Inside the center of the female fruit fly brain, two small regions (red) play a role in aggression. Photo: Robie et al./Cell 2017 via HHMI

With the help of artificial intelligence, a team of researchers has taken another step in the pursuit to answer a fundamental question: how do nervous systems generate behavior?

To find that answer, scientists first need to identify which neurons form the circuits that bring about particular behaviors. Now, an interdisciplinary team has created detailed maps showing specific regions of the fruit fly brain associated with actions like walking, jumping and backing up.

Those maps were developed with the assistance of a machine-learning program that made 100 billion annotations of fruit fly behavior in more than 225 days of videos of flies, a daunting endeavour that would have taken humans thousands of years to manually complete.

A team of biologists and computer scientists at the Howard Hughes Medical Institute’s Janelia Research Campus took a novel approach to probing the inner workings of the fruit fly brain by focusing on the whole brain at once, rather than studying one neuron or a group of neurons at a time. This led to the development of what the Janelia Research Campus calls “the most comprehensive neural maps of behavior yet created.”

“Our goal was to ascribe function to the entire brain of an organism for many behaviors simultaneously,” Alice Robie, a research scientist at the Janelia Research Campus, told Laboratory Equipment. “This hadn’t been done in any organism, and we did it in fruit fly because of the genetic tools that allow us to manipulate the activity of neurons in flies in a way we can’t do in humans.”

The fruit fly, Drosophila melanogaster, is a model organism for studying links between neural activity and behavior due to its relatively compact size. While the brain of a fruit fly is made up of about 100,000 neurons—compared with the estimated 86 billion neurons in the human brain—it has complex circuits that control social and cognitive behaviors.

Still, the research team notes comprehensive mapping of each neuron is difficult given the hundreds of thousands of interconnected neurons and near-infinite variety of behaviors.

With the use of modern genetic tools and machine-vision and machine-learning methods, however, the research team developed what is described as “a brain-wide atlas of behavior-anatomy maps.”

The maps, the authors say, can serve as a guide for future mechanistic and functional studies of the neural substrates of behavior.

The team studied the behavior of more than 400,000 fruit flies from more than 2,200 fruit fly strains that have been genetically modified to allow researchers to activate specific neurons. By turning up the temperature on these genetically engineered GAL4 fly lines, which were also developed at Janelia, researchers were able to study how different sections of the brain contributed to particular behaviors.

“Our project was made possible by the development of thousands of fly lines that each target gene expression to a different small subset of neurons,” Robie said. “These lines allowed us to reliably and repeatedly target the expression of a protein, in our case a temperature-sensitive neural activator, to a small subset of neurons. Thus, when we heated these flies up we could assay the function of each neural population in the generation of behavior.”

For example, turning up the activity on a set of neurons in one fly line caused the flies to jump, while doing so on another line caused male flies to court.

Using a camera suspended above an enclosed arena shaped like a shallow dish, the research team recorded videos of 10 males and 10 females from each of the 2,204 fly lines. In total, the team collected 20,288 videos of flies—a massive amount of data that required an automated approach to process.

The team selected 14 different types of behaviors, including walking, stopping, backing up, chasing, touching and wing flicking, manually labeled the behaviors the flies were performing in a small set of frames and then started teaching a machine-learning base system how to automatically recognize those behaviors.

The system, which is called the Janelia Automatic Animal Behavior Annotator, or JAABA, has a very fast learning algorithm, Robie said, which allowed her to “add labels and retrain quickly and iteratively.”

“First, I trained classifiers that work well on flies without neuronal activation and then I generalized the classifier to work across all the different behavioral phenotypes we observed in our activation screen,” Robie explained. “I used heuristics to find lines with behavior for which the classifier wasn’t working well and added training examples for those behaviors to quickly update the classifier to work on those lines as well.”

JAABA was used to automatically track the positions and annotate the behaviors of each fly in each frame of the more than 20,000 videos recorded. Researchers estimate that manually labeling that immense trove of video would have taken more than 3,800 years.

A valuable resource
With nearly four millenniums of manual work knocked out with the help of artificial intelligence, the researchers then focused on developing maps that illustrate which parts of the fruit fly brain are associated with particular behaviors. The team segmented the fly brain into more than 7,000 regions and then matched those regions to different behaviors.   

The result is comprehensive neural maps that other researchers are calling a valuable resource for further inquiry into how the brain generates behavior.

While some maps show that a broad number of circuits throughout the brain are correlated with a behavior, others are more specific. The map that shows the regions associated with increased chase behavior in female flies, for instance, is limited to a single bilateral region in the protocerebrum. The map for increased walking, however, shows a link with many regions of the brain.

While the researchers say that there are many possible reasons for the density of a map, one potential explanation is the specificity of the behavior.

“There are many reasons a fly might walk. For example, it might walk because it is hungry or because it is chasing another fly,” the authors note.

For this reason, the authors believe that combining behavior statistics using their Browsable Atlas of Behavior-Anatomy Maps, or BABAM, software will be useful in refining maps.

BABAM, a resource the team has made freely available online, allows users to select behavior measures and generate behavior-anatomy maps they can explore in a number of different ways.

“Further analyses of our dataset will reveal more about the neural correlates of behavior and its organization,” the authors wrote in the Cell paper.

And as researchers further advance their understanding of how the fruit fly brain works, experts say it will help scientists figure out the inner workings of more complex animals, such as humans.

“We’ve seen, over and over again, that these mechanisms of brain function are shared across animals from fruit flies to humans,” said Robie. “Understanding how neurons in the brain of fruit fly control its behavior will give us insight into how the human brain controls our behavior.”