Bio-silicon hybrid computing: how living neurons and chips work together

Bio-silicon hybrid computing is an approach that connects living neurons, grown as cultures or brain organoids, to electronic chips that record and stimulate them in real time. Silicon handles input, output, timing and storage, while the living network supplies adaptive, nonlinear dynamics. Together they form a closed loop that can be trained on simple tasks.

Key takeaways

• Every neuron-based computer built so far is a hybrid: living tissue does part of the processing, and conventional electronics do the rest.
• A working system has six layers: living tissue, life support, an electrode interface, a real-time controller, conventional compute for machine learning, and a software and data layer.
• Real systems include DishBrain (Kagan et al., 2022), Brainoware (Cai et al., 2023), FinalSpark’s Neuroplatform (Jordan et al., 2024) and Cortical Labs’ CL1.
• Research uses such as compound testing, disease modelling and neuroscience are the most established applications. Low-power computing on real workloads is still speculative.
• The hard problems are engineering ones: tissue lifespan, variability, signal quality, I/O bandwidth, latency, scaling and shared standards, alongside ethics.

What is bio-silicon hybrid computing? #

The field goes by several names. You will see “biohybrid computing,” “hybrid biological computer,” “neurons on a chip,” wetware computing and, when the tissue is a 3D brain organoid, organoid intelligence. All describe the same basic arrangement: a living neural network sits on or next to a chip, and the chip lets a computer talk to it.

The word “hybrid” matters. Living neurons cannot read a file or print an answer. They respond to electrical and chemical signals and produce electrical activity. Everything that turns that activity into computation is done by silicon. So the realistic question is not “biology or silicon?” but “which parts of the job does each do best?”

Bio-silicon hybrids sit inside the wider field of biocomputing, which also includes molecular approaches such as DNA computing. Our overview of biological computing from DNA molecules to living neurons covers that broader landscape. This article focuses on the neuron-and-chip systems, how they are built, and what they can and cannot do.

A short history of neurons on a chip #

The idea is older than organoids:

• Silicon-neuron circuits (2001). Zeck and Fromherz reported in PNAS a circuit in which a chip’s capacitive stimulator excited a snail neuron, the signal crossed an electrical synapse to a second neuron, and that neuron’s activity modulated a transistor on the same chip.
• The animat (2001). DeMarse, Wagenaar, Blau and Potter described in Autonomous Robots a “neurally controlled animat,” in which rat cortical neurons on a microelectrode array controlled a simulated animal while a computer fed sensory stimulation back to the culture.
• Hybrots (2004). Bakkum and colleagues described embodying cultured networks in simulated and robotic bodies and used the term “hybrot,” short for hybrid robot.
• MEART (2007). Bakkum and colleagues described in Frontiers in Neurorobotics a cultured network connected to a drawing robot.
• Adaptive training (2008). Bakkum, Chao and Potter showed in the Journal of Neural Engineering that adaptive closed-loop stimulation could shape an embodied network’s behavior in a goal-directed task.

What changed in the last decade is scale and access: human neurons derived from stem cells, 3D brain organoids, CMOS chips with thousands of electrodes, and platforms that researchers can use remotely.

How is a hybrid biological computer built, layer by layer? #

Think of a biohybrid system as a stack. Each layer has its own failure modes, and all must work together.

Layer 1: living neural tissue #

The computing substrate is a network of neurons that fire action potentials and change their connections through synaptic plasticity. Older work used primary rodent cortical cells. Most current work uses human neurons derived from induced pluripotent stem cells, a technique introduced by Takahashi and Yamanaka in Cell (2006), either as flat 2D cultures or as 3D organoids.

Organoids matter because they are denser and more structured than flat cultures. Trujillo and colleagues reported in Cell Stem Cell (2019) that cortical organoids developed increasingly complex oscillatory network activity over months of culture. Our guide on how brain organoids are grown explains the process from stem cell to firing network.

The tissue’s job is to provide rich, nonlinear responses that depend on past input, and to change them with experience.

Layer 2: life support and microfluidics #

Neurons need warmth, controlled gases, fresh nutrients and protection from contamination, around the clock. Life support sets the system’s uptime.

