The Nasal Spray That Wants to Be a Brain Implant

[Written by Claude. Image credit.]

Nanoparticle brain-computer interfaces: how they work, why they’re designed to disappear, what a self-assembling implant would require, and the one thing none of this can do.


Every brain-computer interface company is answering the same question: how do you get close enough to neurons to hear them clearly?

The electrical activity of a single neuron is tiny, and skull and scalp are excellent at smearing it into mush. That physical fact split the field into camps. Neuralink drills through the skull and places electrodes directly in cortex, buying superb signal at the cost of neurosurgery. EEG stays entirely outside, available to anyone, and reads a blurred average of millions of cells at once. Synchron threads the needle by going through a blood vessel.

Subsense, a Palo Alto startup that has raised $27 million, proposes something stranger: skip the surgery, but still put the sensor inside the brain. Not a chip — nanoparticles, delivered through the nose.

Their CEO, Tetiana Aleksandrova, calls the approach “invasive but non-surgical.” It is a useful phrase, and an honest one. Nothing is cut open, but material does end up in your brain.

They are not alone in the underlying wager. A whole cohort — including Sam Altman’s $252 million Merge Labs — is betting you can put a reporter inside the brain and interrogate it from outside the skull. Subsense is distinguished by what it puts there and how it talks to it: injected nanoparticles, addressed by magnetic fields and light, rather than genetically expressed proteins addressed by ultrasound.


How it works

The system, which the company calls NanoBCI, has three parts.

Delivery. Charged nanoparticles are inhaled, roughly the way you’d take an allergy spray. They travel along the olfactory nerve — the sensory nerve for smell — into the olfactory bulb at the front underside of the brain. From there they are steered magnetically toward their targets.

Reading. The sensing particles are gold cores coated in an electrochromic polymer, meaning their optical properties change in an electric field. When a nearby neuron fires, the particle scatters light differently. An external headset shines near-infrared light through the skull and watches for that shift. The particle draws nothing from the neuron and doesn’t interfere with it; it just changes how it looks when the neuron is active.

Writing. The stimulating particles are a different design — under 100 nanometers, with a magnetostrictive core and a piezoelectric shell. The headset generates a magnetic field, the core deforms in response, the deformation squeezes the piezoelectric shell, and the shell produces a local electric field. Each particle becomes a tiny wirelessly powered electrode.

It’s worth noting the asymmetry here, because it shapes everything downstream. Magnetism does two of the three jobs — steering and stimulation — and magnetic fields pass through skull and tissue essentially unattenuated. Light does only the reading, and light does not. Subsense chose different physics for each direction, and the costs land unevenly.

Both particle types are designed to be temporary, tuned to last somewhere in the range of 20 to 60 days before clearing.


Why they can’t just stay

That last detail surprises people. If you’ve gone to the trouble of placing sensors on neurons, why not leave them there? Two answers: the brain actively removes them, and even if it didn’t, staying wouldn’t preserve function.

Extracellular space is not a shelf. An unanchored particle is subject to Brownian motion and bulk interstitial flow, so it drifts. Beyond drift, two systems specifically target foreign material. Microglia — resident immune cells that continuously survey tissue — engulf particles, and microglial extracellular vesicles play a central role in clearing both organic and inorganic nanoparticles from the brain. Separately, the paravascular glymphatic pathway carries particles out in bulk. You would be asking an object to resist machinery evolved precisely to flush things like it.

Phagocytosis doesn’t merely relocate a particle, it disables it. A particle swallowed by a microglial cell sits in an acidic lysosome, not on a neuronal membrane. It’s still in the brain and would still appear on a biodistribution scan, but it’s now reporting on the inside of an immune cell. Retention and function run on different clocks.

Chemistry degrades too. Sensing depends on an intact electrochromic coating; stimulation on an intact core-shell interface. Coatings strip, protein coronas form, particles aggregate. Even a perfectly immobilized particle loses its properties.

And permanence would be a liability rather than a feature. Gold is not biodegradable — 150 nm gold nanoshells have shown no detectable reduction in total body mass more than a year after injection. Permanent particles mean permanent exposure, no way to stop treatment, no way to upgrade, and an indefinite stimulus for chronic glial activation. Regulators are vastly more comfortable with something the body eliminates. Aleksandrova frames this as deliberate: a permanent presence was never the goal, and clearance is treated as tuneable.

