A Melbourne-led team has for the first time shown that 800,000 brain cells living in a dish—DishBrain—can perform goal-directed tasks—in this case, the simple tennis-like computer game, Pong. The researchers’ experiments provide evidence that brain cells in a dish can exhibit inherent intelligence, modifying their behavior over time. Future directions of this work have potential in disease modeling, drug discoveries, and expanding the current understanding of how the brain works and how intelligence arises. The findings also raise the possibility of creating an alternative to animal testing when investigating how new drugs or gene therapies respond in these dynamic environments. The researchers’ next aim find out what happens when the DishBrain system is affected by medicines and alcohol.
“We have shown we can interact with living biological neurons in such a way that compels them to modify their activity, leading to something that resembles intelligence,” said Brett J. Kagan, PhD, CSO at biotech start-up Cortical Labs, which aims to build a new generation of biological computer chips. And while scientists have for some time been able to mount neurons on multi-electrode arrays and read their activity, this is the first time that cells have been stimulated in a structured and meaningful way.
Hon Weng Chong, PhD, Cortical Labs CEO, explained to GEN, “The concept for the system arose from the insight of the work from Isomura et al. who proposed a theoretical framework (The Free Energy Principle) developed by Prof Karl Friston as a driving force for learning in a biological neural network.” As Chong noted, “DishBrain offers a simpler approach to test how the brain works and gain insights into debilitating conditions such as epilepsy and dementia.”
Chong, Kagan, and the Cortical Labs team, together with Friston, FMedSci, FRSB, FRS, who is a theoretical neuroscientist at University College London (UCL), and collaborating researchers affiliated with Monash University, RMIT University, UCL, and the Canadian Institute for Advanced Research, reported their developments in Neuron, in a paper titled, “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world.” In their paper, the team concluded: “Using this DishBrain system, we have demonstrated that a single layer of in vitro cortical neurons can self-organize activity to display intelligent and sentient behavior when embodied in a simulated game world.”
Harnessing the computational power of living neurons to create synthetic biological intelligence (SBI), has “previously confined to the realm of science fiction,” but may now be within reach of human innovation,” the authors suggested. Yet while there have been attempts to develop biomimetic hardware supporting neuromorphic computing, “no artificial system outside biological neurons is capable of supporting at least third-order complexity (able to represent three state variables), which is necessary to recreate the complexity of a biological neuronal network (BNN).” And while scientists have made significant progress in mapping “in vivo neural computation,” the team continued, there are technical limits to exploring this in vitro.
For their reported work, the research team developed Dishbrain, which they described as “a system that harnesses the inherent adaptive computation of neurons in a structured environment.” To do this the scientists took mouse cells from embryonic brains as well as some human brain cells derived from stem cells, and grew them on top of microelectrode arrays that could both stimulate them and read their activity. “Through electrophysiological stimulation and recording, cultures are embedded in a simulated game world, mimicking the arcade game ‘Pong’,” the investigators explained.
The researchers connected the neurons to a computer in such a way that the neurons received feedback on whether their in-game paddle was hitting the ball. Electrodes on the left or right of one array were fired to tell DishBrain which side the ball was on, while distance from the paddle was indicated by the frequency of signals. Feedback from the electrodes taught DishBrain how to return the ball, by making the cells act as if they themselves were the paddle.
The scientists monitored the neuron’s activity and responses to this feedback using electric probes that recorded “spikes” on a grid. The spikes got stronger the more a neuron moved its paddle and hit the ball. When neurons missed, their playstyle was critiqued by a software program created by Cortical Labs. This demonstrated that the neurons could adapt activity to a changing environment, in a goal-oriented way, in real time. Reporting their findings, the investigators stated, “The system provides the capability for a fully visualized model of learning, where unique environments may be developed to assess the actual computations being performed by BNNs. This is something that is long sought after and extends beyond purely in silico models or predictions of molecular pathways alone.”
Kagan further explained, “An unpredictable stimulus was applied to the cells, and the system as a whole would reorganize its activity to better play the game and to minimize having a random response … You can also think that just playing the game, hitting the ball and getting predictable stimulation, is inherently creating more predictable environments … We’ve never before been able to see how the cells act in a virtual environment … We managed to build a closed-loop environment that can read what’s happening in the cells, stimulate them with meaningful information and then change the cells in an interactive way so they can actually alter each other.”
The team chose Pong due to its simplicity and familiarity, but, also, Kagan added, because it was one of the first games used in machine learning, “… so we wanted to recognize that.” However, Pong wasn’t the only game they tested. “You know when the Google Chrome browser crashes and you get that dinosaur that you can make jump over obstacles (Project Bolan). “We’ve done that and we’ve seen some nice preliminary results, but we still have more work to do building new environments for custom purposes.”
