Harvard Researchers Develop AI-Driven Framework To Study Social Interactions, A Step Forward for Autism Research

Harvard researchers have developed a new artificial intelligence-driven framework to track and analyze how rats interact in social environments, offering a new tool for studying autism and other disorders.

The study, “Mapping the Landscape of Social Behavior,” introduces a methodology that uses machine learning to generate detailed reconstructions of interacting rats, allowing researchers to measure social behaviors with more precision.

Bence P. Ölveczky, an author on the study and professor of Organismic and Evolutionary Biology at Harvard, said the research pioneers a “new method for characterizing social behavior in freely interacting pairs and triads of animals.”

“We think of this as an opportunity for the scientific community to start looking in a more rigorous and quantitative way into these social behaviors,” he said. “We are interested in understanding the neural mechanisms that underlie these social behaviors and differences in social behavior.”

Researchers developed a new AI model to analyze hours of recorded interactions between rats, consolidating hours of work into minutes.

“This is basically a machine learning-based algorithm where we train the network and neural network to recognize 3D pose from 2D video cameras very accurately, and then we were able to extend this to two and even three animals interacting,” Ölveczky said.

He added that prior to this model, researchers would individually analyze the videos and “hand annotate” each rat interaction.

“This would a be extremely subjective, be extremely tedious and time-consuming,” Ölveczky said.

Ugne Klibaite, a postdoctoral fellow and first author on the study, said rats offer a uniquely social behavior that researchers leveraged during the study.

“The one thing about mice is that they do not have this beautiful, rich, affiliated behavior that rats do,” she said.

Ölveczky noted that the “real incentive” for the research was to better understand autism spectrum disorder models. Researchers used rats that were genetically engineered to remove the gene associated with the disorder to study differences in social behavior.

“We were interested in whether their social behavior was in any way affected because we know that in human autism, there is indeed differences in social behavior on the spectrum,” Ölveczky said. “We were wanting to validate whether the rat social behavior is a good model.”

Klibaite said that the findings supported the rat social behavior model, citing the AI analysis tracking a wide-ranging difference in rats’ social interactions.

“In the autism model cohorts, some of the knockouts become what we would call more social, and some become less social,” Klibaite said of the genetically modified rats.

While Klibaite said that the results may appear “inconsistent” at first glance, the findings ultimately prove the expected broad range of the rats’ social interactions.

“They’re all different in different directions — and we’re not surprised by that,” she said.

These variations in rat models could be critical for identifying relevant behavioral traits that mirror those with autism, though the researchers caution against over-interpretation.

“These animals are not autistic,” Klibaite said. “I never want to make any sort of claims that if we understand the rat, we can understand autism. But I think this type of basic behavioral work is really important.”

More important than the findings, Ölveczky said the study’s methodology and AI model can be applied beyond the research to study disorders in animals and humans.

“The methodology, as you say, can be applied broadly,” he said. “It can be applied at the level of single animals, can be applied to Parkinson’s, can be applied to movement disorders.”

Klibaite said she hoped future researchers will use the study’s data and tools to advance their own research.

“We’re in an era now of so many different types of tools that it would be really nice to use the different tools on the same data set,” she said.

—Staff writer Ella F. Niederhelman can be reached at ella.niederhelman@thecrimson.com. Follow her on X @eniederhelman.

— Staff writer David D. Dickson can be reached at david.dickson@thecrimson.com.