The tension is real, though. Re-dosing every 20 to 60 days means repeated nasal exposure and cumulative body load over years. The alternative is probably worse. There’s no clean resolution, which is why the clearance characterization the company admits is still outstanding matters so much.


Electrical versus chemical: two routes to a non-surgical interface

Subsense’s particles are transducers. They convert energy between forms — magnetism to electricity, electric field to optical signature — and release nothing into tissue.

A parallel research tradition does release things. Chemical neuromodulation loads neuroactive compounds into liposomes or polymer-coated magnetic particles and triggers release on demand. The most developed version uses heat: magnetic particles dissipate energy in an alternating field, and a heat-sensitive linker lets go of its payload. MIT researchers used this to trigger action potentials with latencies around twelve seconds. Others have released dopamine and chlorpromazine from coated particles to excite and inhibit striatal neurons.

Electrical (magnetoelectric)Chemical (triggered release)
SpeedEffectively instant — tracks the fieldSeconds
SpecificityWhatever neurons are nearbyCell-type specific via native receptors
SustainabilityIndefinite; nothing is consumedFinite payload
Genetic modificationNot requiredNot required

Chemical release has the better specificity story, because it works through receptors a cell actually expresses rather than depolarizing a neighborhood indiscriminately. But a payload runs out, and seconds-scale latency is hopeless for a communication channel.

That is almost certainly why Subsense chose transduction. You cannot build a real-time interface on a particle that runs out of dopamine. Chemical approaches are better understood as a route to wireless, non-surgical therapy — a replacement for DBS hardware — than as a route to a BCI.


Where the physics says this goes

Writing deep is plausible. Reading deep is not. This follows directly from the modality split. Magnetic fields reach anywhere, so depth doesn’t limit stimulation, and magnetoelectric particles have already produced deep-brain effects in animals: stimulation of the subthalamic nucleus, the standard Parkinson’s target, produced effects similar to conventional DBS.

The read channel has no such luck. Near-infrared photons must make a round trip through bone and brain, and scattering is brutal. Diffuse optics struggles to collect photons from beyond about 4 cm below the scalp — roughly the outer layer of cortex. Cortical folds, midbrain, deep cerebellum are out of reach. Nanoparticles brighten the signal where light already goes; they cannot change the scattering budget.

So the realistic architecture is bidirectional at the cortical surface, write-only deeper. This creates a specific problem for the company’s own pitch. Aleksandrova describes therapy that reacts to “the state a person is actually in rather than the state they were in at their last appointment” — but a closed loop requires reading where you write, and the two channels don’t currently overlap in the deep structures that Parkinson’s and epilepsy therapy target. You could drive the subthalamic nucleus without observing it.

Single-neuron sensitivity is not single-neuron resolution. The underlying optics, licensed from UC Santa Cruz, has an impressive in-vitro spec: signal-to-noise above 1,000, claimed single-neuron sensitivity, optical signal four orders of magnitude above quantum dots. All measured in a dish with a microscope pointed at it.

A particle on a membrane does experience one neuron’s field. Reading that particle separately through a skull is a different problem. Diffuse optical resolution is around a centimeter; a cubic centimeter of cortex holds tens of millions of neurons. Unless particles are sparse enough that only a handful land per resolvable volume — which costs coverage — the headset sees a blend.

The same logic applies to stimulation. Magnetic coils focus at centimeter scale at best, so selectivity comes from where particles sit, not where the field points. Precision is a delivery problem, not a hardware problem.

For scale: subdural arrays at 65,536 electrodes with sub-millimeter pitch have been reported in animals, though so far in preprint. And even if the optics worked, thousands of sensors at kilohertz sampling implies multi-megabit-per-second egress off the head. Channel count isn’t the only thing that has to scale.


Who else is betting on this

Subsense may be the only company commercially pursuing nanoparticles, but most of the money chasing the same wager has gone to ultrasound.

Merge Labs is the closest cousin and far better capitalized — though less similar than it first appears. Altman’s company emerged from stealth in January 2026 with a $252 million seed at an $850 million valuation, OpenAI writing the largest check. The approach pairs functional ultrasound imaging with engineered molecular reporters — proteins that make neural activity easier to detect through the skull — descending from Mikhail Shapiro’s Caltech work on acoustic reporter genes, which cause cells to form gas vesicles that scatter ultrasound.