Early work recognized this. Potter and DeMarse (2001, Journal of Neuroscience Methods) introduced membrane-sealed culture dishes designed to support long-term studies of networks on electrode arrays. Modern platforms add automation. Jordan et al. (2024) described FinalSpark’s Neuroplatform using a microfluidics system for fully automated medium flow and change, to reduce disruptions from people opening the incubator. Their incubators log CO2, O2, humidity, pressure and temperature, along with door-opening events, because those disturbances show up in the recordings.

Longer term, researchers are exploring organoid vascularization to deliver nutrients into the core of larger tissues.

Layer 3: the electrode interface #

The physical meeting point of biology and silicon is usually a microelectrode array (MEA): a grid of tiny electrodes on which tissue rests. Each electrode senses small voltage changes from nearby neurons and can deliver brief, charge-balanced current or voltage pulses.

There are two broad families:

• Passive MEAs route each electrode through its own wire to external amplifiers. They are simple and robust but limited to tens or a few hundred electrodes.
• CMOS high-density microelectrode arrays build amplifiers, multiplexers and analog-to-digital converters into the chip under the electrodes. Ballini and colleagues described in the IEEE Journal of Solid-State Circuits (2014) a CMOS array with 26,400 electrodes and 1,024 channels that can be recorded at once. Müller and colleagues (2015, Lab on a Chip) showed such platforms can study neurons at subcellular, cellular and network levels.

The CMOS chip is where “neurons on a chip” becomes literal: the first stage of signal processing happens micrometers below the cells. For a detailed walkthrough of electrode size, sampling and stimulation artifacts, see microelectrode arrays explained.

Layer 4: the real-time controller, encoder and decoder #

Above the chip sits the part that makes the system a computer rather than a recording rig: a real-time controller that runs closed-loop stimulation. It does three jobs:

1. Detect. Find spikes in the raw voltage stream, often with fast threshold crossing. Full spike sorting, which assigns spikes to individual neurons, is usually too slow for the loop and is done offline.
2. Decode. Convert activity in chosen electrodes or regions into an output, such as “move the paddle up.”
3. Encode. Convert the next piece of task information into a stimulation pattern: which electrodes, at what rate and when.

Together, steps two and three are neural encoding and decoding. The encoding scheme is a design choice with real consequences. In DishBrain, Kagan and colleagues encoded a ball’s position by which electrode was stimulated and how often, and decoded spike counts from two “motor” regions in 10 millisecond windows. Wagenaar, Madhavan, Pine and Potter (2005, Journal of Neuroscience) used a closed-loop controller of this kind to suppress spontaneous network bursts in cortical cultures, showing that the controller can shape the tissue’s dynamics, not just read them.

Layer 5: conventional compute and machine learning #

Many biohybrid systems use ordinary computers to train a readout. In reservoir computing, the living network acts as a fixed, nonlinear system with fading memory, and only a simple digital model on top of it is trained to map responses to answers. This is where most of the “learning” in some systems happens, so it deserves scrutiny: a strong digital model can extract structure from almost any rich signal.

Conventional compute also runs simulated environments, agents that choose stimulation, and offline analysis.

Layer 6: software and data layer #

The top of the stack is software: the APIs that let researchers write experiments, schedulers that allocate tissue, and storage for what comes back. Jordan et al. described a Python API for the Neuroplatform and stated that over three years the platform was used with more than 1,000 brain organoids, collecting more than 18 terabytes of data. At that scale, data formats matter. Neurodata Without Borders (NWB), described by Rübel and colleagues in eLife (2022), is a widely used open standard for electrophysiology data.

This layer also carries provenance: which cell line, which culture, which protocol version, which consent terms. Without it, results cannot be reproduced or audited.

The stack at a glance #

How does data flow through a closed loop? #

One cycle of a closed-loop biohybrid experiment looks like this:

1. The task environment produces a new state, for example the ball moves.
2. The encoder turns that state into a stimulation pattern.
3. The chip delivers pulses through selected electrodes.
4. Neurons respond, and the electrodes pick up voltage changes.
5. On-chip amplifiers and converters digitize the signals.
6. The controller detects spikes and blanks or removes stimulation artifacts.
7. The decoder converts activity into an output action.
8. The environment updates, and feedback is computed.
9. Data from every step is written to storage with timestamps.

