← Latest papers
📄 synthetic biology

Generative Neuromorphic Programming of Mammalian Cells

This paper introduces a generative neuromorphic programming framework that combines RNA-targeting endoribonucleases for analog circuit construction with biomorphic neural networks for accurate prediction and a "biocompiler" for automated design, enabling the successful single-pass engineering of complex, previously unseen analog behaviors in mammalian cells.

Original authors: Disset, J., Van De Mark, C., Wachter, G. K. A., Kuo, S.-H., Moorman, A., Xu, W., Palanques Tost, E., Buritica Monroy, A., Belta, C., Weiss, R.

Published 2026-09-22
📖 9 min read🧠 Deep dive

Original authors: Disset, J., Van De Mark, C., Wachter, G. K. A., Kuo, S.-H., Moorman, A., Xu, W., Palanques Tost, E., Buritica Monroy, A., Belta, C., Weiss, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cells are not simple on/off switches; they are sophisticated processors that constantly weigh multiple signals to make decisions. In the natural world, a cell might sense the levels of several different proteins floating nearby and, based on the precise combination and strength of those signals, decide whether to divide, repair damage, or die. This decision-making happens through a dense web of internal connections that produce smooth, graded responses rather than binary yes-or-no answers. For decades, scientists trying to program cells for new tasks have struggled because their tools were built for digital logic, treating biological parts like computer bits that are either fully on or fully off. This mismatch made it incredibly difficult to design circuits that could handle the subtle, analog nature of living systems, often requiring years of trial and error to achieve even simple behaviors.

A team of researchers at the Massachusetts Institute of Technology and other institutions has now bridged this gap by teaching cells to compute in the way they naturally do. They developed a system that treats the cell's internal machinery as a flexible, analog computer, capable of performing complex calculations on multiple inputs simultaneously. Instead of forcing cells to follow rigid digital rules, the researchers built a software engine that designs genetic circuits to produce specific, smooth patterns of behavior. They demonstrated this by creating three entirely new genetic circuits from scratch, each designed to recognize a unique combination of two input signals and respond with a precise output. Remarkably, every single circuit worked exactly as the software predicted on the very first try, without any need for manual tweaking or adjustment.

The core of this achievement lies in a new way of thinking about how genes interact. The researchers identified five basic building blocks that mimic the operations of an artificial neural network, but these blocks are made of real biological molecules. They use specific enzymes that can cut RNA to act as negative inputs, effectively subtracting signals, while other mechanisms add signals together. They also found ways to adjust the strength of these connections, much like turning a volume knob, by changing the genetic sequences that control how much protein is made. By connecting these blocks together, they created a system where one part of the circuit could feed its output into the next, allowing for multi-layered processing. This approach allows the cell to compute complex functions, such as detecting a specific range of signal levels and ignoring everything else, a behavior known as a band-pass filter.

To make this system practical, the team created a software package they call the biocompiler. This program acts as a translator between a desired behavior and the DNA instructions needed to build it. The software relies on a model called a biomorphic neural network, which learns how biological processes work by studying a large collection of existing experiments. Instead of guessing the mathematical formulas that govern how genes interact, the software learns the actual behavior of these processes from real data. It treats every genetic part, such as a promoter or a specific enzyme, as a flexible component with a unique "fingerprint" that the system can adjust. When a scientist wants a cell to perform a new task, they simply describe the desired output pattern, and the software searches through millions of possible combinations of parts and connections to find the one that will produce that result.

The researchers tested this system by asking it to design three distinct patterns of behavior that had never been seen before in their training data. The software proposed three completely new circuit architectures, none of which existed in the library of parts it had studied. The team then built these circuits in the lab, inserting the DNA instructions into human cells. In every case, the cells produced the exact pattern the software had predicted. The match between the design and the reality was so close that the difference was no greater than the natural variation seen when repeating the same experiment multiple times. This success proves that the software can not only predict how a circuit will behave but can also invent new circuits that work immediately, bypassing the long cycles of trial and error that have traditionally slowed down synthetic biology.

