Facial Expressions Decoded in Mice
Facial Expressions Decoded in Mice

Facial Expressions Decoded in Mice

A combination of videography, machine learning, and brain imaging has identified facial expressions and neural activity patterns in mice that are associated with their emotions.

Ruth Williams
Ruth Williams
Apr 2, 2020

ABOVE: Six emotion-linked facial expressions of mice
JULIA KUHL

If a mouse is in a lot of pain, an experienced handler may see it in the animal’s facial expression—its narrowed eyes and bulging cheeks. But, subtler facial expressions may be more difficult to match to their moods. So researchers developed an unbiased machine learning approach to study hundreds of videos of mice and, as a result, have now catalogued a range of emotion-specific facial expressions. These expressions, the researchers show, can serve as handy readouts for studying the neural basis of emotions.

“It’s a tour de force in terms of techniques,” says neuroscientist Sheena Josselyn of the University of Toronto who was not involved in the research. “Using the techniques . . . they are really beginning to give [emotion] a scientific definition, which I think is really important.”

“The results provide an important advance by adding quantitative analysis of facial motor patterns to the repertoire of ‘emotional’ behaviors that can be measured in mice,” David Anderson, a neuroscientist at Caltech, writes in an email to The Scientist. That’s important, he adds, because “facial expressions have been considered as key indicators of emotion state in mammals, but have previously been measured in rodents only in a more qualitative, subjective manner.” Anderson, who studies the neurobiology of emotional behaviors, was also not involved in the project.

Previous investigations of facial expressions in mice and other animals not only lacked objectivity, they tended to focus on just one or two emotions, says Nadine Gogolla of the Max Planck Institute of Neurobiology. “None of those studies looked at a whole spectrum [of emotions] and whether they can be distinguished from each other.”

To address these issues, Gogolla’s team collected hundreds of videos of the faces of mice—their heads locked in position—while the animals were subjected to stimuli intended to induce particular emotions. These included tail shocks to induce pain, sucrose solution for pleasure, quinine solution for disgust, and injections of lithium chloride to cause malaise. The video data were then processed with machine learning algorithms, which led to the identification of six specific facial expressions in the mice associated with six specific emotions—the four mentioned above plus two types of fear: one associated with freezing, the other with escape. These fear states were not specifically induced by the researchers but occurred spontaneously in some of the mice. The researchers published their findings today (April 2) in Science.

To ensure the identified facial expressions were truly representative of emotions and not just sensory reflexes to the stimuli, Gogolla’s team performed a series of experiments to test whether the expressions displayed key attributes of emotions such as scalability, generalizability, and valence (a measure of good or bad).  

To determine scalability, for example, the researchers dialed up the strength of the tail shocks and the concentrations of the quinine and sucrose and observed that the intensity of the animals’ facial expressions also increased. To determine valence and generalizability, the team gave mice various concentrations of sodium chloride, which at low concentrations is desirable to the animals but at high concentrations is unpleasant. Sure enough, low concentrations of the salt solution elicited expressions of pleasure on the animals’ faces—akin to those seen in the sucrose experiments—while high concentrations caused them to display disgust. This showed that more than one stimulus can produce the same expression (generalizability), and that a single stimulus, depending on concentration, can alter valence.

“The authors demonstrate that these reactions exhibit several of the core features of emotions,” writes Anderson, which “is reassuring.”

The team went on to show that optogenetic stimulation of parts of the mouse brain known as hedonic hotspots—the ventral pallidum and anterior insular cortex—elicited expressions of pleasure, while stimulation of the posterior insular cortex—involved in the processing of aversive states—led to the appearance of disgust. The team also recorded neuronal activity in the brains of mice while the animals were being videoed and given stimuli. These experiments revealed specific sets of cells whose activities correlated with expressions of pleasure or disgust.

While “more work will be required” to understand the meaning of these neural correlates of emotions, writes Anderson, “the authors provide an important platform for investigating this” and other questions.

The work “has moved the field forward because it has given us a tool that as neuroscientists allows us . . . to study internal states of fear and pleasure,” says neuroscientist Mark Andermann of Harvard University who did not participate in the research. “This is really an important step.”

N. Dolensek et al., “Facial expressions of emotion states and their neuronal correlates in mice,” Science 368:89-94, 2020.

Ruth Williams is a freelance journalist based in Connecticut. Email her at ruth@wordsbyruth.com or find her on Twitter @rooph.