A neuromorphic biocompiler writes analog circuits in DNA

What the work claims

This is a primary experimental result with a machine-learning method at its core, led by Jean Disset and Charles Van De Mark (co-first authors) with Georg Wachter, Shuo-Hsiu Kuo, Andrew Moorman, Wenlong Xu, Eric Palanques Tost, Andres Buritica Monroy, Calin Belta, and Ron Weiss, from MIT's Department of Biological Engineering with collaborators at Boston University and the University of Maryland.¹ The claim has two parts. First, mammalian cells can be programmed the way neuromorphic hardware is programmed: a small set of molecular primitives implements weighted sums, signed inputs, and nonlinear activation, and those primitives compose into multi-layer analog circuits. Second, and more striking, the inverse problem is solvable: given a desired input-output surface, gradient-based software can design the circuit automatically, and the design works in cells without a single manual tweak.

The evidence for the second claim is three target patterns specified as grayscale maps over two input concentrations. The biocompiler, searching within generic scaffold architectures, returned three distinct circuits: two dual-node and one single-node, none present in its training data. All three were built, transfected into HEK293 cells, and reproduced their target patterns, with model-to-measurement agreement of 0.076, 0.054, and 0.057 gRMSE (a grid-based error metric in rescaled 0 to 1 units, where differences between independent biological repeats of the same circuit average about 0.05). Two of the three patterns showed roughly a 15 percent shift along the x axis, which the authors flag rather than hide.

How it works

The molecular engine is a family of RNA-targeting endoribonucleases, or ERNs: PgU, CasE, and Csy4, drawn from the Cas6 and Cas13 families. Each ERN binds a recognition site on a target messenger RNA and triggers its degradation via 5' cap removal. That single mechanism supplies the two hardest neuromorphic primitives at once. Co-expression of converging elements supplies positive weighted inputs; ERN-mediated removal supplies negative inputs; and because removal rate depends on ERN abundance in a graded, saturating way, the same physical process also provides the nonlinear activation function. Weights are set by promoter strength, by upstream open reading frames (uORFs) in the 5' untranslated region that attenuate translation across nearly two orders of magnitude, and by co-transfection ratios. Surviving RNA output of one node can encode an ERN that cleaves the next node's target, which is how layers compose.

The modeling engine is what makes design possible. A biomorphic neural network, or BMN, represents each biological process (plasmid source, transcription, translation, ERN cleavage, fluorescence output) as a small shared neural block learned from data, with discrete genetic parts embedded in a continuous latent space. Crucially, the blocks are never measured directly: they are inferred jointly from whole-circuit fluorescence measurements, and because the same blocks recur across circuits, every new experiment constrains every process that circuit reuses. Trained on just three corners of an 81-condition uORF tuning matrix plus basic single-part measurements, the model predicted the remaining 78 conditions with a grid RMSE of 0.02, below the roughly 0.05 repeat-to-repeat variability measured across 24 circuits run more than once. For design, the blocks are frozen and the search runs in reverse: gradient descent over part embeddings and transfection ratios, with differentiable pruning to select topology, followed by quantization to real parts. One showcase design took over 760 optimization steps from a densely connected scaffold to the final compact circuit.

Two generalization results matter for reading the strength of the claim. Across topologies, held-out predictions were at or below biological repeat variability for six of eight composition classes; the two serial cascades were qualitatively right but quantitatively worse. Across cell types, a model trained only on HEK293 predicted circuits in ARPE-19 and CHO cells with median grid RMSE of 0.031 and 0.044 respectively, against 0.045 for HEK293 itself, from twelve test circuits spanning all topology classes.

Where a skeptic should push

The single most load-bearing assumption is that learned process blocks transfer to designs outside the training distribution because biology reuses processes, not because the model has learned physics. The evidence supports this within the paper's own design space: unseen uORF combinations, unseen multi-node topologies, unseen cell lines. But everything was measured in the same controlled, transiently transfected, flow-cytometry regime, with co-transfected fluorescent markers as stand-ins for inputs. The authors are direct about the gaps: resource competition and ribosome loading are captured only implicitly; design is confined to supplied scaffold architectures; the recipe-to-graph grammar is hand-authored; and the circuits run in transient transfection with exogenous inputs, not as stable genomic programs reading endogenous signals.