Then the loop repeats, often many times a second. Our article on how neurons learn in a dish explains why closing this loop, rather than stimulating on a fixed schedule, was often what made learning-like changes appear.

Where latency comes from #

Every step adds delay. Acquisition hardware buffers samples before sending them. Data crosses a bus such as USB to a host computer. An operating system may schedule the processing thread late. Spike detection, decoding and the environment update all take time, and stimulation artifacts can blind the recording electrodes briefly after each pulse.

Why it matters: neurons tie plasticity to timing. Bi and Poo reported in the Journal of Neuroscience (1998) that in cultured hippocampal neurons, whether a synapse strengthened or weakened depended on the order of pre- and postsynaptic spikes within a window of tens of milliseconds. This is spike-timing-dependent plasticity. If feedback arrives much later than the activity that produced it, the culture has little chance to associate the two.

That is why designers push time-critical steps (detection, decoding, triggering stimulation) close to the electrodes, onto dedicated hardware or tightly optimized software, and leave training and archival to conventional computers. Jitter, the variation in delay from cycle to cycle, matters too, because it blurs the timing relationships the network might learn from.

What does silicon do, and what does biology do? #

The division of labor in today’s systems is clear, even if it may shift as the field matures.

The last row deserves care. Brains run on very little power, which motivates the field, but no hybrid system has yet shown efficient computation on a useful task once incubators, pumps and electronics are counted. We cover this in biocomputing vs silicon.

Which bio-silicon hybrid systems exist today? #

Only a handful of systems are well documented in peer-reviewed or primary sources.

DishBrain (Kagan et al., 2022) #

Kagan and colleagues reported in Neuron (2022) that mouse cortical cells and human iPSC-derived cortical neurons on a CMOS high-density MEA learned to play a simplified Pong game. The array recorded up to 1,024 of its 26,400 electrodes at 20,000 samples per second, and custom low-latency software linked it to the game. The authors reported improved rally lengths within minutes of real-time gameplay, an effect absent in control conditions. The paper’s use of the word “sentience” drew criticism; we cover the design and the debate in detail in how neurons learn in a dish.

Brainoware (Cai et al., 2023) #

Cai and colleagues reported in Nature Electronics (2023) a system called Brainoware, in which a brain organoid on a high-density MEA acted as a physical reservoir and a trained digital readout interpreted its responses. They tested speaker identification from Japanese vowel recordings and prediction of a chaotic system, the Hénon map. MIT Technology Review reported speaker identification accuracy of about 78 percent after training, and News-Medical reported accuracy rising from 51 to 78 percent over training. MIT Technology Review also reported that Brainoware was less accurate than artificial neural networks on that task. See our full article on reservoir computing with brain organoids.

FinalSpark Neuroplatform (Jordan et al., 2024) #

Jordan et al. described the Neuroplatform in Frontiers in Artificial Intelligence (2024) as an open, remotely accessible system for wetware computing research. According to the paper:

• It uses four MEAs, each holding four organoids with eight electrodes per organoid, in an air-liquid interface setup.
• Recording runs around the clock at 30 kHz with 16-bit resolution through commercial headstages.
• Organoids can have lifetimes of more than 100 days, supported by automated microfluidic medium exchange.
• Researchers access the system through a Python API, which also controls pumps, cameras and UV light for releasing caged molecules.

Cortical Labs CL1 #

Cortical Labs announced its CL1 in March 2025, as reported by IT Brief Australia. On its CL1 product page, the company describes the CL1 as “the world’s first code deployable biological computer.” According to that page:

• Neurons are grown in a nutrient-rich solution across a silicon chip that sends and receives electrical impulses.
• A Biological Intelligence Operating System (biOS) runs a simulated world and sends the neurons information about it, with their responses affecting that world.
• The system is closed-loop, and its life support is designed to keep neurons alive for up to six months.
• Recordings, applications and life support are on the device, with no external compute required.
• It offers a programmable, bidirectional stimulation and read interface, and can connect to cameras, USB devices and actuators.

These are the company’s own descriptions, not independent benchmarks. Cortical Labs also offers a cloud service.

