Cells compute through densely interconnected regulatory networks that produce graded, analog responses. The authors describe a generative neuromorphic programming framework that translates desired analog behaviors into DNA constructs for mammalian cells. The work aims to overcome limitations of common synthetic biology abstractions, which the authors argue poorly match the analog and compositional nature of cellular regulation. By combining composable biological devices, learned models that mirror biological processes, and an inverse design compiler, the approach seeks to enable systematic design of multi-input, multi-layer analog circuits in living cells.
Central to the platform are modular regulatory devices constructed from RNA-targeting endoribonucleases (ERNs). These devices are designed to supply both signed weights and nonlinear activation functions within genetic circuits, enabling layered architectures analogous to artificial neural networks. The ERN-based devices serve as composable parts that can be combined into circuits with diverse analog responses to multiple inputs. The authors present these devices as an enabling technology for building multi-layer circuits that operate with graded outputs rather than purely digital on/off behavior.
To predict the behavior of circuits assembled from these devices, the team developed biomorphic neural networks (BMNs). BMNs are a compositional modeling architecture that explicitly mirrors biological interactions: each elementary biological process—such as transcription, translation, and endoribonuclease-mediated cleavage—is represented as a reusable neural block. These blocks are trained from measurements of whole circuits that contain the process of interest, so the learned component captures the effective behavior of the process within circuit contexts.
Once learned, blocks can be recomposed into circuit architectures that were not present during training. According to the authors, BMNs often predict circuit responses with errors that are comparable to the variability seen between experimental repeats, indicating that the models generalize well to new combinations of parts and topologies.
The modeling approach is inverted in a software package termed the biocompiler, which jointly optimizes circuit topology, choice of parts, and quantitative weights to realize specified target behaviors. Given a desired analog response, the biocompiler searches across architectures and component assignments to propose DNA designs predicted to reproduce the target behavior in cells. The authors emphasize that the biocompiler performs a generative design role: it proposes previously unseen architectures rather than limited permutations of pre-existing designs.
To validate the generative design workflow, the authors challenged the biocompiler with three distinct target behaviors. For each target, the software proposed a novel architecture that had not been used in training. The team built all three proposed designs and tested them in mammalian cells. Each constructed circuit reproduced its respective target behavior in a single design pass without manual tuning, demonstrating the end-to-end capability from specification to experimental realization.
The reported outcomes support the claim that composable ERN devices, BMN-based prediction, and biocompiler-driven design can produce functional multi-input analog circuits directly from computational design.
This work frames cellular programming as an analog computation problem and provides an integrated route to design such computations. By combining modular ERN-based devices that encode weights and nonlinearities with biomorphic models that capture process-level behavior, the framework reduces reliance on ad hoc manual tuning and exhaustive experimental searches. The successful one-pass realization of three target behaviors illustrates the promise of a systematic, generative approach to engineering the analog computation native to living cells.
The authors position this neuromorphic paradigm as addressing a key mismatch in synthetic biology: conventional abstractions and digital-centric design workflows do not naturally capture the graded, context-dependent computations performed by biological regulatory networks.
The work received support from a range of funders listed by the authors, including the Defense Advanced Research Projects Agency (DARPA), the U.S. National Science Foundation (NSF), the U.S. National Institutes of Health (NIH), the Advanced Research and Innovation Agency (ARIA), the U.S. Air Force Office of Scientific Research, the United States Army Research Office, Imperial College London/Schrödinger Scholarship, and others. The preprint is posted on bioRxiv and has not been peer reviewed.
The authors disclose that J.D., G.K.A.W., and R.W. are co-founders of an early-stage company founded to develop technologies related to this work; the remaining authors declared no competing interests.
All statements and claims above are drawn from the preprint abstract and accompanying metadata. Specific experimental methods, quantitative performance metrics, training datasets, model architectures, and sequences or construct details were not reported in the provided source text and therefore are not described here. The linked preprint and supplementary materials should be consulted for full methodological and data details.