A Living Brain in a Petri Dish — Cultured Neurons Succeed at Learning
In this article
Proceedings of the National Academy of Sciences (PNAS) 2026;123(11):e2521560123. doi: 10.1073/pnas.2521560123 · Tohoku University & Future University Hakodate & University of Tokyo
In a laboratory, just a few hundred rat cortical neurons are living on a chip. But these neurons are no longer an ordinary culture: scientists have taught them to produce sine waves, follow triangular and square waves, and even imitate chaotic mathematical patterns such as the Lorenz attractor in real time . The Tohoku University team's study, published in PNAS, opens a new door at the intersection of medicine and artificial intelligence: can living neurons perform machine learning? The answer is a cautious “yes” — and this achievement has broad implications extending from cancer research to brain-computer interfaces.
The Essence of the Story: Cultured Neurons Produce Sine Waves
One of a living brain's most extraordinary abilities is producing meaningful, coordinated muscle movements from seemingly chaotic neural activity. A basketball player's shot toward the basket, the precision with which a pianist's fingers strike the keys, or a child's speech sounds — all are orderly temporal patterns emerging from the chaotic spontaneous activity of billions of neurons. How does this transformation occur? And perhaps an even more interesting question: can a similar transformation be reproduced by growing just a few hundred neurons on a chip?
This study's answer is clear: yes, it can. The Japanese team integrated cultured rat cortical neurons with high-density microelectrode arrays (HD-MEA) to build a system operating under real-time closed-loop feedback. The result: living neural networks learned to follow an external “teacher” signal and produce sine waves with periods of 4–30 seconds, triangular and square waves, and even the three-dimensional chaotic Lorenz attractor .
THE KEY QUESTION: WHY DOES IT MATTER?
Computers are enormous energy-consuming machines. The brain, by contrast, operates on about 20 watts — as much as an LED light bulb. If we can integrate living neurons' computational capacity into artificial intelligence, energy efficiency could be revolutionized. This study is one of the most concrete steps yet toward “wetware”.
What Is Reservoir Computing? How Is Order Extracted from Chaos?
Understanding this study requires familiarity with a machine learning concept: reservoir computing (reservoir computing). The basic idea is to use the transient dynamics of a complex, high-dimensional, nonlinear system (the reservoir) as a “ready-made computational resource”. Rather than trying to change the complex activity inside the reservoir, only a simple linear decoder that reads its output is trained. In artificial neural networks, this approach dramatically reduces training costs and delivers impressive results in tasks such as time-series processing.
The mammalian brain inspired reservoir computing. However, the use of living neural networks as physical reservoirs remained limited for a long time — because neurons cultured in vitro (outside glass, that is, under laboratory conditions) usually display excessively synchronized bursting activity . All neurons fire at once, the exact opposite of the “rich, diverse dynamics” a reservoir requires.
Sono et al., PNAS 2026 — Adapted from Figure 1A
Breaking Excessive Synchronization: A Microfluidic Work of Art
The team's method for overcoming this problem is remarkably elegant from an engineering perspective: creating a modular network architecture with microfluidic devices. A polydimethylsiloxane (PDMS) microfluidic film was placed over the electrode surface. The film confines neurons to square wells measuring 100×100 micrometres; connections between wells are provided only through narrow microchannels (~5 micrometres wide).
The team developed two different microfluidic designs:
- Each well connects to 4 neighbors
- Square lattice topology
- Relatively even connections between wells
- Mean firing rate: 0,56 ± 0,09 Hz
- This group had the best learning performance
- Connections grouped at four scales
- Heterogeneous connectivity pattern
- Sparser communication between wells
- Mean firing rate: 0,32 ± 0,08 Hz
- Lower performance than the lattice
A third group — the control group — used homogeneous neuron cultures (grown without a microfluidic film). The results revealed a striking difference. Homogeneous cultures showed the expected excessively synchronized bursting pattern: the mean correlation between electrodes was 0,45 ± 0,04. But this correlation fell dramatically in the lattice and hierarchical designs, respectively to 0,11 ± 0,01 and 0,12 ± 0,03. These networks were therefore no longer excessively synchronized but displayed much richer and more diverse dynamics.
How Were the Neurons “Taught”? The FORCE Algorithm
To help the neurons “learn” to produce the desired patterns, the team used FORCE (First-Order Reduced and Controlled Error). Its basic logic can be summarized as follows: at each time step, the system calculates the difference between the neurons' output and the target signal, and continuously updates the decoder's weights to minimize that difference. At the same time, it sends feedback stimulation to the neurons, steering the network's activity toward the target.
