A biomimetic synapse built from water droplets and biological ion channels achieves synaptic plasticity and performs machine learning tasks.
(Nanowerk Spotlight) Biological brains compute using ions, not electrons. Every signal between neurons depends on charged atoms passing through protein channels embedded in cell membranes, a system that processes information at extraordinary efficiency. Engineers have tried to replicate this architecture in hardware, most commonly through memristors, electrical components whose resistance changes based on the history of current that has passed through them.
A memristor behaves somewhat like a synapse, the junction between neurons, which strengthens or weakens depending on past activity. The first practical memristor was demonstrated in 2008 using semiconductor materials, and researchers have since built increasingly sophisticated devices that mimic aspects of brain function. But most rely on electron transport, a fundamentally different mechanism from the ion-based signaling that biology uses. That mismatch limits how faithfully these devices can reproduce the energy-efficient dynamics of real neural circuits.
A newer generation of ion-based memristors attempts to close this gap. Some use solid-state nanochannels etched into glass, graphene, or silicon nitride. Others employ metal ions that form and dissolve conductive filaments. These designs come closer to biological reality, but solid-state channels are rigid, difficult to scale into three dimensions, and often incompatible with living tissue. A few groups have experimented with actual biological ion channels, though results have typically stopped at basic demonstrations without clear paths to computation.
A study published in Advanced Materials (“Synaptic Functionality and Neuromorphic Information Processing in Membrane Ion Channel Junctions”) reports a device that both replicates synaptic learning behaviors and performs machine learning tasks using biological ion channels. A team based primarily at Lawrence Livermore National Laboratory constructed what they call a membrane ion channel synapse, or MICS, by embedding gramicidin A, a natural antibiotic peptide, into a lipid membrane connecting two microscopic water droplets.
The membrane ion channel synapse (MICS) device and its electrical properties. Top left: two aqueous droplets filled with potassium chloride solution sit in an oil bath, connected by a lipid bilayer membrane embedded with gramicidin A ion channels. Electrodes inserted into each droplet apply voltage and measure ionic current. The accompanying microscope image shows the droplets forming the bilayer contact. The remaining panels characterize the device’s memristive behavior: current-voltage curves at different sweep frequencies show the pinched hysteresis loops characteristic of a memristor, with the loop area shrinking at higher frequencies. Single-channel conductance measurements confirm that individual gramicidin A channels become more conductive at higher voltages. Microscope images reveal how the droplet interface physically expands under increasing voltage. The bottom right schematic illustrates how gramicidin A monomers on opposite sides of the membrane assemble into conducting dimers, a process enhanced by voltage-induced compression of the bilayer. (Image: Reproduced with permission from Wiley-VCH Verlag) (click on image to enlarge)
The device exhibits memristive behavior, meaning its electrical conductance depends on its voltage history, and it replicates multiple forms of synaptic plasticity.
The researchers then used MICS as the physical core of a reservoir computing system. In this machine learning architecture, only the output layer requires training. The system classified handwritten digits with up to approximately 89% accuracy and played tic-tac-toe against an optimal opponent, drawing 95% of games under the best conditions.
The device architecture consists of two aqueous droplets, each about 0.5 μL and filled with potassium chloride solution, sitting in a bath of hexadecane oil. Each droplet is coated with a single layer of lipid molecules. When brought into contact, their lipid monolayers merge to form a bilayer, a structure identical to the membranes surrounding living cells. This configuration is known as a droplet interface bilayer. Silver/silver chloride electrodes inserted into each droplet allow the team to apply voltage and measure ionic current across the membrane.
Gramicidin A is the component that creates the memory effect. This peptide forms ion channels in lipid bilayers, but unlike many channel proteins, it assembles from two separate halves, one from each side of the membrane. The halves must align head-to-head to create a conducting pore roughly 0.4 nm wide, just large enough for potassium ions to pass through in single file. Applied voltage compresses the membrane, making it easier for the two halves to meet and stabilize. When voltage drops, channels gradually disassemble. This built-in sensitivity to electrical history gives the device its memory.