Memristor-neuron hybrids (Serb et al., 2020) #

A different kind of hybrid joins living neurons to neuromorphic electronics. Serb and colleagues, with authors at the University of Southampton, the University of Padova, TU Dresden and the University of Zurich and ETH Zurich, reported in Scientific Reports (2020) memristive links between rat hippocampal neurons and silicon spiking neurons. A memristor paired with a titanium oxide microelectrode connected a silicon neuron to a biological one, producing responses resembling excitatory postsynaptic potentials, and a microelectrode-memristor pair carried the reverse link. They demonstrated a three-neuron brain-silicon network in which the memristive synapses underwent long-term potentiation or depression driven by firing rates. This points toward hybrids where the “synapses” between living and artificial neurons are themselves physical, adaptive devices, a meeting point with neuromorphic computing and spiking neural networks.

The systems side by side #

What are the applications of organoid intelligence and biohybrid computing? #

Organoid intelligence applications are often described in a single breath, from drug testing to robots to green AI. They are not equally mature. Below, each is labeled:

• Demonstrated: shown in peer-reviewed work.
• Plausible: supported by existing components, but not yet shown as a working hybrid application.
• Speculative: an open research goal with no current evidence of practical advantage.

Drug discovery and neurotoxicity: demonstrated as a model, plausible as a closed-loop assay #

Human brain tissue models are already used to study toxicity. Pamies and colleagues described in ALTEX (2017) an iPSC-derived human brain microphysiological system for studying neurological diseases and toxicity. MEAs add a functional readout: firing rates, bursting and network synchrony, not just whether cells survive. What a hybrid computing setup could add is testing whether a compound changes a network’s ability to adapt in a closed-loop task. That is plausible but not yet established as a validated assay. Cortical Labs markets the CL1 for compound testing and as an alternative to some animal testing, which is a company claim.

Disease modelling and personalized medicine: demonstrated for modelling, plausible for personalization #

Because iPSCs can be made from any donor, organoids can carry a patient’s genetic background. Lancaster and colleagues reported in Nature (2013) that cerebral organoids derived from a patient with microcephaly showed premature neural differentiation, modelling features of the disorder. Birey and colleagues reported in Nature (2017) that in assembled forebrain spheroids, an assembloid approach, interneurons from Timothy syndrome patient cells showed abnormal migration. Using patient-derived tissue in a hybrid system to choose treatments for an individual is plausible, but it has not been shown.

Robotics and embodiment: demonstrated in simple forms, practical use speculative #

Cultured networks have controlled simulated animals (DeMarse et al., 2001), robotic and drawing devices (Bakkum et al., 2004 and 2007) and a game paddle (Kagan et al., 2022). These are simple, slow and tightly controlled tasks. A biohybrid controller for a real robot doing useful work in an uncontrolled environment is speculative.

Low-power edge inference: speculative #

The idea is that living neurons might one day perform some inference tasks with less energy than digital chips, perhaps in devices far from a data center. Today there is no evidence for this. Life support alone needs energy and infrastructure, tissues degrade, and systems like Brainoware were less accurate than conventional neural networks on their benchmark. Smirnova and colleagues, who defined organoid intelligence in Frontiers in Science (2023), framed efficiency as a long-term research goal.

Neuroscience research: demonstrated #

This is where hybrids already earn their keep. Closed-loop systems let researchers study how networks learn and how plasticity depends on timing, in human tissue and under precise control.

What are the challenges of bio-silicon hybrid computing? #

The obstacles are concrete, and most are engineering problems.

1. Lifespan. Tissue eventually degrades. Platforms report months, not years, of useful life, so any learned behavior is temporary and experiments must plan for replacement.
2. Variability. No two cultures are identical. Velasco and colleagues reported in Nature (2019) that some organoid protocols reproducibly generate cortical cell diversity, but Bhaduri and colleagues reported in Nature (2020) that cell stress in cortical organoids impaired molecular subtype specification. Hybrid systems need quality control and calibration per culture.
3. Signal quality. Extracellular signals are microvolts and easily buried in noise. Planar MEAs mostly sample the bottom of a 3D organoid. New interfaces such as the self-folding shell MEAs reported by Huang and colleagues in Science Advances (2022) aim to record from around the tissue.
4. I/O bandwidth. Even an array with tens of thousands of electrodes addresses a small fraction of the neurons in an organoid, and stimulation is coarse compared with the precision of natural synaptic input.
5. Latency. As described above, delays and jitter in the loop weaken the timing relationships that plasticity depends on.
6. Scaling. Adding more organoids multiplies life support, electronics and data, and there is no established way yet to wire multiple organoids into one larger, coherent computer.
7. Standards. Protocols, encodings and metrics differ from lab to lab, which makes results hard to compare. Shared formats like NWB help with data; shared benchmarks are still missing.
8. Ethics. Human cells come from donors, so consent and provenance must be tracked. The Baltimore declaration, published by Hartung and colleagues in Frontiers in Science (2023), called for ethically responsible development of organoid intelligence. We cover these questions in the ethics of organoid intelligence.

