Bio-interfacing artificial neurons: revision 1 to 2

Revision 1

Revision 2

categories: ["cybernetics"]

tags: ["neuromorphic", "memristors", "organic electronics", "bioelectronics", "neural interface", "neuroprosthetics"]

summary: "Physical electronic devices that generate neuron-like spikes at biological voltages and timescales so they can exchange signals directly with living nerve cells and tissue."

summary: "Physical electronic devices that generate neuron-like spikes at biological voltages and timescales so they can exchange signals directly with living nerve cells and tissue."

updated: "2026-10-27"

updated: "2026-10-29"

humanEvidence: "No human study has been run; every coupling to living tissue has been acute, in mouse brain slices, cultured rat neurons, stem-cell-derived heart cells in a dish, or the vagus nerve of anaesthetized mice."

humanEvidence: "No human study has been run; every coupling to living tissue has been acute, in mouse brain slices, cultured rat neurons, stem-cell-derived heart cells in a dish, or the vagus nerve of anaesthetized mice."

access: "Nothing to obtain: these are laboratory prototypes with no approved device, no commercial product, and no reported human trial."

access: "Nothing to obtain: these are laboratory prototypes with no approved device, no commercial product, and no reported human trial."

reversibility: "difficult"

"Artificial neuron" names three different things, and news coverage regularly blurs them.

The oldest is mathematical. The units of a software neural network, descended from the

McCulloch–Pitts model of 1943, are weighted sums passed through a nonlinearity. They have no

McCulloch-Pitts model of 1943, are weighted sums passed through a nonlinearity. They have no

physical form and no electrical contact with anything biological, and they are what the phrase

physical form and no electrical contact with anything biological, and they are what the phrase means

means in machine learning and in debates about [[artificial-general-intelligence]].

in machine learning and in debates about [[artificial-general-intelligence]].

The second is neuromorphic hardware: circuits that emulate spiking neurons physically, mostly to

compute more efficiently. Mahowald and Douglas built an analog "silicon neuron" in 1991 that

hardware since then has been designed for computing rather than for contact with tissue, and most

artificial neurons have operated at signal amplitudes of half a volt or more.[^fu2025]

The third, the subject here, is the bio-interfacing subset: devices whose purpose is to sit in or

The third, the subject here, is the bio-interfacing subset: devices built to sit in or next to

next to living tissue and exchange signals with it. That imposes requirements computing does not.

living tissue and exchange signals with it. That imposes requirements that computing does not.

Biological action potentials are on the order of 100 millivolts and last about a millisecond; the

Biological action potentials are on the order of 100 millivolts and last about a millisecond, the

device must work in salty, wet surroundings; and, ideally, it responds to the chemical signals

device must work in salty, wet surroundings, and ideally it responds to the chemical signals neurons

neurons use as well as the electrical ones.

use as well as the electrical ones.

## How it works

**Analog silicon neurons** model each ion channel with a transistor circuit. A University of Bath

team fitted conductance models to recordings from rat hippocampal and brainstem respiratory neurons

and used the fitted parameters to configure silicon neurons, which responded nearly identically to

the biological cells; the completed models matched recorded membrane voltage with 94–97% agreement

the biological cells. The completed models matched recorded membrane voltage with 94-97% agreement

across 60 current-injection protocols.[^abuhassan2019] The authors propose such circuits as a

across 60 current-injection protocols.[^abuhassan2019] The authors propose such circuits as a route

route to implants that repair failing neural circuits; the study itself compared chip and neuron

to implants that repair failing neural circuits, but the study itself compared chip and neuron

offline and connected nothing to tissue.

**Memristive neurons** use a component whose resistance switches abruptly when a conductive filament

forms and then dissolves, producing a spike in a simple circuit. The difficulty has been that most

filamentary memristors switch above half a volt, several times the amplitude of a biological spike.