Sweeping a triangular voltage waveform across the device produced the signatures of a memristor: a pinched hysteresis loop in the current-voltage curve, a loop area that decreased with increasing sweep frequency, and linear behavior at high frequencies. The characteristic memory time was approximately 100 s, far longer than simple ion diffusion through the short channel would predict.
The team attributes this to two interacting mechanisms: voltage-dependent kinetics of channel formation and disassembly, and voltage-dependent conductance of individual channels, which increased by roughly 22% between 50 mV and 200 mV. Neither mechanism alone accounts for the full conductance change of up to approximately 100%, suggesting that their interplay, amplified by elastic reorganization of the droplet interface under voltage, produces the observed behavior.
The device reproduced several forms of synaptic plasticity. Paired voltage pulses separated by varying time intervals produced paired-pulse facilitation and paired-pulse depression, with time constants of approximately 0.26 s and 0.19 s respectively. These values fall close to the roughly 0.1 s timescales observed in biological synapses.
The system also demonstrated spike-rate-dependent plasticity, where higher-frequency stimulation produced stronger responses. It exhibited spike-timing-dependent plasticity following the Hebbian learning rule as well: the order and timing of pre- and post-synaptic signals determined whether connections strengthened or weakened. The device maintained stable performance over 1 000 set/reset cycles.
The researchers also demonstrated associative learning, modeled on the proboscis extension response of honeybees. Low-voltage pulses represented a conditioned stimulus (an odor), while high-voltage pulses represented an unconditioned stimulus (nectar). After training with paired stimuli, the conditioned stimulus alone triggered a response. This learned association then gradually faded, mimicking extinction, the process by which a brain discards outdated information.
For the reservoir computing demonstrations, the team encoded information as sequences of voltage pulses and read out the resulting conductance states. In the handwritten-digit task, they cropped and binarized images from the MNIST dataset, then divided each into four-bit segments fed sequentially to the device. Because the reservoir converts spatial pixel data into temporal conductance sequences, it compresses each image from 440 input features down to 110 conductance readings. These served as input to a single-layer software neural network with only 1 100 trainable weights, a fourfold reduction compared to a conventional approach that would require 4 400 weights. To account for physical device variability, the team added Gaussian noise calibrated to experimental measurements from four separate MICS devices.
Classification accuracy reached approximately 89% after 100 training epochs, approaching the approximately 91% upper bound of a single-layer network operating on full uncompressed data. Energy consumption ran approximately 2 to 7 nJ per pulse, lower than comparable solid-state memristor systems, though still orders of magnitude above the 0.1 to 100 fJ range of biological synapses.
The tic-tac-toe experiment tested the system’s capacity for tasks requiring precise state discrimination. With pulse durations of 2.0 s, the trained agent drew 95% of 10 000 games against an optimal opponent. The paper notes that forcing a draw is the best possible outcome against such an opponent. Shorter pulses of 0.2 s dropped this figure to 24%, illustrating how the temporal parameters of the physical reservoir directly shape computational performance.
The work amounts to a functional proof of concept rather than a deployable technology. The devices operate on timescales of seconds rather than the milliseconds or microseconds of electronic processors. Scaling to arrays of interconnected droplets, reducing energy consumption, and extending device longevity all remain open challenges. The researchers describe MICS not as a competitor to high-performance silicon electronics but as a complement, particularly suited to applications where biocompatibility and ionic signaling matter.
By demonstrating that a computing element built from biological ion channels and assembled through a flexible droplet-based process can perform meaningful computation while reproducing synaptic dynamics, the study establishes what the authors call a promising platform for neuromorphic ionic computing.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
ORCID information
Aleksandr Noy (Lawrence Livermore National Laboratory)
, 0000-0003-4924-2652 corresponding author
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