Two differences from Subsense matter more than the modality. First, the reporters are genetically encoded. Merge hasn’t disclosed delivery, but the widely discussed hypothesis is AAV gene therapy — meaning neurons are permanently modified to produce the reporters themselves. Second, Merge descends from Forest Neurotech, whose ultrasound device sits beneath the skull and is in a UK NHS safety trial with 30 patients planned. So despite the non-surgical framing in coverage, Merge may be the more invasive approach on both axes: irreversible genetic modification plus an implant.

To its credit, Merge is unusually blunt about timing, with coverage reporting the technology arrives in “decades rather than years.”

Focused ultrasound is the actual funding wave. Nudge reportedly raised $100 million. In China the category went from nonexistent to crowded in a year: Gestala, founded in Chengdu in January 2026, raised $21.6 million two months after launch — the largest early-stage round in China’s BCI industry — and roughly $84 million by mid-year. BCI-Sonics raised about $14 million, with Sisheng Technology and Neurorevix behind.

The nanoparticle science lives in academic labs. Ali Yanik’s group at UC Santa Cruz developed the plasmonic optics Subsense licenses. Polina Anikeeva’s lab at MIT drove much of the magnetothermal work. Both do the physics; neither is a company.

So the real contest is optical readout of injected particles versus ultrasound readout of genetically expressed proteins, and neither clearly wins. Ultrasound reaches deeper with better spatial resolution — Merge reports roughly 100 microns, cortical surface to deep brain center. But functional ultrasound is hemodynamic: it reads blood flow, on a timescale of seconds. The plasmonic approach reads electric fields at spiking speed. Subsense’s genuine advantages are temporal resolution and reversibility. Merge’s are depth, spatial precision, and about ten times the capital.

One structural caveat on Subsense’s side: the core optics is licensed from UC Santa Cruz rather than solely owned, so the moat is partly rented.


The interesting version: a self-forming implant

Here is where the idea gets genuinely compelling, and where the honest difficulties are most interesting.

If particles could self-organize — anchor to chosen neurons, connect to each other, build structure in place — you’d have something better described as a self-forming implant than as a drug. All the placement advantages of surgery, none of the craniotomy.

Part of this is already demonstrated. In 2020 the Deisseroth and Bao labs at Stanford genetically instructed specific living neurons to guide the synthesis of conductive or insulating polymers at their own plasma membranes, confirming by electrophysiology and behavior that neurons stayed viable while membrane properties were remodeled in freely moving animals. The mechanism is elegant: precursors diffuse through intact tissue, then polymerize at a targeting enzyme’s reactive center, and because radical cations have short diffusion lengths and the products are poorly soluble, material deposits exactly where the enzyme sits. The cell becomes its own assembly template.

Note the word doing the work: genetically. Specificity came from expressing an enzyme in chosen cells — which lands you in exactly the territory Merge occupies, and abandons the injectable, reversible premise that distinguishes Subsense.

And particles connecting to each other is much harder than connecting to cells. Extracellular fluid is a salty ionic conductor, so a chain of conductive particles leaks current along its whole length — this is why axons evolved myelin, and self-assembling an insulating sheath along a chosen path is a far larger ask than coating a membrane. Worse is power: these particles have no internal energy source by design, so an assembled network that routes or computes needs power at every node. And nanoscale radio is out, because antenna efficiency collapses when the antenna is far smaller than the wavelength.

Self-assembly also makes clearance dramatically worse. Structures large enough to be networks are too large for microglial or glymphatic removal. They’d be permanent — the liability established above. Aggregates are more immunogenic, can occlude capillaries, and present a standing stimulus for chronic gliosis.

So how could it actually work? The most coherent theoretical stack I can assemble has five layers.

Encode position in frequency, not space. MRI achieves sub-millimeter resolution through an intact skull because a field gradient makes each position resonate differently — spatial information lives in the frequency domain. Plasmonic resonance depends on particle geometry; magnetoelectric resonance on geometry and field. So you could manufacture populations with distinct spectral signatures, or apply a gradient so a particle’s response frequency reports its position. Readout becomes demultiplexing rather than imaging, and resolution is set by linewidth rather than photon scattering. This is my extrapolation, not a company claim. The limit is that resonance linewidths in warm, protein-coated environments are broad, and nanoparticle manufacturing tolerances are loose.