A University of Massachusetts Amherst group used protein nanowires harvested from the bacterium *Geobacter

A University of Massachusetts Amherst group used protein nanowires harvested from the bacterium

sulfurreducens* to build memristors that switch at about 60 millivolts and 1.7 nanoamps, with a

*Geobacter sulfurreducens* to build memristors that switch at about 60 millivolts and 1.7 nanoamps,

spiking energy of about 37 picojoules at a typical membrane capacitance, within the range the

with a spiking energy of about 37 picojoules at a typical membrane capacitance, within the range the

authors cite for real neurons.[^fu2025] A Northwestern team instead printed graphene and

authors cite for real neurons.[^fu2025] A Northwestern team instead printed graphene and molybdenum

molybdenum disulfide nanosheet inks onto flexible substrates, producing memristive neurons that

disulfide nanosheet inks onto flexible substrates, producing memristive neurons that fire single

fire single spikes, tonic trains and bursts at frequencies up to 20 kHz.[^hadke2026]

spikes, tonic trains and bursts at frequencies up to 20 kHz.[^hadke2026]

**Organic electrochemical neurons** are built from transistors whose polymer channel conducts both

ions and electrons, so ions in the surrounding fluid alter the firing directly. Linköping University

ions and electrons, so ions in the surrounding fluid alter the firing directly. Researchers at

researchers first made fully printed versions,[^harikesh2022] then built a version whose firing approaches 100 Hz and

Linköping University first made fully printed versions,[^harikesh2022] then a version whose firing

can be modulated by sodium, glutamine or neurotransmitters.[^harikesh2023] A Max Planck Institute

approaches 100 Hz and can be modulated by sodium, glutamine or neurotransmitters.[^harikesh2023] A

for Polymer Research group built a two-transistor organic neuron that operates in liquid and changes

Max Planck Institute for Polymer Research group built a two-transistor organic neuron that operates

its excitability in response to dopamine or small shifts in ion concentration; placed behind a

in liquid and changes its excitability in response to dopamine or small shifts in ion concentration.

layer of intestinal epithelial cells, its firing was suppressed by the layer's barrier to ion

Placed behind a layer of intestinal epithelial cells, its firing was suppressed by the layer's

flow.[^sarkar2022]

barrier to ion flow.[^sarkar2022]

## Development history

  { "year": "2020", "title": "Brain and silicon linked online", "text": "Memristive synapses connect a cultured rat neuron in Padova with silicon neurons in Zurich, routed through Southampton over the internet." },

  { "year": "2020", "title": "Brain and silicon linked online", "text": "Memristive synapses connect a cultured rat neuron in Padova with silicon neurons in Zurich, routed through Southampton over the internet." },

  { "year": "2022", "title": "Organic neurons meet biology", "text": "A printed organic neuron closes a Venus flytrap; another has its firing modulated in real time by a layer of epithelial cells." },

  { "year": "2022", "title": "Organic neurons meet biology", "text": "A printed organic neuron closes a Venus flytrap; another has its firing modulated in real time by a layer of epithelial cells." },

  { "year": "2023", "title": "Mouse nerve stimulated", "text": "An organic electrochemical neuron that senses ions and glutamine stimulates the vagus nerve of anaesthetized mice and changes heart rate." },

  { "year": "2023", "title": "Mouse nerve stimulated", "text": "An organic electrochemical neuron that senses ions and glutamine stimulates the vagus nerve of anaesthetized mice and changes heart rate." },

  { "year": "2025–2026", "title": "Biological amplitudes and timescales", "text": "Protein-nanowire neurons match biological voltage and energy; printed MoS2 neurons stimulate Purkinje neurons in mouse cerebellar slices." }

  { "year": "2025-2026", "title": "Biological amplitudes and timescales", "text": "Protein-nanowire neurons match biological voltage and energy; printed MoS2 neurons stimulate Purkinje neurons in mouse cerebellar slices." }