Initially, the neurons displayed spontaneous, irregular activity. When FORCE learning began, the decoder weights started fluctuating rapidly, causing the neurons in the network to adopt a periodic firing pattern matched to the target signal (for example, a sine wave with a period of 30 seconds). Over time, the weight fluctuations decreased and the system reached a stable matrix configuration capable of generating the target function. Even more strikingly, after learning was switched off, the system could continue its oscillatory activity on its own .
Sono et al., PNAS 2026 — Compiled from Figure 3C · MSE: Mean Squared Error
As the chart shows, networks shaped with microfluidics — especially the lattice design — exhibited lower errors at every sine-wave period than homogeneous cultures and even the cell-free control condition. A notable finding: for the rapid sine wave with a period of 4 seconds, homogeneous cultures failed at a level indistinguishable from the cell-free control — the neurons were present as if they were not there. Lattice networks consistently outperformed the others at every period.
What Was Learned? Sine, Triangular, Square, and a Chaotic Attractor
The study's practical strength lies in the ability to retrain the same culture for different target signals. The same neuronal network could learn the following tasks in sequence:
📈 1. Sine Waves (Periods of 4, 10, and 30 Seconds)
The same culture successfully adapted to sine waves of different frequencies. Its easy retraining for different frequencies is the strongest evidence of the system's flexibility.
📐 2. Triangular Wave
A more complex signal containing multiple frequency components and nondifferentiable points. The neural network followed this successfully too — although some error was observed at sharp corners.
▅ 3. Square Wave
The most challenging periodic target — abrupt transitions and constant plateaus. The current system's 332 ms delay limited faithful reproduction of very high-frequency square waves.
🦋 4. Lorenz Attractor (Chaotic Signal)
A three-dimensional, nonlinear, chaotic time series. The team achieved this too at a proof-of-concept level — correlations exceeded 0,8 in all three dimensions. Although some error appeared at high-amplitude peaks in the chaotic trajectory, the system retained its chaotic oscillatory behavior even after learning was switched off.
Are the Neurons Really “Learning”, or Merely Responding?
Perhaps the study's most philosophical question is whether the living neurons are actually performing learning , or are simply a component of a feedback loop controlled by an external computer. The answer partly includes both.
Technically, learning occurs in two stages. First, an external algorithm (FORCE) continuously updates the decoder weights. Second — and more interestingly — the neural network's own intrinsic dynamics reorganize alongside this feedback. The study showed that enabling feedback shifted neural activity from an irregular, high-variance state to more structured, low-dimensional dynamics. In this structured state, neurons made stable transitions between two attractor states corresponding to the positive and negative peaks of the sine wave. This is a miniature imitation of the process believed to occur when the living brain generates real motor commands.
Not Everything Is So Smooth: Limitations and Open Problems
The researchers transparently stated several limitations of their work. This honesty strengthens the scientific quality of the results:
- Closed-loop delay of 332 ms — insufficient for high-frequency signals
- Performance deteriorates after learning (MSE increased in 99% of trials)
- Nondifferentiable transitions of triangular and square waves cannot be reproduced clearly
- High-amplitude peaks of the Lorenz attractor are not fully accurate
- The system is vulnerable to reservoir perturbations
- Rat neurons were used — human neurons may behave differently
- Culture lifespan of 14-31 days — long-term stability is uncertain
- Only 128 wells and a few hundred neurons — a very limited scale
- The system depends on a nutrient medium; autonomous operation is limited
- Short-term dynamics of synaptic plasticity could not be incorporated into the model
Where Could This Technology Lead in the Future?
In their conclusion, the authors highlight several exciting applications:
From the perspective of oncology and neuro-oncology, one of the most interesting possibilities is using neurons derived from human induced pluripotent stem cells (iPSCs) in similar systems. This could provide a platform capable of replacing animal experiments. Personalized drug-response tests and neuroprosthetic designs using neuronal samples from patients with brain tumors may also become possible.
🎓 DROZDOGAN Academy — Commentary and Discussion
BIOLOGICAL NEURAL NETWORKS · RESERVOIR COMPUTING · NEUROMORPHIC TECHNOLOGY · ETHICAL AND MEDICAL DISCUSSION
The Importance of This Study
This study represents an important threshold at the intersection of artificial intelligence and biology. Demonstrating that living neurons can be trained by an artificial intelligence algorithm makes clear that “intelligence” is not exclusive to silicon-based systems. For medical science, its greatest value is showing that the mechanisms of motor-command generation in the human brain — the transformation of a chaotic signal into consistent motor movement — can be modeled under laboratory conditions. This offers a long-term horizon for treatment strategies in conditions such as Parkinson's disease, ALS, post-stroke loss of motor function, and spinal cord injuries.
