Rational design of computable CRNs. Our research presents a crucial experimental stepping stone to gear current design efforts in chemical reaction networks (CRNs)—interconnected webs of chemical reactions—into the direction of neuromorphic systems. The ability to use CRNs to rationalize ‘learning’ will break major new ground in the domain of autonomous molecular materials and out-of-equilibrium systems chemistry. Our work impacts (bio)chemical sciences at large because insights into this type of in-material computing will stimulate new impetus for predicting drug-target interactions, optimizing diagnostics, and interfacing with soft robotics. Ultimately, our experimental platform can be tailored to the complexity observed in theoretical, biological, or societal contexts that are governed by differential equations, and uncover unconventional methods to go beyond current approaches in the realm of machine learning.
background
Computing with chemistry. Approaches to computing using molecules have relied heavily on nucleic acid-based polymers (DNA), sequence-specific polymers, or molecular logic gates but have proven impractical. Other more recent attempts incorporate elements of synthetic biology, chemical reaction network, and control theory to mimic the capabilities of the human brain, underscoring how molecular systems could become reconfigurable under dynamic environments. For the next-generation chemical computer, we need a general design strategy to exploit the potential of the nonlinear and nonbinary nature in chemical reactions. Novel strategies are emerging in systems chemistry—a field that creates and studies networks of interacting molecules—but rationalizing the ability to learn and adapt goes beyond current synthetic designs. Most demonstrations focus on showing how the chemical synthesis of feedback loops, which are sets of chemical reactions representing the smallest possible control module in biochemical networks, could form the basis for behavior necessary for living systems (e.g. bistability, oscillations, and other transient assemblies). The common challenge many efforts faces is that the underlying design remains relatively simple, limited to a set of connected feedback loops: Strategies to rationally scale the complexity are scarce.
state of the art
Out-of-equilibrium chemical reaction networks as chemical neurons. In our research group, we use so-called continuously stirred flow tank reactors, CFSTRs, to force networks of enzymatic reactions to adapt to their environment (Fig. 1a). Under continuous flow conditions, trypsinogen (Tg) acts as a fuel for generating the catalyst trypsin (Tr), creating a positive feedback loop (+). Its process is regulated by the inhibitor (I), which can suppress the activation of the autocatalytic reaction (-). Importantly, when we used a gradient in the input concentration ([I]o) to probe the stability of the network, we found that the network exhibited memory effects. We measured the pathways by which activation and inhibition can be reached, and discovered various history-dependent responses (such as hysteresis, synchronization, and adaptation, Fig. 1b). Hence, this work demonstrates that a CRN, with as little as three components, is sufficient to emulate the working principles of neurons.
|
Fig. 1. History dependency in chemical reaction networks can drive nonlinear activation patterns. (a) Illustration of an autocatalytic network based on three enzymes (trypsin (Tr), trypsinogen (Tg), and soybean trypsin inhibitor (I)) and the three corresponding main processes. (b) Phase portraits of the network under different gradients, d[I]0/dt. Hysteresis occurs when the reaction rates outcompete the rate of the slope. |
current research
Autonomous reconfiguration in networks of CRNs. Building on this foundation, we are currently developing a microfluidic platform to scale our concept of neuron-like activity in CRNs and combine them with surface functionalization techniques to incorporate local reactions. This integration creates conditions for large-scale CRNs which we envision to produce complex spatiotemporal patterns. Fig. 2 illustrates the multi-dimensional output, comprising a set of steady states from which a subset can be selected dynamically. In greater detail, in our experiments, we use chaotic time sequences as input with the duration and the amplitude of a ‘write pulse’ defining the change in control parameters (e.g. inhibitor concentration). Under continuous perturbations, various nodes in the network of CRNs will be affected, and yield an output that we measure by a set of defined variables (x1, x2, x3, …, xn, typically these are concentrations of components produced). The process is the continuous reconfiguration within the CRNs, which we aim to capture as multi-layered patterns.
|
Fig. 2. Proposed computing scheme, in which connected CRNs are forced to dynamically reconfigure, resulting in multiple decision schemes. |
