Washington, D.C.—Your favorite chatbot may one day run not on silicon microchips, but on real human neurons. Both systems use electricity to compute, transforming inputs into outputs, but neurons—for reasons that are not fully understood—are orders of magnitude more efficient than semiconductor chips. They also require orders of magnitude less energy—not to mention fewer rare earth minerals, no vulnerable supply chains, and negligible water for cooling.
Those advantages have some scientists thinking the data center of the future might sit in a lab dish, leveraging the natural computing abilities of organoids—miniature models of the brain grown from stem cells—or other brain cell configurations. Last month, at the inaugural International Conference on Organoid Intelligence here, researchers discussed that prospect and other ways to harness these tiny tissue blobs for “biocomputing,” electrically stimulating neurons and interpreting their firing patterns via complex physical principles and algorithms.
This nascent field faces a paradox: No one fully understands exactly how the brain works, in physical terms. Even though artificial neural networks, the computational arrangements behind current artificial intelligence (AI) platforms, are modeled on neuronal connections, the analogy may say little about how real neurons compute. But that may not matter, says Brett Kagan, chief scientific officer at the biotech Cortical Labs. “We can build something that is not really like the human brain at all, but still leverages some of the physical properties of neural networks to achieve something new.”
Computing is far from the original purpose scientists envisioned for brain organoids. They have become a standard fixture in neuroscience labs in large part because they offer a window into how the human brain develops, and how that process can go awry. But once these knots of brain tissue reach a certain level of complexity, their neurons begin to exchange electrical signals, which biocomputing researchers are eager to exploit.
Many of the approaches described at the conference work roughly like this: Researchers train a group of interconnected neurons to perform some function—say, guiding an agent through a virtual maze—by giving them electrical or chemical inputs reflecting features of the maze and interpreting their electrical activity. An algorithm transforms an organoid’s firing patterns into actions, such as turning its avatar in a specific direction. Depending on the response, researchers dole out rewards and punishments that, over time, encourage the organoids to make good choices and deter errors.
“We have no idea what they like and dislike,” says Corey Rountree, lead scientist at the biotech company 28bio. But his team has trained pairs of interconnected groups of cells called “connectoids” to play Pac-Man by giving them a rest period if they reach a fruit in the virtual maze and hitting them with a nasty burst of stimulation if they get nabbed by a pursuing ghost.
At a time when new policies at the U.S. National Institutes of Health are discouraging the use of animal models in research, such trained organoids might serve as substitutes, enabling researchers to estimate how genetic mutations, environmental toxins, or drugs would affect an organism’s behavior. Johns Hopkins University molecular biologist Lena Smirnova, for example, is training organoids to play video games and do other tasks to study how such factors are likely to influence cognitive functions such as learning and memory. At the October conference, Smirnova presented unpublished work showing organoids with a genetic mutation linked to autism and epilepsy are more susceptible to developing abnormal firing patterns when stressed by heat, infection, or certain drugs.
But others at the meeting were pursuing a different prospect: using organoids to replace not organs or animals, but actual computers. Fred Jordan, CEO of the biotech FinalSpark, uses combinations of neural signaling molecules such as dopamine and serotonin to control organoids that sit on a bed of electrodes. He concedes such approaches may be “naïve”: Although brain circuits see dopamine as a reward that influences future decisions, for instance, those circuits don’t necessarily exist in organoids.
Still, the workings of real brains hold some lessons. In 28bio’s experiments, for instance, spiking the cells with a molecule called BDNF that improves learning in real brains enhanced their ability to remember their training for longer—up to 48 hours.
Researchers at the meeting also weighed another challenge: defining the relationships between neurons and the algorithms that drive and interpret them. One approach, known as reservoir computing, uses neurons as a largely unchanging “substrate” and trains an algorithm to interpret their complex electrical outputs in the context of some task or goal. It can be difficult to determine whether the organoids or the algorithms are actually doing the calculations, however. Another approach, closed-loop computing, depends on the neurons themselves learning to interpret the stimuli and reorganize their own activity to improve performance on the task. But the way the cells do this might not be easily understandable to human observers.
Despite the uncertainties, biocomputing is starting to look like a junior version of familiar technology. In a preprint published on bioRxiv last month, a group led by Feng Guo at Indiana University Bloomington used organoids to drive a humanoid robot that could chase a laser or move blocks around to spell words. Guo says his team is now exploring ways to use sensory organoids, such as those based on sound-detecting hair cells, to trigger neuronal activity with real-world inputs.
Jordan’s company, meanwhile, provides organoid computing as a service: leasing remote access to labs that want to run their own experiments on FinalSpark’s organoids-on-chips. He says about 20 research teams are using the system. To test the system’s capabilities, one group taught a touch-sensitive robotic arm to read Braille with 83% accuracy, using the robot’s sense of texture as an input and interpreting the organoids’ computation into an output letter.
Cortical Labs sells a biocomputing device called the CL1. It’s a white box containing jars of nutrients and tubes that pump liquid through a thin layer of neurons arrayed on electrodes. At least two companies in Singapore and Australia are already building data centers using CL1 devices. Kagan says they are using the devices for applications such as drug discovery, robotics, and improvements in AI.
Unlike computer chips, each organoid is unique—a source of variability in their behavior that Kagan hopes to turn into a strength. His team has been linking CL1 boxes together to create a hive mind that can “vote” on problems and reach a better answer than a single neural network could. At the meeting, Kagan showed how a hive of 15 machines could control a robot dog. The voting system was 252% more accurate at moving the dog toward a target than a comparable all-electronic machine learning model, he said.
Investigators also hope to tinker with the biology of organoids to make them into better computers. Experiments could help sort out questions such as which brain cell types are necessary for cognition. Support cells called astrocytes, for instance, play an as-yet-unclear role in memory. Besides yielding better biocomputers, solving such mysteries could yield insights about real brains along the way. “If we can understand ourselves,” Kagan says, “that’s a mighty good achievement for humans.”