Second, scale honestly. These are one- to two-layer circuits with two inputs and a fluorescent protein readout, evaluated as population-level dose-response surfaces from single cells pooled by flow cytometry. That is real analog computation, but it is a long way from anything anyone would call intelligence: the dynamic ranges (10-fold to 30-fold) are respectable for gene circuits and tiny by computing standards, and the readout latency is days, gated at 60 hours post-transfection. Third, note the incentive structure: three of the authors, including the senior author, are co-founders of an early-stage company built around this technology, disclosed in the preprint. None of this undermines the data, which is unusually well controlled with repeat-variability baselines, but it calibrates how much of the roadmap is demonstrated versus promised.

What compiling cells means for organoid intelligence

The non-obvious implication is not that gene circuits compete with neurons. It is that this paper quietly prototypes the toolchain organoid intelligence has been missing. The field's core frustration is symmetrical to the one this paper solves: you can grow a rich analog substrate, but you cannot yet program it, because you cannot predict how a chosen intervention will compose, and the design space grows faster than experiments can cover. The biocompiler's doctrine is the transferable part: separate what you know (the graph structure of biological information flow, written as declarative rewriting rules) from what you don't (the quantitative behavior of each process), learn the unknowns jointly as reusable blocks from whole-system measurements, and then search the learned model in reverse. Every organoid lab faces exactly this epistemic split when it stimulates tissue and records population activity it cannot fully instrument. A BMN-style model trained on a corpus of stimulation-recording experiments across circuit configurations would let each new experiment constrain every process it shares with past ones, turning organoid characterization from per-device curve fitting into cumulative, composable knowledge.

And the funding trail confirms the intent, not just the analogy. This work was supported in part by NSF award 2422282, the PROGENIC program (Programmable Organoid Intelligence using Neuronal Networks implemented by Gene Circuits), a roughly $2 million EFRI grant to the same MIT group, explicitly aimed at teaching non-brain organoids to compute by installing genetically encoded neural networks in their cells, with microrobots delivering the programming signals and a neuroethics co-investigator on the award.² Read that against this preprint and the roadmap is legible: circuits like these are the intended weight-and-activation substrate, organoids are the intended host, and the biocompiler is the intended programming interface.

The threat is substrate competition from below. If non-neural cells can be compiled into analog computers with defined I/O and single-pass design, they offer everything organoids offer for low-dimensional analog computation (native biochemistry, self-maintenance, energy thrift) without the organoid field's two hardest problems: uncontrolled developmental variability and the lack of genetic handles on computation. A liver organoid that senses toxins through its own receptors and computes a graded response through engineered circuits, as PROGENIC proposes, is a biocomputer with an immediate application, while cortical organoid intelligence still struggles to define its first one. The dual-use and governance shadow is real and acknowledged in the funding itself: programmable analog decision-making in mammalian cells is one adjacency away from programmable cell behavior in therapeutic or environmental contexts, and the consciousness question is raised by the funders, not just the ethicists. The honest bottom line for organoid intelligence is that this work raises the bar for what counts as a programming interface for living matter, and right now, synthetic gene circuits have a compiler and neural tissue does not.

The bottom line

Established: composable multi-layer analog gene circuits in mammalian cells built from three orthogonal ERNs; learned process models that predict unseen weight combinations, unseen topologies, and unseen cell lines at or near biological repeat variability; and one-pass generative design that produced three working circuits with no manual tuning. Not established: stable genomic implementation, endogenous inputs, scalability beyond two-input one-to-two-layer circuits, or behavior under the resource competition of real tissues. The claim is confirmed if the biocompiler designs circuits that read endogenous molecular signals in stable cell lines, and broken if prediction error degrades sharply as circuits densify and shared-resource coupling strengthens. For organoid intelligence, the durable import is the doctrine: specify structure, learn process, search in reverse. That is the shape a programming interface for living tissue will have to take, whoever builds it first.