Tracking behaviour to understand the brain

“I'm trying to understand the neural basis of what makes us able to learn and move.” This is how Mackenzie W. Mathis, a professor at EPFL, sums up the central question of her research. Her work spans the fields of neuroscience, machine learning and engineering. She teaches mice to play video games while recording their brain activity and behaviour during the process, and critically, she also develops the computer algorithms necessary to analyse the data obtained. Her outstanding work has earned her the Swiss Science Prize Latsis 2024. Faced with this unexpected recognition, she feels “very humbled, especially as the laureates form a long and prestigious line of esteemed scientists, so those are big shoes to fill.”

“If the tools to answer a question don’t exist, we shouldn’t change the question, but rather create those tools.

But Mathis did not wait for this award to be demanding of herself in her research: “If the tools to answer a question don’t exist, we shouldn’t change the question, but rather create those tools.” Her words set the tone for the scientific path she has taken. She has made a name for herself in the field, particularly by developing artificial intelligence-based open-source algorithms that are helpful in neuroscience. One of these algorithms, DeepLabCut, the first version of which was released in 2018, makes it possible to analyse animal videos, ranging from the quivering of a mouse's whiskers, to a horse's gallop to the undulatory movement of an eel. Described as a “breakthrough in life sciences” by the Eric Kandel Young Neuroscientists Prize jury in 2023, the software had been installed more than 700,000 times by the summer of 2024 and continues to make an impact in the field.

A window into the mouse brain

However, her research goes beyond developing software tools: “The people who develop them sit next to the people who use them.” One of the major questions her team is trying to solve is understanding what allows our brains to adapt in real time to changes in the environment. “When you drink coffee, the weight of the cup changes with each sip as the volume is reduced,” she says. “But we don't actively notice this because our brain adapts the strength of our muscles without us even thinking about it.” This adaptability is even more apparent in motor skill learning or sports.

To study this phenomenon, she uses mice that have been genetically modified to be able to observe the activation of their neurons and teaches them to play video games. The mice learn how to move a joystick towards a particular area to get rewards or to navigate in virtual reality worlds displayed on screens. The environment is sometimes modified, for example by applying external forces to the joystick, to see how the animals adapt. During their gaming sessions, the mice are filmed, and their neural activity is recorded.

Linking neural activity to behaviour

This is where the machine learning algorithms that Mathis is helping to develop come into play. To draw conclusions, it is necessary to extract the relevant signals from the recordings of the brain activity and to have quantifiable data on the precise movements of their bodies, hands and even their tiny digits.

The DeepLabCut software makes it possible to track the animals’ movements. And it is further helped by the latest extension released in 2024, which includes “SuperAnimal” models. These make it possible to automate and standardise analyses for species such as mice that are often used in behavioural neuroscience, increasing the reliability of the results and enabling the reuse of data between different research groups.

The aim is then to relate the activity of the neurons to the observed behaviours. To this end, in 2023, Mathis and her team published a new algorithm, called CEBRA, with the idea of identifying the underlying neural dynamics of the brain. This mathematical model makes it possible to decode information from the brain and can combine data from different animals. For example, Mathis and her group were able to use it to find out what a mouse was seeing or where a rat was heading based on records of neuronal activity in their brains. “In the future, this model could lay the foundations for neuroprosthetics in humans to restore vision or mobility after injury,” she says.

Neuroscience is still in its infancy

Although we are beginning to understand in which regions of the brain changes are taking place and when, details of the types of neurons and the precise mechanisms at play are still a mystery. “In neuroscience, we are in a pre-Newtonian era,” Mathis says. “Thanks to artificial intelligence, the networks formed by a few hundred neurons are increasingly well understood. But in one mouse alone, there are 70 million neurons, so there's still a lot of work to do!” But this doesn't discourage her: “For me, science isn’t just a job, it's my life. And it’s a real privilege.” She therefore enthusiastically strives to unravel, one step at a time, the complexity of biological neural networks aided by the lens of artificial intelligence.