Build hubs, not networks. Rather than nanoparticles talking to the outside directly, have them aggregate into mesoscale relays. “Neural dust” supplies the precedent: millimeter-scale motes with a piezoelectric crystal, two electrodes and a transistor, where extracellular voltage modulates the transistor gate, changing current through the crystal and altering the reflected ultrasound — so the echo encodes the recorded voltage. Completely passive, no battery, no antenna. StimDust later ran at 2.2 mm³ on a single ultrasonic link carrying both power and bidirectional data, at one fifth the FDA acoustic limit. Recent work reports 0.8 mm³ implants at 50 mm depth. The original thesis argued the approach scales to hundreds, possibly tens, of microns.

Power and exit by ultrasonic backscatter. The hub layer solves both at once, and ultrasound penetrates where light cannot. The caveat is significant: propagating ultrasound through the skull is hard, which is why neural dust targeted peripheral nerves first.

Anchor by chemistry. Particles carrying a polymerization catalyst could nucleate structure only where they’ve bound their target, self-limiting by the same diffusion-length argument Stanford exploited.

Design the exit before the entrance. Build with bonds that break on command — pH shift, cleavable linker, specific field — releasing fragments below the renal clearance threshold of 5–10 nm. Simpler versions exist: gold formulations with pH-triggered decomposition have achieved over 85% liver elimination in three months. This restores explantability without surgery, which was the main thing a self-forming implant otherwise surrenders.

LayerStatus
Transduction at the membraneDemonstrated in vitro
Chemical anchoring to targetsDemonstrated, genetically gated
Spectral or gradient addressingSpeculative; physics plausible
Self-assembled mesoscale hubNot demonstrated at all
Ultrasonic power and backscatterDemonstrated in fabricated motes
Triggered disassemblyDemonstrated in simpler systems

Three of six have real precedent. One has none. And they come from entirely different labs, organisms, and material systems — integration is usually where stacks like this die.

It’s worth naming what you’d be trading. A surgical implant’s quiet advantage isn’t placement, it’s the wire: power in, data out, local amplification. Neuralink runs 1,024 electrodes because a chip sits right there. A self-formed implant without a wire still has to push signal through the skull. You’d have built the array and skipped the cable. You’d also lose explantability, the underrated safety feature of anything surgical, trading bounded procedural risk for unbounded permanent risk.

This is a spectrum, not a binary: open craniotomy, then endovascular, then injectable mesh, then self-assembled material, then free particles. Moving down, invasiveness falls and so do bandwidth, control, and reversibility. The question isn’t whether avoiding surgery is good — obviously it is — but whether what survives at the far end is still worth having.

My read: the hub-assembly step is the wall, and it’s a chemistry problem more than a neuroscience one. A crude but functioning in-vivo self-assembled backscatter element, even in a mouse, would substantially revise my skepticism. Nothing currently underway is aimed at that. Subsense is solving transduction, which was the layer least in doubt.


Can any of this replace dead neurons?

No. Not in real time, not eventually, not with any amount of refinement. The distinction gets blurred constantly in coverage of this field, so it’s worth stating precisely.

A dead neuron is gone. It took with it a cell body, a specific axonal projection, thousands of synapses with particular weights, and a position in a circuit tuned over years. A nanoparticle is a transducer in extracellular space. It can make a surviving neuron fire. It cannot synthesize dopamine, grow an axon, form a synapse, or occupy the role of a cell that no longer exists.

Stimulation compensates; it does not replace. Deep brain stimulation works by driving remaining circuitry harder — it manages symptoms and does nothing to stop neuron death. Parkinson’s motor symptoms typically aren’t detectable until 50 to 80% of dopaminergic neurons are already lost. A nanoparticle BCI aimed at Parkinson’s does DBS’s job by a gentler delivery route. Worthwhile, but not replacement.

The actual replacement technology exists, and it’s further along than most people realize — though earlier than headlines suggest. Researchers grow dopaminergic neurons from pluripotent or induced pluripotent stem cells and transplant them, where they graft into host tissue, sprout axons, and secrete dopamine. In March 2026 Japan granted conditional, time-limited approval to Amchepry, developed by Sumitomo Pharma with Kyoto University — the world’s first approved iPS-cell-derived therapy. The supporting trial was small: seven patients aged 50 to 69, receiving five or ten million cells implanted bilaterally, monitored two years with no major adverse effects and four showing symptom improvement. Full approval requires roughly seven more years of data across a target of 35 patients. A separate Phase I trial of human embryonic stem cell–derived neurons was published in Nature in 2025. BlueRock Therapeutics has a Phase 3 trial underway in about 100 people, and a USC trial of iPSC-derived neurons holds FDA fast-track designation.

