In a new study in Neuron (PDF), researchers from Naoshige Uchida’s MCB lab provide an unprecedented look at how the brain makes foraging decisions—choices between sticking with a dwindling resource or venturing elsewhere. The work, led by former graduate student Michael Bukwich (now a postdoctoral researcher at University College London) and postdoctoral researcher Malcolm Campbell, combined virtual reality, high-density neural recordings, and behavioral modeling to uncover how the brain solves foraging problems by integrating noisy evidence over time.
“Foraging behavior is everywhere in the animal kingdom—and even in our daily lives,” said Campbell. “Anytime we decide whether to continue doing what we’re doing or switch to something new, that’s a stay-or-leave decision. It’s different from choosing between two distinct options. We wanted to understand the neural mechanisms that support this kind of decision.”
To study this, the team designed a virtual reality environment where head-fixed mice ran on a wheel through a digital landscape. Along the way, the mice encountered virtual “patches” that delivered drops of water as a reward. But there was a twist: the reward rate diminished over time, forcing the animals to decide when to leave a patch in search of a better one.
“The key was to simulate a foraging environment where staying too long would cost the animal opportunities elsewhere,” said Bukwich. “We set up patches where the probability of getting water decreased over time. The mice had to figure out when it wasn’t worth waiting anymore.”
The classic framework for such decisions is the Marginal Value Theorem (MVT), a 1970s-era model that predicts that animals should leave a resource patch when their intake rate drops below the environmental average. However, when the team attempted to apply MVT to the mice’s behavior, they discovered inconsistencies—real mouse behavior deviated from these idealized predictions in subtle yet meaningful ways.
To better capture what the mice were doing, the team developed a behavioral model based on integration. In this framework, the mouse builds up a quantity in its brain—a kind of internal counter—over time. Once that integrator reaches a threshold, the mouse leaves the patch. Importantly, each water reward reduces the integrator’s value, effectively setting back the clock from threshold. “More rewards keep the mouse around longer by knocking the integrator back down,” Bukwich explained.
The team then analyzed neural data collected with Neuropixels probes, which can record activity from hundreds of neurons simultaneously across multiple brain regions. They trained models on one part of the data to see if they could predict the integrator’s value from neural activity at later times.
“It was striking,” said Campbell. “We could decode these integrator signals directly from the neural data. We even saw single neurons ramping up over time and then dropping back down in response to rewards—exactly what the integrator model predicted.”
Importantly, the study doesn’t discard the Marginal Value Theorem but instead proposes that integration may be the mechanism the brain uses to approximate the MVT’s predictions.
“The MVT tells us what animals should do in theory,” Bukwich noted. “But it doesn’t tell us how the brain pulls it off. The integrator model gives us a biologically plausible process that could implement those choices.”
The project also overcame some unexpected hurdles. The first neural recordings were done just weeks before the COVID-19 shutdown, and early trials were failing—mice were behaving randomly during recording sessions. “We thought the project might be doomed,” said Bukwich. The breakthrough came when Edward Soucy, Head Neuroengineer of the Neuroengineering Core Facility in the Center for Brain Science, designed a 3D-printed shield to block the mice’s view of the experimenters during setup. Success rates jumped from zero to 100% in a matter of weeks.
Bukwich and Campbell also credit a stunning animation by Kristian Herrera of Mark Fishman’s lab, which visualizes the task and has been used in scientific talks. “It captures the essence of the experiment beautifully,” adds Bukwich.
Looking ahead, the Uchida Lab is continuing to explore how integrator mechanisms might adapt to different environments. “If the environment changes—say, patches are more or less predictable—you might want to tweak your internal integration strategy,” said Campbell. “Other researchers in the lab are now looking at whether the brain adapts its decision-making processes based on these statistical cues.”
By identifying a neural signal that closely tracks a foraging-style integrator model, the study bridges the gap between high-level theories of decision-making and real-time brain activity. Similar activity patterns could be at play in our own brains whenever we ask: should I stay, or should I go?