]

```

Each claimed connection to living tissue is narrower than its headline.

**Into tissue.** The printed Northwestern neurons stimulated Purkinje neurons in slices of mouse

cerebellum, work done with Indira Raman's neurobiology laboratory.[^hadke2026] The Linköping

cerebellum, work done with Indira Raman's neurobiology laboratory.[^hadke2026] The Linköping organic

organic neuron, coupled to a nerve cuff, stimulated the right cervical vagus nerve of anaesthetized

neuron, coupled to a nerve cuff, stimulated the right cervical vagus nerve of anaesthetized mice,

mice, lowering heart rate by about 4.5%; its authors state explicitly that the demonstration does not amount to a

lowering heart rate by about 4.5%. Its authors state explicitly that the demonstration does not

new therapy.[^harikesh2023] The earlier printed organic neuron triggered closure of a Venus flytrap,

amount to a new therapy.[^harikesh2023] The earlier printed organic neuron triggered closure of a

a plant.[^harikesh2022]

Venus flytrap, a plant.[^harikesh2022]

**Out of tissue.** The UMass neuron processed signals from cardiomyocytes, heart muscle cells

derived from human embryonic stem cells and grown on a graphene sensor mesh. It stayed silent at the

tissue's baseline beating rate and fired once noradrenaline raised it, so it distinguished two

tissue's baseline beating rate and fired once noradrenaline raised the rate, so it could distinguish

physiological states. The signals passed through an amplifier first, and the cells were heart cells

two physiological states. The signals passed through an amplifier first, and the cells were heart

rather than neurons.[^fu2025]

cells rather than neurons.[^fu2025]

**Both directions.** The clearest two-way demonstration is a 2020 network in which memristive

synapses linked a single cultured rat hippocampal neuron to silicon spiking neurons, with the

connection strengthening or weakening according to firing rates. The three sites communicated over

the internet, and the biological side was read through a patch-clamp pipette, a method that does

the internet, and the biological side was read through a patch-clamp pipette, a method that does not

not scale beyond the laboratory bench.[^serb2020] A 2019 system took the same idea toward repair:

scale beyond the laboratory bench.[^serb2020] A 2019 system applied the same idea to repair.

dissociated neurons were grown as two connected modules, the link between them was cut with a

Dissociated neurons were grown as two connected modules and the link between them was cut with a

laser, and a field-programmable gate array running a spiking network restored bidirectional

laser; a field-programmable gate array running a spiking network then restored bidirectional

activity between them or stood in for the missing half.[^buccelli2019]

activity between the modules or stood in for the missing half.[^buccelli2019]

Coverage of the April 2026 Northwestern paper ran under the headline that artificial neurons

Coverage of the April 2026 Northwestern paper ran under the headline that artificial neurons "talk"

"talk" to real brain cells, "paving way to better brain implants". The same Live Science report

to real brain cells, "paving way to better brain implants". The same Live Science report quoted

quoted Timothée Levi, a bioelectronics researcher at the University of Bordeaux who was not

Timothée Levi, a bioelectronics researcher at the University of Bordeaux who was not involved, as

involved, as saying such devices can control living neurons "for a short time but not yet for a long

saying such devices can control living neurons "for a short time but not yet for a long time", and

time", and so are not fit to be permanent additions to a human brain.[^guenot2026]

so are not fit to be permanent additions to a human brain.[^guenot2026]

> [!caution] What "communicating with brain cells" has meant

> None of the studies here implanted an artificial neuron chronically in any animal, let alone a

> person. The brain-tissue results are acute slice and culture experiments, the in vivo result is an

> acute experiment in anaesthetized mice, and several "bio-integration"

> acute experiment in anaesthetized mice, and several "bio-integration" demonstrations used heart

> demonstrations used heart cells, epithelial membranes or a plant. Durability is also unmeasured at

> cells, epithelial membranes or a plant. Durability is also unmeasured at implant scale: the

> implant scale: the printed MoS2 neurons were stable for more than a million cycles, which at ten

> printed MoS2 neurons were stable for more than a million cycles, which at ten spikes a second is

> spikes a second is roughly a day of firing.

> roughly a day of firing.

## Limitations

The hard problems are the ones that limit every chronic [[brain-computer-interface]], plus some of

their own. Any implant provokes a foreign-body response that encapsulates electrodes over months, as

documented for the [[utah-array]]; the cited artificial-neuron papers do not report tissue

documented for the [[utah-array]]. The cited artificial-neuron papers do not report tissue responses

