Teachable Neurons Could Improve Preclinical Drug Screening
By Deborah Borfitz
December 20, 2022 | When sent electrical signals, a living model of the brain succeeded in “playing” Pong, a simple, tennis-like computer game popular in the 1970s. It was a notable feat for a single layer of neural cells in a dish, recapitulating how a real brain functions better than a computer or artificial intelligence ever will, according to Brett Kagan, chief scientific officer of biotech start-up Cortical Labs (Melbourne, Australia).
The disembodied neurons, dubbed DishBrain, represent a new capacity to teach cell cultures to perform goal-directed tasks—in this case, by controlling a paddle to return a ball. As described in a recent article published in Neuron (DOI: 10.1016/j.neuron.2022.09.001), the system can learn and exhibits sentience in a simulated gameplay environment.
While no more than a simple biological structure, DishBrain provides the means to stimulate neural cells in a structured and meaningful way to observe how they process information and respond and adapt to the world around them. Researchers can “harness the innate intelligence that’s inside of neurons,” says Kagan.
Even small neural systems are breathtakingly brilliant, as should be appreciated by anyone who has ever wielded a flyswatter, he notes. “A fly just won’t be killed. Even with the best technology, we can’t build a drone that navigates the environment this well and does it with the barest fraction of energy used by a smartphone.”
With DishBrain, it may be possible to work in a biomimetic sandbox with the same neuronal elements as an actual brain rather than create digital twins to test the effects of drugs and genetic variants, says Kagan. Past models of the brain have typically been based on information technology such as silicon computing.
‘Synthetic Biological Intelligence’
The apparent learning of DishBrain in response to voltage stimulation has been termed “synthetic biological intelligence,” heretofore a concept limited to science fiction and the “least bad option” of all the considered terms to describe the breakthrough, Kagan says. The idea here is to reproduce the key attributes of the brain, outside of anybody, to understand how it enables activity such as movement, food-seeking, and predator avoidance.
DishBrain meets the formal definition of sentience because the neural clusters respond to sensory information by altering their activity, says Kagan. It could enable testing for pseudo-cognitive responses as part of drug screening with neurons (including organoids) in a dish and the development of high-performance silico-biological computational platforms, as well as improve machine learning approaches with a better understanding of how planning happens in the brain.
For the study in the simulated game-world mimicking Pong, 800,000 neural cells were derived from embryonic rodent and human-induced pluripotent stem cells on high-density multielectrode arrays that could both electrically stimulate them—from different firing locations and signal frequencies—and read their activity in a closed-loop environment.
The cultures effectively made their world more predictable by acting upon it, says Kagan, despite having no sense of reward and punishment like a pet or a person. Results were validated across multiple cell types and with different types of feedback setups, he says.
The technology is ideal for personalized medicine, says Kagan, including testing drugs for people with a genetic form of epilepsy. “We can find something that works in a dish and take that forward.”
Time will tell if the approach makes a measurable difference, he adds. But preliminary results look promising and suggest it could better inform treatment decisions.
The study in Neuron was led by Kagan and Andy Kitchen, co-founder of Cortical Labs, which is building a new generation of biological computer chips. Collaborators are affiliated with Monash University, RMIT University, University College London, and the Canadian Institute for Advanced Research.
Separate from the science is the business side of Cortical Labs, which offers access to its Biological Intelligence Operating System (biOS), a simulation environment where researchers can program tasks, challenges, and objectives to learn how brain cells respond to electrical signals. Potential uses include testing a hypothesis, building a better digital health app, or assessing the effects of a drug on the activity of neural cells.
Currently, biOS is being used for preclinical screening of drug candidates to help guide decision-making about which ones to move forward with clinical trials. This may be done in vitro to look at whether cells live or die, the genetic changes they undergo, the proteins they express—and even if a compound alters the firing activity of neurons in a way that’s useful in fighting disease.
Preclinical screening has long been an intrinsically flawed process, says Kagan, citing the less-than-1% success rate on identified novel mechanisms of action for neurological diseases that proceed to human testing. The expectation is that DishBrain will improve the situation by factoring in the information-processing functions of brain cells when calculating the odds.
Cortical Labs is inviting solicitations from developers interested in using biOS for neural computational experimentation, says Kagan. The company will soon be selecting participants for alpha testing of the platform from among applicants who have expressed an interest in using this system, including for novel ideas, via contact with a team member on the website.
Biology Trumps Robots
DishBrain is now being used to assess the effects of medicines and alcohol on neural cells, in collaboration with labs in the U.S. and Europe, says Kagan. “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.”
While still a rudimentary system, DishBrain is nonetheless considered by some to be the “most advanced system that’s been built,” says Kagan, who prefers to compare it to the first-ever transistor. Initially “big and ugly,” transistors now—after seven decades of continuous research and development—populate a smartphone by the thousands.
“We think we will help change the world for the better ... across a number of different industries,” including robotics, in the future, Kagan says. “It will be the result of running rigorous science, hard work, and the multiplier effect of people coming together.”
Cortical Labs, meanwhile, is campaigning around the notion that biological neurons—the basis for every intelligence on the planet and therefore incredibly adaptive—can’t be adequately replicated with robotics. To make the point, visitors to its corporate website can find a robodog prancing stiffly and unnaturally around their computer screen, not at all resembling the grace, beauty, and purpose-driven movement of a real dog.
“Robots can’t interact with the real world in real time in a dynamic way,” Kagan say. “They have to follow incredibly strict, rule-based procedures, so you can’t send a robot into an environment it hasn’t been to before and expect it to be able to achieve anything useful.”
Biology, on the other hand, is amazing, he continues, referencing a cat that has just jumped on his lap and is not at all confused by the fact that the tables and chairs have all been recently moved around. The cat’s ability to reorganize its behavior isn’t shared with even the best machine learning algorithms, which would most likely fail to maneuver around the room even after thousands of training sessions if just one piece of furniture is moved a quarter inch to the left.
Viewed from that perspective, the fact that DishBrain performed much better than chance on a simple game of Pong is not really all that surprising, says Kagan. “Neurons are good at learning ... and have capabilities we still don’t fully understand despite a huge amount of work over decades.”