Friston also noted: “The beautiful and pioneering aspect of this work rests on equipping the neurons with sensations—the feedback—and crucially the ability to act on their world … Remarkably, the cultures learned how to make their world more predictable by acting upon it. This is remarkable because you cannot teach this kind of self-organization; simply because—unlike a pet—these mini-brains have no sense of reward and punishment. The translational potential of this work is truly exciting: it means we don’t have to worry about creating ‘digital twins’ to test therapeutic interventions. We now have, in principle, the ultimate biomimetic ‘sandbox’ in which to test the effects of drugs and genetic variants—a sandbox constituted by exactly the same computing (neuronal) elements found in your brain and mine.”
The research supports the free energy principle developed by Friston. Kagan pointed out, “We faced a challenge when we were working out how to instruct the cells to go down a certain path … We don’t have direct access to dopamine systems or anything else we could use to provide specific real-time incentives so we had to go a level deeper to what professor Friston works with: information entropy—a fundamental level of information about how the system might self-organize to interact with its environment at the physical level. The free energy principle proposes that cells at this level try to minimize the unpredictability in their environment.”
One of the exciting findings was that DishBrain did not behave like silicon-based systems, he continued. “When we presented structured information to disembodied neurons, we saw they changed their activity in a way that is very consistent with them actually behaving as a dynamic system. For example, the neurons’ ability to change and adapt their activity as a result of experience increases over time, consistent with what we see with the cells’ learning rate.”
Chong said the reported discoveries are just the beginning. “This is brand new, virgin territory. And we want more people to come on board and collaborate with this, to use the system that we’ve built to further explore this new area of science. As one of our collaborators said, it’s not every day that you wake up and you can create a new field of science.”
By building a living model brain from basic structures in this way, scientists will be able to experiment using real brain function rather than flawed analogous models like a computer. As the authors concluded in their paper, “Ultimately, although substantial hardware, software, and wetware engineering are still required to improve the DishBrain system, this work does evince the computational power of living neurons to learn adaptively in active exchange with their sensorium. This represents the largest step to date of achieving SBI that responds with externally defined goal-directed behavior.”
Kagan and colleagues aim next to investigate what effect alcohol has when introduced to DishBrain. “We’re trying to create a dose-response curve with ethanol—basically get them ‘drunk’ and see if they play the game more poorly, just as when people drink,” commented Kagan.
That potentially opens the door for completely new ways of understanding what is happening with the brain. “This new capacity to teach cell cultures to perform a task in which they exhibit sentience—by controlling the paddle to return the ball via sensing— opens up new discovery possibilities which will have far-reaching consequences for technology, health, and society,” pointed out Adeel Razi, PhD, director of Monash University’s Computational & Systems Neuroscience Laboratory. “We know our brains have the evolutionary advantage of being tuned over hundreds of millions of years for survival. Now, it seems we have in our grasp where we can harness this incredibly powerful and cheap biological intelligence.”
Kagan continued: “We have also shown we can modify the stimulation based on how the cells change their behavior and do that in a closed-loop in real time … This is the start of a new frontier in understanding intelligence. It touches on the fundamental aspects of not only what it means to be human but what it means to be alive and intelligent at all, to process information and be sentient in an ever-changing, dynamic world.”
Speaking to GEN, Chong further elaborated, “Overall the company believes that there are two long-term goals that we are striving to achieve. First, and related to the alcohol experiments, is the ability to use the DishBrain system as a way of performing cognitive testing of brain cells in a dish. When these brain cells are grown from the stem cells of consenting donors, they would be genetically identical to those in the donor’s brain. Hence they should theoretically respond the same way to various drugs as their donors. Currently, there are a myriad of drugs that affect the central nervous system, and especially the cognitive system, that we currently have no way of testing in a laboratory the side effects before it goes into a human. This results in ~90% of drug candidates that fail because the cognitive effects could not be predicted before live trials and also a significant amount of trial and error that patients undergo and thus resulting in a decrease in quality of life.
“Second, given that we have now proven that it is possible to train the same substance that gives rise to intelligence in the biological world e.g., flies, dogs, octopuses, humans, etc., the question arises of what we might be able to build if we were to increase the complexity of the systems built using this substrate. Will we be able to build machines that exhibit the same strengths that we see in biological intelligent systems such as fluid/flexible intelligence, low power consumption, and high information efficiency? As such our team is currently exploring various ways of improving the performance of these neural systems from better ways of bridging digital information with biological systems, with potential ramifications in the Brain-Computer Interface (BCI) space to ideas on how we could increase complexity such as 3D organizational structures.”
