Listening to Molecules Speak: How Spiking Networks Interpret Chemical Signals

Researchers at the University of Oslo report a pilot study exploring whether neuromorphic processors can be used to estimate gamma-aminobutyric acid (GABA) concentration from dielectric relaxation spectroscopy measurements. The work combines laboratory permittivity measurements, machine learning models, and spike-based hardware to evaluate how different computational approaches process biochemical sensor data.
GABA is a major inhibitory neurotransmitter in the central nervous system, and its concentration is relevant for studying conditions such as epilepsy, anxiety, and depression. Existing measurement methods, including magnetic resonance spectroscopy, provide reliable results but require expensive equipment and specialized infrastructure. The study investigates an alternative approach based on dielectric spectroscopy and data-driven modeling, with a focus on whether neuromorphic hardware can process such data under constrained numerical precision.
Measuring GABA with Dielectric Spectroscopy
The researchers prepared aqueous GABA solutions with concentrations ranging from 10 to 200 micromolar. Using an open-ended coaxial probe connected to a vector network analyzer, they measured the dielectric properties of each solution across frequencies from 200 MHz to 14 GHz. All measurements were conducted at a constant temperature of 22 °C.
Dielectric relaxation spectroscopy characterizes how materials respond to electromagnetic fields over a broad frequency range. From the measured spectra, four properties were extracted. those being dielectric constant, dielectric loss, conductivity, and loss tangent. These values vary systematically with concentration, providing a numerical basis for predicting GABA levels without relying on electrochemical reactions.
Mapping Electrical Properties to Concentration
To test whether these dielectric features contain enough information to estimate concentration, the group trained two machine learning models. One known as a feedforward neural network and the other a two-dimensional convolutional neural network. Both models were trained using normalized data and evaluated on a held-out test set.
The convolutional model produced the strongest results, achieving a minimum mean absolute error of approximately 0.86 micromolar. Given the overall concentration range of the dataset, this error is small relative to the scale of the measurements and indicates that the dielectric data encode concentration-dependent information in a consistent way. These results confirm that convolutional architectures can model the relationship between permittivity features and GABA concentration with high numerical accuracy.
Quantization and Deployment on Neuromorphic Hardware
The next stage of the study examined whether this trained convolutional model could be deployed on a neuromorphic processor. The researchers used the Akida 1000 platform, which operates using spike-based communication and integer-only arithmetic.
Before deployment, the model underwent quantization, converting 32-bit floating-point values into 8-bit unsigned integers. This step is required for compatibility with the hardware but reduces numerical resolution. After quantization, the convolutional model maintained reasonable predictive performance, with a mean absolute error of approximately 1.48 micromolar.
The quantized model was then converted into a spiking neural network using Akida’s CNN-to-SNN conversion framework. At this stage, predictive accuracy decreased substantially, with a mean absolute error of approximately 49 micromolar. This reduction reflects both the limited precision of the hardware and the difficulty of representing continuous permittivity values using discrete spike events.
The study focused on feasibility and predictive accuracy rather than detailed measurements of processor power consumption.
Interpreting the Accuracy Gap
The performance gap between the floating-point model and its spike-based counterpart highlights a key compatibility issue. Permittivity measurements are continuous, high-resolution numerical signals with a wide dynamic range. In contrast, spiking systems encode information using discrete events governed by firing thresholds and integer weights.
Quantization compresses decimal-rich values into fewer representable states, which can obscure subtle variations in the input data and reduce sensitivity to concentration-dependent changes. In this study, that loss of resolution limited how effectively the spiking network could preserve concentration-dependent information.
The authors note that this issue is not unique to GABA sensing. It reflects a broader tension between analog-like scientific measurements and event-driven computational frameworks. Improving how continuous values are encoded for spike-based processing remains an active area of research.
One proposed direction involves rate-based encoding, where numerical magnitude is represented by spike frequency over time rather than by single threshold crossings. This approach more closely matches how biological neurons represent graded signals and may retain more information after quantization.
Software Integration and User Interface
Alongside the modeling work, the researchers developed a graphical interface using the Tkinter library. The interface allows users to manage data transfer between the measurement system and the neuromorphic processor, store results, and visualize outputs.
The application is lightweight and cross-platform, supporting Windows, macOS, and Linux. Its role is practical rather than analytical: it simplifies interaction with the hardware and supports reproducible experimentation. While it does not alter model performance, it provides a usable framework for future development and testing.
Implications for Biosensing Systems
The study establishes that convolutional neural networks can accurately estimate GABA concentration from dielectric spectroscopy data under floating-point computation. It also shows that current neuromorphic hardware struggles with this type of continuous input when strict quantization and spike-based conversion are applied.
These findings are informative for the design of future biosensing systems. They indicate that successful deployment of spike-based processors for biochemical analysis will depend on advances in numerical precision, encoding strategies, and hardware flexibility. At present, the main contribution of the neuromorphic experiments lies in identifying where existing platforms fall short.
Outlook
The authors emphasize that this work is exploratory. It clarifies which aspects of permittivity-based sensing are compatible with current machine learning methods and which aspects pose difficulties for neuromorphic processors. Future research may benefit from newer generations of neuromorphic hardware that support higher precision, programmable activation behavior, and temporal encoding schemes.
By examining both successes and constraints in a controlled pilot study, the research provides a clear reference for assessing how spike-based systems interact with biochemical measurement data.
In summary, the study shows that dielectric spectroscopy combined with convolutional neural networks can yield accurate GABA concentration estimates. It also shows that converting such models into spiking form on current neuromorphic hardware introduces substantial numerical constraints. Rather than presenting a finished solution, the work defines the technical conditions under which neuromorphic computing may or may not be suitable for biochemical sensing applications.
This article is based on the research paper: Hou, J., Ali, A. H., & Martinsen, Ø. G. (2025). Using neuromorphic computing in prediction of GABA concentration – a pilot study. Journal of Electrical Bioimpedance, 16, 146–149. https://doi.org/10.2478/joeb-2025-0019
