Predator robots help researchers uncover how larval zebrafish rapidly learn

A novel system developed by Janelia researchers that chases larval zebrafish around an arena with predator robots is enabling scientists to understand how these days-old fish quickly learn in the real world.

Larval zebrafish are an invaluable tool for neuroscientists, who use the tiny, transparent fish to probe how the brain controls behavior, but it’s been difficult for scientists to study learning in these developing vertebrates – an important part of understanding how the brain works.

Previous research found young zebrafish can learn simple associations. But this type of learning happens slowly and often unreliably, and it was still unclear whether days-old zebrafish can learn fast enough to use their memory in natural situations, like recognizing and avoiding new predators.

Janelia researchers led by Postdoctoral Scientist Dhruv Zocchi and Senior Group Leader Misha Ahrens thought that traditional ways of testing learning in larval zebrafish in the lab -- where the conditions were far from what the fish would encounter in the wild -- might not be effective for uncovering how the fish learn.

Instead, the team decided to simulate a real-life experience: being chased by something that does not initially look like a predator. To model this, the researchers used small robotic cylinders, with some programmed to show predator-like characteristics.

Using this novel system, the researchers made the unexpected discovery that not only could larval zebrafish learn robustly and extremely quickly in a more natural context, but they could also do so just five days after beginning their lives as single cells. The researchers showed that the larval zebrafish rapidly learned to recognize non-predator and predator robots and learned to avoid the latter.

“It was an open question: how smart larval zebrafish were in terms of being able to learn rapidly,” Ahrens says. “Dhruv had the right intuition for how to do it and the right audacity to try something very different.”

Simulating real-life learning

In the wild, a zebrafish’s predators aren’t always the same: they can vary from generation to generation as zebrafish and predators migrate. In these situations, the fish need to quickly learn who in their environment to avoid, so the team thought this would be an ideal context for testing the learning capabilities of larval zebrafish.

To simulate this experience, the researchers first placed a robot in an arena with a freely swimming zebrafish. While the robot was stationary, the fish would explore the entire arena, including the area around the robot.

Next, the researchers had the robot chase the fish for about a minute before becoming stationary again. Just one minute of chasing enabled the fish to learn that the robot could be dangerous, resulting in the fish avoiding the area around the robot for more than an hour – a big change from the non-avoidant behavior before the chase experience.

Further, when the researchers introduced a second robot that didn’t chase the fish, the fish would avoid only the robot that was chasing it, showing a well-developed ability to distinguish dangerous from benign entities in the environment.

Together, these experiments suggest that after only about a minute of training, the fish learned to avoid the predator robot, a memory that persisted for more than an hour. This was particularly surprising given the fact that a developing zebrafish larva contains just 1 percent or so of the neurons in its adult counterpart.

“When you are dealing with an organism like the young larval zebrafish, which is still in development and might not yet have its full cognitive capabilities, it turns out you can’t always rely on these more standardized techniques and it’s useful to go back to more naturalistic, ecologically relevant tasks that they can perform,” Zocchi says. “That was the motivation for taking this less standard and, in some sense, messier approach with these robots moving around. But as we saw, that unlocked behavior that we hadn’t seen in the past.”

A multiregional brain network

Whole-brain imaging of the zebrafish brain revealed two linked signals that are required for the fish to learn to recognize and avoid the predator robot.

A fast, teaching signal comes from the fish’s noradrenergic system, with cells in the hindbrain – a region that controls essential functions – responding to the approaching predator. A slower signal distributed across the forebrain – a region associated with learning and planning – encodes the presence of the predator robot. Both regions are necessary for learning, and silencing either one of them removes the ability of the fish to learn. The researchers found that the habenula, a brain area known to be involved in signaling aversive outcomes, was also necessary for learning.

The new work suggests that this multi-regional brain network underlies the ability of young vertebrates to rapidly learn to recognize predators within their first week of life. Because this happens before the fish learns to hunt or accomplish other types of learning, the research suggests that there may be a staggered emergence of different associative learning capabilities and that some abilities that emerge very early – like learning to determine which fish are predators and which are benign – might be the most important learning modalities for survival.

The findings could help scientists better understand how learning happens in brains with large networks of neurons. Neuroscientists are increasingly discovering that even simple learning requires input from across large swaths of the brain, which is difficult to study in other animals but can be accomplished in zebrafish.

“In order to study these more global phenomena, you need systems where you can cover very large spatial distances over the whole brain while at the same time resolving dynamics in single cells,” Zocchi says. “We now have the possibility of probing these things brain-wide in a relatively unbiased manner.”