What comes next for biohybrid computing? #

Progress is likely to come from every layer of the stack rather than a single breakthrough: longer-lived, more mature tissue, 3D and flexible electrodes, faster controllers closer to the chip, and shared data standards.

Kagan and colleagues set out a broader framing of the field in Biotechnology Advances (2023), which they call synthetic biological intelligence, and the 2023 Frontiers in Science papers sketched a long-term roadmap for organoid intelligence. Our article on the future of biocomputing looks at that roadmap and its open problems in more detail.

A recurring theme is that the software layer is underbuilt. Experiments are often written as one-off scripts tied to a specific rig, which makes them hard to share or repeat. Treating stimulation protocols, encoders, decoders and feedback rules as versioned code, the practice we call organoid programming, is one way to make hybrid experiments reproducible. That software and data layer is what we are building at Organoid Programming.

Conclusion #

Bio-silicon hybrid computing pairs the adaptive, nonlinear dynamics of living neurons with the precision, timing and storage of silicon. Every working system today, from the 2001 animat to DishBrain, Brainoware, the Neuroplatform and the CL1, combines tissue, life support, electrodes, a real-time controller, conventional compute and a software and data layer. The chip does far more than listen: it encodes, decodes, keeps the loop on time and records everything.

The most credible organoid intelligence applications today are research tools for neuroscience, compound testing and disease modelling. Robotics has been shown only in simple forms, and low-power computing remains speculative. Lifespan, variability, signal quality, bandwidth, latency, scaling, standards and ethics are the problems to solve next. For the bigger picture, start with what organoid intelligence is.

Frequently asked questions

What is bio-silicon hybrid computing?

Bio-silicon hybrid computing is an approach in which living neurons, grown as cultures or brain organoids, are connected to electronic chips that record their electrical activity and stimulate them in real time. The silicon handles encoding, decoding, timing, storage and control, while the living network supplies adaptive, nonlinear dynamics. Together they form one closed-loop system that can be trained on simple tasks.

Is a hybrid biological computer the same as a biological computer?

Not quite. A purely biological computer would do all its processing in living or molecular material, as some DNA computing schemes aim to. Every neuron-based system demonstrated so far is a hybrid: living tissue does part of the work, but conventional electronics are needed to feed it inputs, read its outputs, keep the loop timed and run the machine learning that interprets its responses.

What does the silicon chip actually do in a neurons-on-a-chip system?

The chip, usually a microelectrode array, senses tiny voltage changes from nearby neurons, amplifies and digitizes them, and delivers small electrical pulses back. In CMOS high-density arrays, amplifiers and converters are built under thousands of electrodes. Attached computers detect spikes, translate task data into stimulation patterns, decode responses into outputs and log everything for later analysis.

What is the difference between biohybrid computing and neuromorphic computing?

Neuromorphic computing builds brain-inspired circuits entirely from electronics, such as spiking neuron chips and memristive synapses. Biohybrid computing uses real living neurons as part of the system. The two can meet: Serb and colleagues reported in Scientific Reports (2020) a small network in which memristive synapses linked rat hippocampal neurons and silicon spiking neurons in both directions.

What are the main applications of organoid intelligence and biohybrid systems?

The most established uses are research tools: measuring how compounds affect human neural network activity, modelling disorders with patient-derived cells, and studying how networks learn. Embodied control of simulated or simple robotic bodies has been demonstrated in basic forms. Low-power computing for real-world tasks remains speculative, because no biological system has yet matched silicon on a practical workload.

How long can neurons live on a chip?