Note that these are surgical implantations, and the field has its own hard problems — graft survival, immune rejection, a sobering history of failed fetal-tissue transplants in the 1990s, and the likelihood that non-motor symptoms won’t respond to dopamine cell replacement at all. But it is the branch actually attempting to put new cells where old ones died.

The two are complements. Cell therapy restores tissue, slowly and surgically. A BCI reads and writes signals, fast and non-invasively. A plausible future uses both. But if someone tells you a nanoparticle interface will replace your dying neurons, they are confused or selling something.

One honest caveat the other way: there’s a weaker sense in which a BCI replaces lost function. If a stroke destroys the pathway from motor cortex to your hand, a BCI that reads intent and drives a prosthetic has replaced the dead circuit’s function without replacing a cell. That’s real, it’s what BCIs are for, and it’s the claim worth getting excited about. It’s just a different claim from regenerating tissue.


What’s still unanswered

Subsense is preclinical, running mouse studies. Aleksandrova is reasonably candid: between mice and humans there’s work remaining in reproducibility, manufacturing, toxicology, biodistribution, clearance, and regulation. The questions I’d want answered, roughly in order of how much they’d move me:

Does the optics survive the skull? Every impressive number is in vitro. Transcranial signal-to-noise, real resolvable volume, and separation from hemodynamic and scalp artifacts are unknown.

How does the loop close? Stimulation reaches deep; reading doesn’t. Either there’s a cortical proxy strategy, or a second read modality is coming, or the closed-loop pitch has a hole in it.

Does magnetic steering scale? A human brain is orders of magnitude larger than a mouse’s and field gradients fall off with distance. Targeting precision at human scale hasn’t been published, nor has off-target fraction. Since selectivity comes from placement rather than field aiming, this is load-bearing.

What’s the long-term body burden? Inorganic particles commonly accumulate in liver and spleen for months to years. Renal clearance needs sub-5–10 nm; the stimulation particles are under 100 nm. Some formulations disassemble into excretable fragments. Whether these do isn’t public, and 20–60 days is a target, not a measurement.

What are the magnetoelectric particles made of? Magnetoelectric cores in the literature often involve cobalt ferrite, and cobalt neurotoxicity isn’t trivial. The formulation is undisclosed and drives much of the safety case.

What does the nasal route cost? Repeated dosing through olfactory epithelium carries its own risk, including anosmia. I haven’t seen it addressed.

Does it treat anything? The company aimed to show neural reading and symptom relief by end of 2026 — roughly now, making it the nearest checkable milestone. Safety assessment is slated for 2028, approval around 2031.

And the unglamorous one. $27 million is seed money, all of it from a co-founder’s own fund. Carrying a novel nanomaterial through toxicology, GMP manufacturing, and first-in-human trials plausibly costs a hundred times that. The company is advertising for a CFO.


The bottom line

The nanoparticle approach is a clever piece of physics aimed at a real gap. If it works, the likely near-term product is wireless deep-brain stimulation without an implant — a meaningful advance for Parkinson’s and epilepsy patients, and a genuinely gentler path than drilling the skull. Its distinctive virtues are speed and reversibility, and reversibility may prove the more valuable of the two in a field where every competitor is proposing something permanent.

The high-bandwidth, read-your-thoughts version runs into scattering physics that nanoparticles don’t solve. The self-forming implant is the most interesting idea in the space and the furthest from demonstration — not because any single step is impossible, but because it needs six of them to compose inside a body actively dismantling the work. And neuron replacement isn’t on this roadmap at all; that’s a cell therapy story, happening in a different building and moving faster.

The broader pattern is worth sitting with. Enormous capital is chasing non-surgical high-bandwidth BCI on the strength of physics not yet demonstrated in a human head by anyone, in any modality. The best-funded company in the category says decades. The honest version of the Subsense story says something similar.

Worth watching. Worth reading the fine print.

Leave a Reply