Blog

Researchers at the Institute for Molecules and Materials at Radboud University, the Netherlands, have demonstrated that a complex self-organizing network of chemical reactions can perform a variety of computational tasks, such as nonlinear classification and complex dynamics prediction.

The field of molecular computing interests researchers who want to harness the computational power of chemical and biological systems. In these systems, chemical reactions or molecular processes act as the reservoir computer, transforming inputs into large-scale outputs.

The study, published in Naturewas led by Prof. Wilhelm Huck of Radboud University.

Researchers have taken advantage of the potential chemical and biological networks offer because of their sophisticated computational capabilities. However, the implementation of molecular computing presents an engineering and design challenge.

Rather than trying to design molecular systems to perform specific computational tasks, Prof. Huck and his team are investigating how naturally complex chemical systems can exhibit emergent computational properties.

“I’m very interested in the chemical driving forces that led to the origin of life. In this context, we search for mechanisms by which chemical evolution can shape the properties of complex reaction mixtures. This study led us to consider how molecular systems can process information,” he explained to Phys.org.

The formic reaction

The formose reaction is a chemical reaction that synthesizes sugars from formaldehyde in the presence of a catalyst, calcium hydroxide. This reaction was chosen because of its unique properties.

Prof Huck explained: “Although chemistry may seem complicated to outsiders, most reaction sequences are rather linear. The formose reaction is the only example of a self-organizing reaction network with a highly nonlinear topology containing multiple positive and negative feedback loops.”

In other words, the reaction is not simple and produces multiple intermediates that react further to form new compounds. These dynamic reactions can lead to a diverse array of chemical species and are nonlinear in nature.

In addition, the network includes positive feedback loops that amplify response outcomes and negative feedback loops that decrease response outcomes.

The network is known as “self-organizing” because it naturally evolves and responds to chemical input without the need for external intervention, producing a diverse range of outputs.

Computational capabilities derive from the inherent properties of the network rather than being explicitly programmed, making the computation very flexible.

Implementation of the tank computer

The researchers used a continuous stirred tank reactor (CSTR) to carry out the formose reaction. The inlet concentrations of four reactants—formaldehyde, dihydroxyacetone, sodium hydroxide, and calcium chloride—are controlled to modulate the behavior of the reaction network.

The starting molecule is identified using a mass spectrometer, which allows them to track up to 106 molecules. This setting can be used to perform calculations with the concentrations of the reactants as the input value for each function to be calculated.

But first, the system must be trained to find the result of this calculation, which is done using a set of weights.

“We need to find a set of weights that convert the traces in the mass spectrometer to the correct calculation value. It’s a linear regression problem and it’s computationally simple. Once that’s done, the tank computer calculates the score for this function for each new input,” he explained. Prof. Huck.

Weights are coefficients that determine the influence of each input on the output. This learning step is essential because it allows the tank to learn and predict how changes in the input affect the output so that it can predict the output for a new set of inputs.

Computational capabilities

The researchers used the tank computer to perform several tasks. The first was the performance of non-linear classification tasks. The tank computer can emulate all Boolean logic gates and even handle more complex classifications such as XOR, checkers, circles and sine functions.

The team also showed that it could predict the behavior of a complex metabolic network model of E. coli, accurately capturing both linear and nonlinear responses to input fluctuations over different concentration ranges.

Furthermore, the system demonstrates the ability to predict future states of a chaotic system (the Lorentz attractor), accurately predicting two out of three input dimensions several hours into the future.

The research team also found that certain chemical species in the system exhibit short-term memory, retaining information about past inputs.

They also demonstrated a proof of concept for fully chemical sensing using colorimetric reactions, showing how the state of a system can be interpreted without electronic measuring devices.

In other words, the state of the system can be interpreted using color changes from chemical reactions, eliminating the need for electronic measuring devices.

Origin of Life, Neuromorphic Computation and Beyond

This new approach to molecular computing could bridge the gap between artificial systems and the information-processing capabilities of living cells.

It offers a more scalable and flexible approach to molecular computing, opening up possibilities for creating autonomous chemical systems that can process information and respond to their environment without external electronic control.

Prof Huck expressed his team’s interest in this area, saying: “Can we build reservoir computing into chemical systems that sense their environment, process that information and take the right actions?

“This would require connecting the reservoir to other elements that could translate the output of the chemical brain into some form of mechanical response or into interaction with living cells, for example.”

The research also has intriguing implications for the origins of life. The emerging computational properties of this relatively simple chemical system may provide insight into how early biological systems might have developed information-processing abilities.

Prof. Huck mentioned that this was his main motivation to study reservoir computing.

The research team also sees potential in neuromorphic computing, which mimics the neural structure and functioning of the human brain to improve computational efficiency and power.

“We are very interested in exploring the technological limits of the computational power of the formose tank computer—this is ongoing research in collaboration with IBM Zurich. “Reservoir computing is an example of neuromorphic computing that has attracted interest because it is expected to consume less energy than conventional computers,” explained Prof Huck.

© 2024 Science X Network