Researchers use computer vision machine learning tools to learn about the behavior of the nematode C. elegans.
PHOTO ILLUSTRATION: ORIGINAL IMAGE COURTESY OF STEVEN FLAVELL
Flavell studies the brain’s internal states—partially hidden variables that shape perception, cognition, and action. An internal state such as sadness or joy represents neuronal activity on multiple levels. To fully understand the infinitely complex brain–behavior link, these levels need to be captured from the first inkling of a feeling to shedding a tear or laughing out loud.
“If we could obtain a predictive understanding of how all the brain’s wiring gives rise to its activity dynamics and an animal’s behavior, then we’d be able to make very strong predictions about the neural pathways that control internal states in the brain,” says Flavell, associate professor of brain and cognitive sciences at The Picower Institute for Learning and Memory at MIT. “They would open the door to eventually designing rational approaches to change circuit activity and behavior.”
Such an understanding could one day inform AI or help temper disruptive thoughts in patients with schizophrenia or other disorders, Flavell says. Diagnosing and treating psychiatric diseases with overlapping symptoms can be challenging, leading to a frustrating process of trial and error for physicians and patients. It’s increasingly apparent that neural circuit dysfunctions underlie many common brain disorders, so the need for neural circuit-based therapies is vast.
A C. elegans approach to a complex subject
Flavell works with the nematode C. elegans, whose simple nervous system is more manageable for scientists to study than the millions of individual neurons in a mouse brain. While researchers had previously mapped each of the roundworm’s 302 neurons, they had only a general idea of which sets of neurons activate internal states such as hunger, the drive to reproduce, and others.
Flavell’s lab built a microscope using AI-powered robotics that tracks and records the one-millimeter-long, transparent worms as they move freely around a glass plate. “We can see the inside of its mouth moving as its chews. We can measure everything the animal is doing,” such as wriggling like a sine wave, laying eggs, or foraging for food. “We then use computer vision—machine-learning tools—to extract all the behavioral parameters of the worm over time,” he says.
A second microscope under the glass plate captures the green glow of proteins engineered to fluoresce when a neuron is active. The result is a mountain of raw data. That’s where MIT applied mathematics PhD students Alexander E. Cohen and Alasdair D. Hastewell ’18 come in.
The mathematics of behavior
Cohen is pursuing a PhD in chemical engineering and computational science and engineering and Hastewell in applied mathematics. They developed tools that compress data representing more than 100 points on the worm’s body into around five key parameters. “We represent all that high-dimensional behavioral data in a more compact, usable form,” Cohen says.
“There are different behaviors the worm will exhibit, and types of shapes” that correspond with those behaviors, he says. “The dynamics between the shapes and one set of behaviors will be different than a new set of behaviors. Hopefully, a lower-dimensional representation will allow us to study those different behaviors more easily.”
The output, Hastewell says, is a grid of numbers that “enables us to use tools from mathematics to understand those behaviors in a way that you can’t do if you just have a movie of a worm.”
Cohen and Hastewell found that shape dynamics—specifically, the undulatory locomotion of not just worms, but also centipedes and snakes—are governed by Schrödinger equations, the fundamental basis of quantum physics. Applied to biological systems, the Schrödinger equation can help characterize and predict the roundworm’s behavioral dynamics.
Flavell says, “It’s cool to interact with Alexander, Alasdair, and [their advisor, MathWorks Professor of Mathematics] Jörn Dunkel, who have these amazing skills and models and tools that are not typically applied in neuroscience.”
In a study published in Cell in 2023, Flavell determined that around a third of the worm’s nervous system can flexibly encode different behaviors under different circumstances, allowing the animal to adapt to a constantly changing environment.
Now, with the latest log of the worms’ brain-wide activity, the researchers’ goal is to translate the visual data into predictive and interpretable models documenting a full set of causal interactions between neurons and behavior.
Down the road, researchers hope to illuminate brain–behavior links in other model animals, such as zebra fish or fruit flies. “I can see, in future years, applying these same methods to more complex animals, humans, and disease states,” Hastewell says.