It depends on the culture and the life support. Jordan et al. (2024) reported that FinalSpark’s Neuroplatform runs experiments on neural organoids with lifetimes of more than 100 days, helped by automated microfluidic medium exchange. Cortical Labs states that the CL1’s life support is designed to keep neurons alive for up to six months. Even so, tissue lifespan remains a basic limit for hybrid systems.

Why does latency matter in a closed-loop biohybrid system?

Neurons change their connections based on the relative timing of activity, often on scales of tens of milliseconds. If feedback reaches the culture long after the activity that caused it, the network cannot link cause and effect, and closed-loop learning becomes harder. That is why hybrid systems run spike detection, decoding and stimulation on dedicated, low-latency hardware close to the electrodes.

Can I buy or access a bio-silicon hybrid computer today?

Research access exists. Jordan et al. (2024) described FinalSpark’s Neuroplatform as remotely accessible through a Python API, and Cortical Labs offers its CL1 device and a cloud service. These are experimental research platforms for running stimulation and recording studies on living neurons, not general-purpose computers or replacements for conventional hardware.

Sources

1. Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron.
2. Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.
3. MIT Technology Review (2023). Human brain cells hooked up to a chip can do speech recognition.
4. News-Medical (2023). Brainoware: a breakthrough AI approach using brain organoids for advanced computation.
5. Jordan et al. (2024). Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence.
6. Cortical Labs. CL1 product page.
7. IT Brief Australia (2025). Cortical Labs launches CL1, the first biological computer.
8. Serb et al. (2020). Memristive synapses connect brain and silicon spiking neurons. Scientific Reports.
9. Zeck & Fromherz (2001). Noninvasive neuroelectronic interfacing with synaptically connected snail neurons immobilized on a semiconductor chip. PNAS.
10. DeMarse et al. (2001). The neurally controlled animat: biological brains acting with simulated bodies. Autonomous Robots.
11. Potter & DeMarse (2001). A new approach to neural cell culture for long-term studies. Journal of Neuroscience Methods.
12. Bakkum et al. (2004). Removing some 'A' from AI: embodied cultured networks. In Embodied Artificial Intelligence (Springer).
13. Bakkum et al. (2007). MEART: the semi-living artist. Frontiers in Neurorobotics.
14. Bakkum, Chao & Potter (2008). Spatio-temporal electrical stimuli shape behavior of an embodied cortical network in a goal-directed learning task. Journal of Neural Engineering.
15. Wagenaar, Madhavan, Pine & Potter (2005). Controlling bursting in cortical cultures with closed-loop multi-electrode stimulation. Journal of Neuroscience.
16. Ballini et al. (2014). A 1024-channel CMOS microelectrode array with 26,400 electrodes for recording and stimulation of electrogenic cells in vitro. IEEE Journal of Solid-State Circuits.
17. Müller et al. (2015). High-resolution CMOS MEA platform to study neurons at subcellular, cellular, and network levels. Lab on a Chip.
18. Bi & Poo (1998). Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of Neuroscience.
19. Takahashi & Yamanaka (2006). Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. Cell.
20. Lancaster et al. (2013). Cerebral organoids model human brain development and microcephaly. Nature.
21. Birey et al. (2017). Assembly of functionally integrated human forebrain spheroids. Nature.
22. Pamies et al. (2017). A human brain microphysiological system derived from induced pluripotent stem cells to study neurological diseases and toxicity. ALTEX.
23. Trujillo et al. (2019). Complex oscillatory waves emerging from cortical organoids model early human brain network development. Cell Stem Cell.
24. Velasco et al. (2019). Individual brain organoids reproducibly form cell diversity of the human cerebral cortex. Nature.
25. Bhaduri et al. (2020). Cell stress in cortical organoids impairs molecular subtype specification. Nature.
26. Huang et al. (2022). Shell microelectrode arrays (MEAs) for brain organoids. Science Advances.
27. Rübel et al. (2022). The Neurodata Without Borders ecosystem for neurophysiological data science. eLife.
28. Smirnova et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science.
29. Hartung et al. (2023). The Baltimore declaration toward the exploration of organoid intelligence. Frontiers in Science.
30. Kagan et al. (2023). The technology, opportunities, and challenges of Synthetic Biological Intelligence. Biotechnology Advances.