Lab-Grown Brain Tissue Shows Potential for Problem Solving

A few clusters of lab-grown brain tissue have revealed an interesting possibility: living neural circuits can be motivated to tackle a classic control challenge through structured feedback.

In a closed-loop system providing electrical feedback based on performance, cortical organoids improved their ability to balance an unstable virtual pole, a common engineering task.

Although this doesn’t create a fully functioning hybrid biocomputer, it serves as a solid proof of concept. It suggests that neural tissue can adapt through structured feedback, offering insights into how neurological diseases might affect the brain’s learning abilities.

“We’re trying to get to the core of how neurons can be tuned to solve problems,” says Ash Robbins, a robotics and AI researcher at the University of California (UC) Santa Cruz.

“If we can understand what drives this process in a lab setting, it opens new avenues for exploring how neurological diseases impact learning,” he adds.

The cartpole problem is fairly straightforward. Picture balancing an object like a ruler on your hand: if it’s not perfectly aligned, you have to constantly adjust your hand to keep it upright.

In this iteration, a virtual cart moves left or right to keep a hinged pole balanced. The concept is simple, but small errors can quickly lead to failure, making it an example of an unstable control problem.

This problem is frequently used in reinforcement learning studies because it’s easy to simulate and run, yet it demands continuous, fine-tuning adjustments as opposed to just hitting a single correct answer.

For Robbins and his team, the cartpole was a clear and efficient way to evaluate the capabilities of culturing brain organoids.

The organoids themselves were derived from mouse stem cells developed into clusters of cortical tissue capable of signaling. While they weren’t developed enough for thought or consciousness, they could transmit and receive electrical signals, and their internal connections could be modified based on external stimulation.

The experiment involved a virtual cartpole, where different electrical signals indicated the pole’s tilt. The organoids’ reactions were interpreted as forces pushing the cart to counterbalance the pole’s movement.

Importantly, the organoids had no real grasp of the task at hand. The focus was on whether their neuronal connections could be adjusted via feedback, specifically if bursts of electrical signals could encourage the network to improve its functioning.

Each attempt to balance lasted until the pole tipped beyond a set angle, with performance monitored over five attempts at a time. The organoids were put into three categories: no feedback, random feedback to specific neurons, or adaptive feedback based on previous performance.

The adaptive feedback was the significant part. If performance over five attempts dropped against a recent average, a quick burst of high-frequency stimulation was applied. An algorithm determined which neurons received stimulation, based on whether past patterns had led to better control.

“You can think of it like an artificial coach saying, ‘you’re doing it wrong, adjust this way,'” Robbins explains. “We’re learning the best ways to provide these coaching signals.”

To verify that the organoids were indeed enhancing performance rather than merely experiencing luck, the researchers set a standard based on random performance. If the organoids’ best performances during a session exceeded random expectations, that session was deemed proficient.

The results were notable. Organoids without feedback succeeded just 2.3 percent of the time, while those receiving random signals managed 4.4 percent. Under continuous adaptive feedback, they achieved proficiency in 46 percent of the cycles.

“When we can directly select training stimuli, we can shape the network to resolve the problem,” Robbins adds. “This shows we’re capable of short-term learning, shifting the state of an organoid consistently over time.”

However, “short-term” is indeed accurate. After just 45 minutes of inactivity, the organoids reverted to their baseline performance, essentially ‘forgetting’ their training. Future studies might explore ways to enhance the organoids’ memory, potentially increasing their complexity.

“Ash’s work could foster a larger community centered on adaptive organoid computation. But we want to emphasize that our objective is to advance brain research and treatment for neurological conditions, not to replace robotic controllers with lab-grown animal brain tissues,” states bioinformatician David Haussler of UC Santa Cruz.

“While the latter might seem fascinating, it raises significant ethical concerns, especially if human brain organoids were involved.”

The study findings have been published in Cell Reports.