This work represents a significant shift in how scientists approach the engineering of life. Previously, designing a cell to perform a complex task was like trying to build a clock by guessing which gears would fit together, often requiring many failed attempts to get the timing right. The new method is more like using a blueprint that is generated by a system that already understands the physics of gears. By learning from a vast array of previous experiments, the software has built a deep understanding of how genetic parts interact, allowing it to generalize to new situations it has never seen. The system is robust enough to work across different types of human cells, suggesting that the rules it learned are fundamental to how these cells function, rather than being specific to just one type.

The implications of this work extend beyond just making better circuits. It establishes a framework where biological engineering can become a cumulative science, where every new experiment adds to a shared library of knowledge that makes future designs more accurate. As the system learns from more data, it becomes better at predicting the behavior of increasingly complex networks. The researchers have shown that it is possible to program cells with the same precision and reliability that engineers use to design electronic devices, but using the native analog language of biology. This opens the door to creating cells that can perform sophisticated tasks, such as identifying and destroying cancer cells based on a complex signature of markers, or guiding the growth of tissues with precise spatial patterns, all without the need for constant human intervention to fix what goes wrong.

The success of the biocompiler relies on a specific type of learning that differs from traditional machine learning. Instead of training a single massive model to recognize patterns in images or text, this system learns the behavior of individual biological processes, such as how a specific enzyme cuts RNA or how a promoter drives gene expression. These learned behaviors are then recombined to form new circuits. This means that the system does not just memorize the circuits it has seen; it understands the underlying rules that govern them. When faced with a new design challenge, it can assemble these known rules in a novel way to solve the problem. This compositional approach allows the system to handle a vast design space that would be impossible to explore through experiments alone, as the number of possible combinations of genetic parts grows far too quickly for any lab to test them all.

In the experiments described, the researchers used a technique called poly-co-transfection to test their designs. This method allows them to deliver varying amounts of different genetic inputs to a single population of cells, creating a wide range of conditions in a single experiment. By measuring the output of each cell, they can map out the entire response surface of a circuit, seeing how it behaves across all possible input combinations at once. This high-resolution view of cellular behavior provided the detailed data needed to train the software and verify its predictions. The fact that the software could predict the behavior of circuits it had never seen, and that these predictions held true in the lab, confirms that the model has captured the essential logic of the cell's internal network.

The researchers also tested whether the system could transfer its knowledge from one type of cell to another. They trained the software using data from human embryonic kidney cells and then asked it to predict the behavior of circuits in two other cell types: retinal pigment epithelial cells and Chinese hamster ovary cells. The software performed nearly as well in these new cell types as it did in the one it was trained on, suggesting that the fundamental rules of how these genetic parts interact are consistent across different human cell lines. This portability is crucial for practical applications, as it means a design created in one cell type can likely be adapted to work in another without starting from scratch.

While the current system is a major step forward, the researchers acknowledge that there is still work to be done. The software currently operates within a set of pre-defined structural templates, or scaffolds, and cannot yet invent entirely new types of connections from nothing. Future versions of the system may learn to generate these structural rules directly from data, further expanding the range of possible designs. Additionally, the current models capture the behavior of individual processes well but do not yet explicitly account for global effects, such as competition for cellular resources, which can sometimes alter circuit behavior. However, the framework is designed to be flexible, and these limitations can be addressed as the system learns from more complex data.

The ability to program cells with such precision transforms the potential of synthetic biology. It moves the field from a discipline of trial and error to one of rational design, where engineers can specify a function and receive a working genetic circuit in return. This capability is essential for the next generation of cell therapies, where cells must be programmed to detect and respond to the complex, changing environment of the human body. By harnessing the native analog computing power of cells, this approach offers a path to creating living machines that are as sophisticated and adaptable as the biological systems they are built from. The work demonstrates that with the right tools, we can finally speak the language of cells fluently, turning their natural complexity into a programmable resource for solving some of the most difficult challenges in medicine and biology.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →