Computers that combine living human brain cells with silicon chips to form a neural network can now be used by researchers to conduct experiments in different fields.
Until recently, accessing a biocomputer usually was not within reach of most researchers, since it is expensive to culture brain organoids in the lab (called wetware computing) and would require in-house experts. However, a few companies now are developing biocomputing platforms that are being made available for researchers to use.
“It’s huge because it allows us to do things that we just couldn’t do otherwise,” said Ben Ward-Cherrier, a senior lecturer at the University of Bristol in the U.K.
Artificial intelligence (AI) has long tried to emulate the workings of the human brain, resulting in the current generation of chatbots such as ChatGPT, which are capable of sophisticated responses. Now, however, there is a growing interest in incorporating brain cells themselves into machines to create biocomputers.
“Brain cells form the only known ground truth of intelligence,” said Brett Kagan, the Chief Scientific Officer of Australian tech company Cortical Labs. “People have invested literally trillions of dollars, especially recently, into silicon computing and machine learning, but no one has generated anything that even comes close to generalized intelligence over the course of 60 years of history in this area.”
Brain cells are thought to learn more efficiently than their artificial counterparts, since they are better at adapting over time, and they also consume much less energy. The massive amount of power required to train AI models is a growing concern: training the large language model (LLM) that powers ChatGPT 3, for example, consumed almost 1,300 megawatt hours of energy, according to one estimate. That much power would sustain 130 American homes for one year.
“The brain operates on about 20 watts and is able to do lots of very complex computations,” said Ward-Cherrier. “That’s much more power efficient than any kind of computing that we’re able to design currently.”
Ward-Cherrier and his colleagues, who are working on replicating human touch in an artificial system, have always wanted to bring down the amount of power required to run their simulations. In a new project, he and his team are using a biocomputing platform developed by Swiss startup company Final Spark to see if it can perform tactile tasks. They are among a select group of researchers who have been given free access to the system via the cloud.
The platform consists of 16 lab-grown mini-brains, housed in a microfluidics system that provides them with water and nutrients within four arrays, each connected to eight electrodes that can be used to send and receive electric signals. The behavior of the organoids also can be altered by releasing chemicals into the system that transmit messages between neurons, called neurotransmitters, and others that change neuron activity, called neuromodulators.
However, the system is in its early stages. Fred Jordan, a co-CEO of Final Spark, and his colleagues are still trying to figure out how to electrically stimulate their organoids so that a specific input generates a desired output. Similar to training an artificial deep learning network, it involves using algorithms that learn from large amounts of data. So far, using Boolean logic functions, the team has been able to manipulate its biological network so that it stores one bit of information.
“That means modifying the way the neurons answer in order to change their state for a few hours,” said Jordan.
In an initial proof of concept, Ward-Cherrier and his colleagues were successful in getting the platform to master a touch-related task. They demonstrated it was able to recognize characters of the Braille alphabet.
“We touch every single symbol [with a sensor] and send it to the organoid,” said Ward-Cherrier. “The organoid outputs data, which has been processed, and from the data that’s output, we can classify the different letters of the alphabet.”
Eventually, the team hopes to be able to use the platform as a robot’s brain. For example, data from a tactile sensor could be sent to the organoids for interpretation. They would then send a signal back to a robot to get it to perform a specific movement.
Although Ward-Cherrier is pleased with the platform, adding more electrodes is one improvement that he thinks would help advance their work. The electrodes act as channels that allow them to interact with the organoids by stimulating them and getting readings.
“It would give us more data about what’s happening inside that organoid,” he said. “But it would also allow us to choose the electrodes that are most informative so we can work just with the most useful locations on the organ.”
A similar biocomputer being developed by an Australian company is initially targeting medical researchers studying therapies for brain diseases. Called Cortical Labs, its biocomputer, called CL1, can now be ordered and will be widely available for purchase later this year for about $35,000, or otherwise by paying for access to it via the cloud. It uses human brain cells grown in the lab and derived from stem cells alongside silicon chips.
Cortical Lab’s biocomputer will provide a new alternative to existing methods for testing potential drug therapies. For example, current lab-based methods that use brain organoids in a dish focus on looking for molecular or structural changes to cells. However, using a biocomputer would allow researchers to model how a mini-brain processes and acts on information, which can better predict whether a drug will be effective or not. In some cases, it could replace animal testing, a practice that raises ethical concerns and which may not replicate how a drug would function in a human.
“Neurological and psychiatric diseases are arguably the most impactful class of diseases and we [often] have no real idea [during testing] whether or not a particular drug will help or not,” said Kagan.
In recent work, Kagan and his colleagues tested three anti-seizure medications in a previous version of their biocomputing platform called DishBrain, which made headlines in 2022 when it was trained to play Pong. All of the drugs, which are used to treat people with epilepsy, changed the spontaneous firing of neurons, and one of them, carbamazepine, significantly improved the system’s game-playing abilities when administered in a specific dosage. It was the first demonstration of altered synthetic biological ‘intelligence’ after a drug treatment in a closed-loop simulation outside of the human body.
“This is cool because it basically shows us differences between drugs targeting a disease that we could never discover otherwise,” said Kagan.
Since the technology is so new, Kagan expects many new discoveries to be made once the platform is available for use. The company has already had more sign-ups than expected from startups and researchers wanting to pay for remote access. Just as hobbyists shaped the early development and popularization of personal computers, Kagan thinks a similar approach will lead to rapid improvements in biocomputing.
“There are a lot of people who have ideas and now they’ll finally be able to explore them,” said Kagan. “It will help the whole field grow.”
Sandrine Ceurstemont is a freelance science writer based in London, U.K.