responses because none was implanted for longer than an acute experiment. Filament formation is

because none was implanted for longer than an acute experiment. Filament formation is partly random,

partly random, so firing varies from device to device and from cycle to cycle. And the coupling itself still runs

so firing varies from device to device and from cycle to cycle. The coupling itself still runs

through conventional electrodes, pipettes or cuffs, so an artificial neuron does not remove the

through conventional electrodes, pipettes or cuffs, so an artificial neuron leaves the interface

interface problem; it changes what sits behind it.

problem in place and changes only what sits behind it.

A single neuron is also not a circuit. Hersam, the senior Northwestern author, described the

"frontier problem" as integrating devices that mimic separate elements of the brain into circuits

with full functionality.[^guenot2026] The brain-slice and nerve results each test one device driving

tissue, not a network of artificial neurons maintaining a computation with a network of real ones.

tissue. None tests a network of artificial neurons maintaining a computation with a network of real

ones.

## Risks

Because no device has been implanted, the risks are prospective. They are the risks of any

penetrating neural implant: surgical injury, infection, and the long-term consequence of

penetrating neural implant: surgical injury, infection, and the long-term consequence of stimulation

stimulation that is not identical to natural input, which in [[deep-brain-stimulation]] includes

that is not identical to natural input, which in [[deep-brain-stimulation]] includes mood and

mood and behavioural side effects. The UMass design depends on silver filaments forming and

behavioural side effects. The UMass design depends on silver filaments forming and dissolving, and

dissolving, and none of the cited studies tests what such materials do in tissue over months. A device that

none of the cited studies tests what such materials do in tissue over months. A device that adapts

adapts its firing to local chemistry is also harder to audit than a fixed stimulator, which bears on

its firing to local chemistry is also harder to audit than a fixed stimulator. That bears on the

the questions raised under [[mental-privacy]] and [[neurorights]] about who can read or alter neural

questions raised under [[mental-privacy]] and [[neurorights]] about who can read or alter neural

activity.

## Outlook

Proponents argue that artificial neurons could make prostheses that speak the nervous system's own

language: a [[retinal-implant]] or [[cochlear-implant]] whose output is shaped like natural spikes, a

language: a [[retinal-implant]] or [[cochlear-implant]] whose output is shaped like natural spikes,

[[memory-prosthesis]] that replaces a damaged circuit with a matched one, or a closed-loop

a [[memory-prosthesis]] that replaces a damaged circuit with a matched one, or a closed-loop

stimulator for the vagus nerve that responds to the body's chemistry. Some, as Live Science

stimulator for the vagus nerve that responds to the body's chemistry. Some go further, Live Science

reported, go further and suggest such devices could replace neurons lost in diseases like

reported, and suggest such devices could replace neurons lost in diseases like

[[alzheimers-disease]].[^guenot2026] None of these has been tested.

[[alzheimers-disease]].[^guenot2026] None of these applications has been tested.

The nearer competitors are not standing still. Conventional electrodes driven by digital

The nearer competitors are also advancing. Conventional electrodes driven by digital electronics,

electronics, refined by better [[neural-decoding]], already restore function in small numbers of

refined by better [[neural-decoding]], already restore function in small numbers of people. In

people; [[optogenetics]] achieves cell-type specificity in animals; and wireless miniaturization of

animals, [[optogenetics]] achieves cell-type specificity, and wireless miniaturization of the kind

the kind pursued under [[neural-dust]] attacks the tethering problem directly. A further line uses

pursued under [[neural-dust]] attacks the tethering problem directly. A further line uses living

living neurons rather than electronic ones, as in [[organoids]] and biohybrid circuits.

neurons rather than electronic ones, as in [[organoids]] and biohybrid circuits.

The longest-range claims connect bio-interfacing neurons to piecemeal neuron replacement as a path

toward [[mind-uploading]] and [[substrate-independence]]. That argument depends on electronic

toward [[mind-uploading]] and [[substrate-independence]]. That argument requires electronic neurons

neurons reproducing not just spike timing but plasticity, neuromodulation and the full repertoire of

to reproduce plasticity, neuromodulation and the full repertoire of a cell's chemistry as well as

a cell's chemistry, over decades, in place. The experimental record reaches spike timing in a dish.

spike timing, over decades, in place. The experimental record so far reaches spike timing in a dish.

## See also