Link to the code: brain-emulation GitHub repository

Printed Artificial Neurons Communicate with Living Brain Cells

Whole brain emulation has a hardware prerequisite that rarely gets explicit attention: artificial neurons that can actually talk to biological ones. Electrode arrays record and stimulate. Flexible mesh implants conform to tissue. But none of these devices compute the way a neuron computes. They are input-output transducers, not processing nodes.

A team at Northwestern University has changed that framing. In April 2026, they reported the first printed artificial neurons, fabricated from molybdenum disulfide (MoS₂) and graphene, that successfully communicate with living mouse brain cells. The printed devices are not passive electrodes. They generate action potential-like outputs in response to biological inputs, and those outputs trigger responses in adjacent biological neurons. The communication runs in both directions.

This places printed semiconductor neurons in a different category from everything that came before, including the bacterial protein nanowires developed at UMass Amherst that established biological-voltage operation last year. The Northwestern work uses inorganic materials, fabricated through printing processes rather than biological growth, with implications for manufacturing scale that protein-based approaches cannot easily match.

The Materials

Molybdenum disulfide is a transition metal dichalcogenide, a class of two-dimensional semiconductor materials that has attracted intense interest in electronics research over the past decade. In thin-film form, MoS₂ exhibits semiconducting properties that silicon cannot match at nanometer thicknesses. Its biocompatibility has been demonstrated in cell culture studies, and its surface chemistry allows functionalization with biological molecules.

Graphene provides the conductive backbone. Combined with MoS₂, the composite material produces a structure that can be patterned through printing. The Northwestern team used this approach to fabricate neuron-shaped devices where the geometry itself contributes to the signal processing characteristics, as it does in biological neurons where axon diameter and myelination determine conduction velocity and integration properties.

The critical performance metric for any device interfacing with neural tissue is voltage compatibility. Biological neurons operate in the 100 millivolt range. Silicon electronics traditionally run at voltages an order of magnitude higher, requiring signal conditioning circuitry that adds complexity and potential sources of noise and failure. The MoS₂-graphene composite operates within the biological voltage range without conversion circuitry, directly addressing the compatibility problem that has limited earlier hybrid interface attempts.

The Experiment

The Northwestern team tested their printed neurons on acute mouse brain slices, a standard ex vivo preparation that preserves local circuit connectivity while allowing experimental access to tissue that would be inaccessible in a living animal. Brain slices maintain viable neurons for several hours under appropriate conditions and allow simultaneous recording from multiple sites.

The printed neurons were positioned in contact with the tissue. When biological neurons in the slice generated electrical activity, the printed devices detected those signals. When the printed devices were driven electrically, they produced outputs that propagated to neighboring biological neurons, triggering responses measurable by standard electrophysiology.

The demonstration satisfied the two conditions required for any candidate replacement neuron technology: afferent communication (receiving signals from biological neighbors) and efferent communication (generating signals that biological neighbors respond to). Both functions were confirmed in the same preparation, establishing bidirectional operation rather than one-way sensing or stimulation.

Comparison to Existing Approaches

The landscape of neural interface technologies has expanded considerably in recent years, and the Northwestern work occupies a specific and previously empty niche within it.

Flexible neural implants using graphene and polymer substrates have demonstrated months-long stable recording in animal models, significantly outperforming rigid silicon arrays that provoke tissue responses within weeks. These devices record and stimulate, but they do not compute. They relay signals between brain and external processors.

The bacterial nanowire neuron developed at UMass Amherst operates at biological voltage and interfaces with living cardiomyocytes, demonstrating that non-silicon artificial neurons are achievable. That work used Geobacter sulfurreducens protein filaments, a biological material grown through microbial culture. The approach is scientifically compelling but faces manufacturing challenges. Printed inorganic semiconductors can be patterned using processes already developed for flexible electronics manufacturing, a supply chain that exists at industrial scale.

Neuromorphic chips such as Intel Loihi and IBM NorthPole implement spiking neural network computation in silicon but are not designed for direct biological tissue integration. They process signals from sensors and produce outputs for actuators. The gap between a neuromorphic chip and a device that sits inside biological tissue and participates in local circuit computation has not previously been bridged with a manufacturable technology.

The Northwestern printed neurons occupy the gap: devices that compute locally, operate at biological voltages, and establish genuine chemical and electrical communication with biological neurons.

Relevance to the Moravec Transfer

Hans Moravec’s 1988 gradual neuron replacement proposal remains theoretically significant because it sidesteps the copy problem that undermines destructive scanning approaches. The Moravec Transfer posits replacing biological neurons one at a time with artificial equivalents that preserve the same function. Consciousness, on this account, migrates with the gradual replacement rather than being interrupted and restarted.

The proposal requires three capabilities. An artificial neuron must receive inputs from biological neighbors. It must process those inputs according to the same computational rules as the neuron it replaces. It must produce outputs that biological neighbors respond to.

The Northwestern work demonstrates the first and third. The second remains the harder problem. Faithful computation requires not just signal exchange but the correct transformation of inputs to outputs, including the plasticity rules that govern how that transformation changes over time with experience.

Behavioral timescale synaptic plasticity, recently identified as the primary learning rule in hippocampal circuits, is one example of the computational rules an artificial replacement neuron must eventually implement. Compartment-specific plasticity in motor cortex, where apical and basal dendrites follow different rules, is another. The hardware challenge of printed neurons is being addressed. The computational fidelity challenge remains open.

Limitations of the Current Work

The April 2026 results come from ex vivo brain slices, not living animals. The difference matters for several reasons. Slice preparations lack the vascular environment, immune system activity, and long-range circuit connectivity of intact brain tissue. Inflammation responses, glial encapsulation, and long-term material degradation, all of which determine whether a device remains functional over months or years in vivo, cannot be assessed in acute slice experiments.

The devices demonstrated in the paper do not implement synaptic plasticity. A fixed transfer function can replicate a neuron’s baseline behavior but cannot adapt to learning. For any application in a living animal over timescales longer than a single session, plasticity is not optional.

Scaling from individual devices to the billions of neurons in a human brain requires manufacturing consistency that has not been demonstrated. Individual printed neurons in laboratory conditions and wafer-scale production of identical devices with controlled variation are different engineering problems.

TRL Assessment

TRL 3-4. Proof of concept validated in living tissue (ex vivo). The technology demonstrates bidirectional communication between printed neurons and biological neurons under controlled conditions. Transition to in vivo animal models, long-term stability testing, and integration of plasticity mechanisms would advance the readiness level toward TRL 5-6.

Path Forward

The immediate experimental next steps are in vivo validation in rodent models and long-term stability characterization. The more difficult medium-term challenge is implementing plasticity in the printed devices, whether through material properties, analog circuit design, or hybrid approaches combining printed hardware with embedded digital processing.

The result from Northwestern is notable not because it solves the problem of building a replacement neuron but because it demonstrates that the materials and fabrication approaches required to build such a neuron are accessible with current manufacturing technology. The field has moved from the question of whether inorganic materials can communicate with biological neurons to the question of what computation those communications must implement.

Official Sources