Programming a living cell means working with regulatory machinery whose responses are continuous, variable and tightly coupled to the rest of the cell. Designers can assemble DNA parts into a circuit, but predicting what several parts will do together is difficult. A September 2026 preprint proposes a different engineering workflow: tell a compiler which behaviour you want, and have it search for a regulatory circuit likely to produce it.
The study, Generative Neuromorphic Programming of Mammalian Cells, was led by Jean Disset and colleagues in Ron Weiss's research group. It appeared on bioRxiv on 22 September. It has not passed peer review as of this site's 24 October update.
The components are RNA-based regulators
The group's circuit elements use RNA-targeting endoribonucleases (ERNs). They allow regulatory interactions to have positive or negative effects and to produce nonlinear relationships between input concentration and cellular output. Combining these elements permits multi-input, multilayer analogue behaviours.
The compiler does not translate an arbitrary Python programme into DNA. It works inside the chosen library of genetic parts and the behaviour those parts have displayed in the researchers' experimental system. The proposed output is a candidate biological circuit architecture, ready for construction and measurement.
How the design loop works
- Measure parts and circuits. The team collects experimental responses from genetic regulatory combinations.
- Learn reusable processes. Biomorphic neural networks model transcription, translation and RNA cleavage as components whose behaviour can be recombined.
- Specify a target response. A desired analogue relationship between input and output becomes the objective.
- Search the design space. The biocompiler chooses topology, regulatory parts and weights predicted to approximate the requested function.
- Build and test. The proposed DNA circuit is constructed in cells, then its actual response is compared with the target.
This is a substantial use of conventional AI in biological engineering. The learning occurs in a digital model, while the resulting computation is executed through gene regulation inside living cells.
The experimental result
The authors challenged the system with three target behaviours. They report that the biocompiler proposed three architectures absent from its training set, and that all three reached their targets after construction in mammalian cells without a manual circuit redesign. The paper also compares the model's predictions with biological variability across experimental repeats.
One-pass success on three designs is encouraging because manual iteration can dominate synthetic-biology projects. It is not yet a general reliability estimate. Validation on larger design sets, different cellular contexts and independently run experiments will be essential.
What could come next?
Reliable automated design would help build cellular sensors, feedback responses and more complex living materials. Some applications might tolerate the slower time scale of gene expression if the system can operate directly in a chemical or biological environment.
There are engineering limits: resource competition, cellular evolution, unwanted interactions, context-dependent gene expression and distribution of responses across cells. Safety and containment requirements also depend strongly on the intended application. The present preprint demonstrates a design method inside a bounded set of parts. It does not solve those general constraints.
The parallel 2026 printed-bacteria study is useful for contrast. It reuses five pre-engineered strains and changes their arrangement. This biocompiler generates different genetic architectures within mammalian cells. They may eventually complement one another, but they are not competing benchmarks of processor speed.