end-to-end computational units, resulting in concise and efficient reaction network implementations. We demonstrate this efficiency by assembling a highly streamlined supervised learning procedure that requires only two clock phases. We further validate the robustness of our framework by constructing Neural CRN circuits for several linear and nonlinear regression and classification tasks. Furthermore, a minimal linear regression circuit is assembled using only 13 reactions and 15 species. Our nonlinear modeling circuits significantly advance the state-of-the-art through compact and simple implementations. The synergistic nature of our framework with the analog chemical computing hardware leaves ample room for optimizations and approximations in the computational model, several of which are discussed in this work. Our work introduces a novel paradigm for chemical computing and learning, providing a foundational platform for future adaptive biochemical circuits with applications in fields such as synthetic biology, bioengineering, and adaptive biomedicine.