When attempting to entice a mate, male fruit flies serenade females with a courtship song. Courtship songs are well-known features of birds and teen-romance movies, but they are also indispensible to flies, which beat their wings to create an alluring tune. These ballads are surprisingly complex — flies alternate between two major modes of song, called sine and pulse, and modulate how loudly they sing, making each individual courtship session highly varied. Females are correspondingly choosy, scrutinizing a suitor’s efforts for minutes before making a decision.
Mala Murthy and her collaborators at Princeton University want to know what drives this diversity. The statistics of fly song appear to be random, meaning that variability might simply result from a noisy underlying neural network. But researchers suspected something more sophisticated was taking place, with males tailoring their songs to the object of their affection. “Males make moment-to-moment adaptive decisions about what to sing in response to sensory feedback,” Murthy says.
Understanding this process in fly courtship will shed light on a much broader conundrum: How do animals, including people, alter their behavior using social sensory signals? “We don’t yet have a handle on how nervous systems achieve this amazing flexibility — to rapidly incorporate sensory feedback and change ongoing behavior in response,” she says. Fruit flies’ compact nervous system and molecular toolbox make it possible to connect patterns of neural activity with sensorimotor strategies, she says.
To decipher flies’ complex courtship behavior, Murthy’s team records males’ songs and other behavioral data, such as the fly couple’s positions, and uses statistical models to make predictions about behavior. Male flies use information about how far away the female is and how fast she is moving to decide which song mode to produce at each moment in time, according to a 2014 Nature study analyzing more than 100,000 fly love songs. (To hear a fly love song, watch this video.)
Researchers could account for roughly 70 percent of the variability in the male’s song by considering the sensory feedback he receives during courtship. To do this, they built generalized linear models, a statistical framework that identifies the relationship between sensory cues and motor outputs. Generalized linear models are often used to analyze neural activity, specifically to predict what information is represented in neural codes. But Murthy’s lab was among the first to apply this approach to predict behavioral output from sensory cues.
Researchers noticed that how well the model captured song variability itself seemed to vary over time. “There are times when the model gets it right, and times it gets it wrong,” says Adam Calhoun, a postdoctoral researcher in Murthy’s lab and a fellow with the Simons Collaboration on the Global Brain. Calhoun wanted to better understand why, speculating that the fly changes his song strategies over time. “We know he uses sensory information, but we don’t know if he uses it in the same way all the time,” Calhoun says.
To answer that question, Calhoun, in collaboration with SCGB investigator Jonathan Pillow, employed a combination of two modeling approaches — a generalized linear model and a hidden Markov model, which can identify whether the animal is in different states. (A hungry fly, for instance, reacts differently to food than a sated one.)
On its own, the generalized linear model identifies just one sensory motor transformation, a function that predicts how an animal chooses an action based on sensory input. Adding hidden Markov modeling allows for multiple sensory motor transformations within a given courtship session.
Calhoun identified three basic states in courting males. In one state, the male is close to the female. In another, the male is actively chasing his target. In the third, he is fairly still and often oriented away from his potential mate. Each state has its own sensory motor function, meaning the male uses sensory information to guide his song in distinct ways depending on the state. “In one state, he will sing more of the pulse song when she’s moving faster,” Calhoun says. “In another state, he’ll sing more of the sine song when she’s moving fast.” A model incorporating the three states can account for more than 95 percent of song variability.
The approach provides a new way to explore the role of hidden states in behavior, helping to find patterns in otherwise noisy data. “Overall, it looks like there is no relationship between his song and her speed,” Calhoun says. “But if you break [behavior] into states, there is a relationship.”
“Adam’s work sets the stage for us solving the neural mechanisms underlying internal states,” Murthy says. “This is incredibly exciting.”
Calhoun’s research is part of a growing effort to develop computational approaches to tracking and understanding behavior. “This is one of a handful of powerful techniques people are exploring for characterizing behavior and generating models of it,” says Ben de Bivort, a biologist at Harvard University, who was not involved in the research.
These efforts are still in the early stages. De Bivort speculates that the most successful techniques for analyzing behavior will be those that can successfully integrate neural activity. “People are coming up with lots of clever ideas, and I think it’s an open question which will be most useful to the field,” he says.
Researchers don’t yet know why males seem to have different courtship strategies in different states. To understand that, they’ll need to figure out what female flies find most attractive in a potential partner’s song. Other researchers in the lab are pursuing this question, tracking female behavior in response to different songs. “She likes a lot of song and variety in song, but we don’t have a good sense of what his strategy should be,” Calhoun says.
Researchers also want to understand what pushes flies from one state to another. Sensory information is one driver — when the female is farther away, the male will change his state accordingly. But internal dynamics also contribute. “What is changing in the brain?” Calhoun says. “Is it a whole brain pattern of activity or are small subsets of neurons controlling these states?” They are now examining how activity of individual neurons affects transitions between states.
The researchers are also building a virtual reality setup where they can track neural activity in a moving fly as he responds to a virtual female behaving in different ways. Experimenters can manipulate sensory factors that drive changes in the male’s state, such as the female fly’s distance and velocity, and look for corresponding changes in neural activity.