The organoid intelligence project that throws out the neurons

What the award proposes to build

Know the genre. This is a grant abstract, which is a statement of intent written to persuade reviewers, and nothing in it has been demonstrated.¹ It is a Continuing Grant under the Emerging Frontiers in Research and Innovation office, obligated at $1,107,816 in fiscal 2024 and $890,589 in fiscal 2025 against a total of $1,998,405, running to 31 August 2028. The principal investigator is Ron Weiss, a founding figure in synthetic biology. The co-investigators tell you as much as the abstract does: Calin Belta, a control theorist, Sambeeta Das, who builds microscale robots, and Laurie Zoloth, a bioethicist. It sits in the EFRI BEGIN OI programme, which stands for Biocomputing through EnGINeering Organoid Intelligence, from solicitation NSF 24-508.

The premise is an argument by omission. Biocomputing has fixated on neurons because the brain is the obvious biological computer, and in doing so it has ignored that every organ senses inputs, integrates them and produces responses. Livers make decisions constantly, about metabolism, detoxification and protein synthesis. So the proposal takes liver organoids as the substrate and asks whether intelligence can be engineered into them directly.

The mechanism is the interesting part. Each engineered cell is to carry a multi-node artificial neural network built from synthetic gene circuits operating on DNA, RNA and protein. Learning means changing the weights of that network, and the weights are to be changed by on-demand modification of the genetic circuits in the cell's chromosomes. The signal that triggers those edits comes from "designer guide cells" ferried to specific locations inside the three-dimensional organoid by magnetically controlled microrobots. Performance is then assessed by perturbing the organoid with insults such as viral infection and toxins and measuring whether organ function holds up, quantified as albumin and urea secretion. Three aims follow that structure: engineer the in-cell networks, build the microrobot delivery route, then integrate both into an organoid that adapts. The award also funds explicit work on whether consciousness or sentience could arise in the resulting network.

The strongest case for taking it seriously

The steelman is not that liver organoids will out-think cortical ones. It is that the proposal attacks a real and unacknowledged weakness in the neural approach. Computation in neural tissue is carried by synaptic weights, and synaptic weights are volatile, invisible and unaddressable. You cannot read the weight matrix of an organoid, you cannot write a chosen value into a chosen synapse, and whatever state a culture reaches decays and drifts. Every result in the field is therefore obtained by coaxing a system whose parameters are inaccessible.

Relocating the network from between cells to inside cells changes all three properties at once. A weight encoded in a genetic circuit has a physical address: a locus, a sequence, a copy number. It can in principle be read by sequencing and written by editing. Weiss is a reasonable person to attempt this, and the choice of a control theorist as co-investigator suggests the team knows that the hard part is not building one circuit but making a population of them behave as a specified dynamical system.

The choice of readout is also more considered than it first looks. Albumin and urea secretion are the standard functional assays for hepatocyte competence, so "did the organoid keep working after the insult" is a measurable, meaningful objective rather than an arbitrary benchmark. The system is being asked to maintain function under perturbation, which is a defensible operational definition of adaptive intelligence for a non-neural tissue.

Where a skeptic should push

Start with what has been delivered. Exactly one publication is reported under this award: a paper from Das's laboratory on closed-loop acoustic-magnetic propulsion of microrobots for precision cell manipulation.² It is real work, published in Advanced Intelligent Systems in 2025, and it concerns acoustically powered, magnetically steered microrobots demonstrated by moving mammalian cells. It contains no liver organoid, no gene circuit and no artificial neural network. Two years into a four-year award, the reported output covers the delivery vehicle for Aim 2 and none of the biocomputing claims. That is not misconduct and it is not unusual for a project of this shape, but anyone citing this award as evidence that non-neural organoid intelligence works is citing a plan.

The load-bearing assumption is that a weight written into chromosomal DNA is a stable weight. It is not, and the distinction between persistent and stable is where this design is most exposed. Mitotic heritability is genuinely bought: unlike a synapse, an edited locus is copied to daughter cells and does not decay on a synaptic time constant. Everything else pushes the other way. Integrated circuits are routinely silenced by promoter methylation and heterochromatin spread, so the sequence persists while the weight it encodes quietly goes to zero. In a dividing population, any weight configuration that imposes a metabolic burden is selected against, which means the stored value carries a survival bias that no synapse suffers. Recombinase and nuclease-based editing leaves scars, consumes a limited supply of orthogonal recognition sites, and is often not cleanly reversible, so rewriting degrades the substrate. And DNA edits are discrete, so graded weights must be approximated through copy number or promoter strength, both noisy at low copy number.

Worst of all for this particular architecture, the write mechanism guarantees heterogeneity. Guide cells delivered by microrobots to selected locations in a three-dimensional organoid produce, by construction, a spatially incomplete edit. The organoid-level weight is then a population average over a distribution the experimenter does not control and cannot easily measure. Mosaicism is not a risk here; it is the expected outcome of the delivery method.

