An international research team led by Heidelberg University in Germany and KU Leuven in Belgium has used artificial intelligence to reconstruct how the genetic switches that control development of the cerebellum changed over mammalian evolution. The study, published in the journal Science and announced by the universities on January 30, trained AI models on genomic data from six species and used them to predict the activity of regulatory elements in 240 mammals.
The cerebellum is the brain region associated with movement and balance, and it is also thought to contribute to some higher cognitive functions. Its unusual expansion in humans has long drawn the attention of evolutionary biologists.
A regulatory element near the THRB gene
Among the findings, the researchers identified a previously unknown genetic control element located near the THRB gene, which encodes a thyroid hormone receptor. According to the team, this element allows THRB to become active in cerebellar stem cells, a process they say may have contributed to the growth and complexity of the human cerebellum.
Genetic control elements, sometimes called enhancers, do not encode proteins themselves. Instead they regulate when and where nearby genes are switched on. Changes in these elements, rather than in the genes they control, are widely considered a major driver of evolutionary differences between species.
A stable “grammar” over 160 million years
The researchers also report that the regulatory logic governing the cerebellum’s cell types has remained largely stable across roughly 160 million years of mammalian evolution. In their interpretation, ancient genes were repeatedly repurposed for new functions over time without altering the underlying rules by which they are controlled. That conservative strategy, the authors suggest, may be part of how a brain as complex as the human one could emerge.
Why AI was needed
Measuring regulatory activity directly in hundreds of species is impractical. By training deep learning models on experimental data from six species, the team was able to predict how regulatory sequences would behave in many more, turning a limited experimental dataset into a broad comparative picture. The approach illustrates how machine learning is increasingly being used to answer questions in evolutionary biology that would otherwise require decades of laboratory work.
The authors say the methods could also inform research on neurodevelopmental disorders and other brain conditions, though the study itself is focused on evolutionary history rather than clinical applications.
