AI framework decodes how brain regions talk to one another

Neuroscientists have long been able to record brain activity during various behaviors, but they struggled to understand how these regions communicate with each other and identify which regions play the main role. A new artificial intelligence framework called Current-Based Decomposition (CURBD) can now reveal these intricate connections. Developed by Harvard Medical School researchers and their colleagues, CURBD is detailed in the journal Neuron.

According to Kanaka Rajan, an associate professor of neurobiology and co-senior author of the study, "No brain region works in isolation; they're all constantly sending signals to each other, shaping each other's activity." CURBD allows researchers to observe these conversations, identifying which region is directing which and how strongly. This newfound ability to map the brain's internal communication has significant implications for both basic science and clinical applications.

CURBD addresses a limitation of existing tools, which can identify active brain regions but cannot explain what causes those neurons to behave in certain ways or how activity in one region influences activity across the brain. To uncover these signals, CURBD works backward from observed neuronal activity to determine what signals from other regions triggered the activity.

The researchers trained an AI model, a recurrent neural network, to accurately reproduce the patterns of electrical activity in brain regions. Once the model learned to mimic the brain's behavior, CURBD analyzed its internal connections to determine which regions were sending signals to others, the strength of those signals, and the timing of their arrival.

CURBD was tested on neural recordings from various species, including mice, nonhuman primates, and humans, as well as in several disease-relevant conditions. In some cases, CURBD uncovered new insights from recordings originally collected for unrelated research, shedding light on brain circuits underlying conditions like depression and memory loss.

In one experiment, CURBD was applied to data from a model of depression in which animals exposed to inescapable stress gradually stopped moving. CURBD confirmed the role of the habenula in driving this behavioral shutdown but also revealed a previously unknown mechanism in which the telencephalon gradually increased its influence over the habenula as the animal transitioned into passivity.

These findings have implications for understanding treatments like ketamine and deep brain stimulation, which have shown promise in treating severe, treatment-resistant depression. CURBD is now being used to compare depression-related circuit patterns in humans to those observed in animal models, potentially revealing why these treatments are effective at the level of brain circuits.

In another set of experiments, CURBD was applied to recordings from human epilepsy patients with electrodes implanted across multiple brain regions. The researchers found that recalling a familiar memory is driven primarily by signals flowing from the frontal cortex to the hippocampus and amygdala, while encountering a novel image recruits the entire network simultaneously. Understanding these memory-related circuit dynamics could prove valuable in researching conditions like Alzheimer's disease.

Across various experiments, CURBD consistently identified the specific regions driving each behavior, findings that would have remained hidden to previous methods. Rajan concluded, "If we've built something that walks like a duck and quacks like a duck, then the operating principles inside it are likely to match those in the real brain that we can't otherwise access."