Second, the claim that this escapes the interface bottleneck needs qualifying. It is true that no electrodes are required, and therefore that electrode count stops being the binding constraint. But albumin and urea secretion are bulk, population-averaged scalars accumulated over hours. Neural organoid intelligence is limited by how many channels you can read and write; this design is limited by how few numbers come out and how slowly. That is a trade between two observability problems, not an escape from the class, and for anything resembling computation the low-dimensional slow readout may well be the worse of the two.

Third, timescale, stated carefully because a flat version of this criticism is easy to rebut. For inference, the gap is enormous: a synaptic event takes about a millisecond, while a transcription and translation cycle reaches steady state over tens of minutes to hours, and an induced chromosomal edit takes longer still. That is roughly six orders of magnitude, and it means this substrate cannot compete on throughput or latency for anything. But the honest counterpoint is that biological learning is not fast either. Late-phase synaptic consolidation is itself transcription-dependent and unfolds over hours, so as a mechanism for writing durable memory, gene-circuit editing is closer to par than the raw comparison suggests. The right way to state the limitation is that PROGENIC targets adaptation and homeostatic decision-making, not computation in any throughput sense, and the word "intelligence" should be read accordingly.

What this means for organoid intelligence

The non-obvious implication is one the abstract never states, and it is the most important thing in the award. If weights live in chromosomal DNA, then a trained configuration is clonable.

Consider what that would fix. The deepest structural problem in organoid intelligence is not electrode count or culture longevity; it is that every organoid is a unique, non-transferable object. Two organoids from the same protocol differ in cell composition, connectivity and developmental trajectory. A result obtained on one cannot be reproduced on another except statistically, a trained tissue cannot be copied, and nothing can be shipped to another laboratory except a protocol and a hope. This is why silicon retains a decisive advantage in reproducibility, and why the field's demonstrations remain artisanal.

Genetic weights break that constraint in principle. A weight matrix encoded in chromosomes is a cell line: expandable, freezable, bankable, shippable and re-instantiable in another laboratory next year. Training would become a one-time cost amortized across every organoid subsequently grown from the bank, and biological computing would move from artisanal to manufacturable. No neural organoid approach offers anything comparable, because you cannot clone a synapse population. I have not seen this argued in the organoid intelligence literature, and it is a stronger reason to watch this project than anything in its own framing.

The same analysis identifies exactly what would kill it, which is why the argument is worth making precisely. Clonability requires the weight to survive copying, and the failure modes listed above are precisely failures of copying: epigenetic silencing changes the weight without changing the sequence, mosaic writing means there is no single weight to copy, and clonal selection means the copies that thrive are the ones whose weights cost least to carry. Whether PROGENIC delivers the biggest prize in the design therefore turns on questions of epigenetic stability and edit uniformity that its abstract does not mention at all.

There is a threat here for the neural organoid community, and it is a threat to identity rather than to funding. The field's implicit claim is that neurons are special, that their plasticity is what makes tissue computational. A funded, plausible programme to make liver cells learn says the interesting property may be regulatable state rather than neural identity. If it succeeds even partially, the category "organoid intelligence" stops meaning "brain organoids" and starts meaning any tissue with an engineerable internal state, and the neural approach must justify itself on performance rather than on the assumption that it is the only game.

This award also clarifies something about write channels generally, once set beside the rest of the field. There are three ways into living tissue and each trades addressability against rate. Electrodes are fast and precise but confined largely to a surface and limited by channel count. Fields are fast and reach everywhere but address nothing. Microrobots are the only option that is genuinely addressable in three dimensions, and they are extraordinarily slow. PROGENIC chooses the third, which explains why its one publication so far is the microrobot paper. Aim 2 is not a supporting detail; it is the load-bearing element, and the team appears to know it.

On ethics the award does something valuable and slightly misaimed. Funding a bioethicist to ask whether a de novo network could be sentient is right, but the framing inherits the word "neuronal" from a design in which nothing is a neuron. The sharper question the project actually raises is whether the functional criteria we use for moral status could be satisfied by non-neural tissue at all, and a neuron-framed ethics aim is structurally unable to ask it. That question will outlive this grant regardless of whether the science works.

The bottom line

Nothing here is an established result. What exists is a funded four-year plan, a credible team, and one published paper on the microrobot delivery hardware that touches none of the central claims. Treat every capability described above as hypothesis.

The hypothesis worth tracking is narrow: that a multi-node computational function can be encoded in synthetic gene circuits inside hepatocytes, that its parameters can be rewritten on demand in situ, and that the resulting organoid measurably improves its functional resilience to a toxin or infection it has encountered before. Improved albumin and urea retention after a repeated insult, against properly matched unedited controls, would be the first real evidence. Everything else in the abstract is scaffolding around that claim.

What would break it: evidence that edited weights drift or silence over passages faster than they can be written, or that microrobot-delivered editing cannot achieve uniform enough coverage for organoid-level behaviour to be reproducible. Either finding would leave the delivery technology intact and useful while removing the reason to care about the computing claim. For readers tracking this programme, the neural counterpart award analysed in an earlier piece in this stream makes a useful contrast: same funding call, opposite bet on what the substrate should be.