Strengths
The study's most original contribution is combining microfluidic engineering with fine control of biological network dynamics. Solving excessive synchronization through modular connections is an elegant approach not only in biology, but also in network science and complex systems theory. The ability of the same culture to learn signals of different frequencies and types is strong evidence of system flexibility. The researchers' open sharing of data and software will also contribute substantially to progress in the field. The Tohoku University team's approximately 15 years of work in this area is another important factor increasing methodological maturity.
Limitations and Open Questions
The study remains at the proof-of-concept level, with substantial obstacles to real-world applications. First, scale: it is unknown whether this experiment involving just a few hundred neurons would work with the same consistency at the scale of thousands, tens of thousands, or millions. Second, stability: the cultures live only a few weeks, and performance rapidly deteriorates after learning. Third, functional significance: producing a sine wave is impressive, but these neurons are not yet performing real computation in the sense of “carrying out a task”. Fourth, and perhaps most importantly, the ethical questions: what does it mean to give living neurons “memory” or “experience”? Could these neural networks acquire “consciousness”? The ethical framework for brainoids and organoid intelligence is not yet fully mature.
Implications for Clinical Practice and Neuro-Oncology
This technology will not directly affect oncology practice today; however, it could open several important applications in the medium to long term. First, miniature networks created from normal cortical neurons after tumor removal in patients with brain tumors could serve as platforms for personalized drug-response testing. Second, they could model in vitro the electrophysiological effects of tumors such as glioblastoma on surrounding healthy tissue. Third, systems of this kind have the potential to replace animal experiments in evaluating neurotoxicity after chemotherapy and radiotherapy. The most exciting possibility is making it possible to test the mechanisms and treatment strategies of conditions such as radiotherapy-induced cognitive decline or chemotherapy-induced peripheral neuropathy in patient-specific neuronal cultures derived from iPSCs.
Questions for the Future
The critical research questions facing this field are: can human iPSC-derived neurons show the same flexibility in this paradigm, and in which disease models could they be used? How can closed-loop delay be reduced to the millisecond scale to learn more complex patterns? When integrated with organoid technology, what advantages can three-dimensional neural networks offer over two-dimensional microfluidic systems? Could work by the Tohoku group and other groups, such as Australia's DishBrain project, combine to create a new standard of “biological computing”? And perhaps the most critical question: could trainable intelligence emerge in living systems at the scale of millions or billions of neurons, and what ethical and regulatory frameworks would then be needed? The answers will profoundly affect not only engineering but also medicine and philosophical anthropology over the next 10-20 years.
Scientific Sources
- Sono Y, Yamamoto H, Nishi Y, Sumi T, Sato Y, Hirano-Iwata A, Katori Y, Sato S. Online supervised learning of temporal patterns in biological neural networks under feedback control. Proc Natl Acad Sci USA 2026;123(11):e2521560123. doi: 10.1073/pnas.2521560123
- Sussillo D, Abbott L. Generating coherent patterns of activity from chaotic neural networks. Neuron 2009;63:544–557.
- Nicola W, Clopath C. Supervised learning in spiking neural networks with FORCE training. Nat Commun 2017;8:2208.
- Yada Y, Yasuda S, Takahashi H. Physical reservoir computing with FORCE learning in a living neuronal culture. Appl Phys Lett 2021;119:173701.
- Sumi T et al. Biological neurons act as generalization filters in reservoir computing. Proc Natl Acad Sci USA 2023;120:e2217008120.
- Cai H et al. Brain organoid reservoir computing for artificial intelligence. Nat Electron 2023;6:1032–1039.
- Kagan BJ et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron 2022;110:3952–3969.e8.
- Yamamoto H et al. Modular architecture facilitates noise-driven control of synchrony in neuronal networks. Sci Adv 2023;9:eade1755.
- Smirnova L et al. Organoid intelligence (OI): The new frontier in biocomputing and intelligence-in-a-dish. Front Sci 2023;1:1017235.
- Beniaguev D, Segev I, London M. Single cortical neurons as deep artificial neural networks. Neuron 2021;109:2727–2739.e3.
This article was created using various editorial tools, including artificial intelligence, as part of the preparation process. It was reviewed by human editors before publication.
Was this article helpful?
Your feedback helps us improve our content.
Reader comments
Share your comment
Your name and comment are published only after administrator approval. Your email address is never shown to visitors.
Be the first to comment.