Leuven, 14 February 2025 – In a new study published in Science, a Belgian research team explores how genetic switches controlling gene activity define brain cell types across species. They trained deep learning models on human, mouse, and chicken brain data and found that while some cell types are highly conserved between birds and mammals after millions of years of evolution, others have evolved differently. The findings not only shed new light on brain evolution; they also provide powerful tools for studying how gene regulation shapes different cell types, across species or different disease states.
Our brain, and by extension our entire body, is made up of many different types of cells. While they share the same DNA, all these cell types have their own shape and function. What makes each cell type different is a complex puzzle that researchers have been trying to put together for decades from short DNA sequences that act like switches, controlling which genes are turned on or off. The fine-tuned regulation of these switches ensures that each type of brain cell uses just the right genetic instructions from the genome to perform its unique role. Scientists refer to the unique patterns of these genetic switches as a regulatory code.
AI to crack the code
Prof. Stein Aerts and his team at VIB.AI and the VIB-KU Leuven Center for Brain & Disease Research study the fundamental principles of this regulatory code, and how it may impact diseases such as cancer or brain disorders. They develop deep learning methods to help make sense of the huge amount of information on gene regulation they gather from thousands and thousands of individual cells.
"Deep-learning models working with the DNA sequence code have helped us enormously to identify regulatory mechanisms across different cell types," explains Aerts. "Now, we wanted to explore whether this regulatory code could also inform us on how these cell types are conserved across species."
One example of where such a question is highly relevant is in the brain. Despite shared developmental trajectories, the brains of mammals and birds display a strikingly different neuroanatomy. Aerts and his team have now applied deep learning models to assess whether the existing differences and similarities are reflected in shared or divergent regulatory codes.
Tool to study evolution
Nikolai Hecker and Niklas Kempynck, respectively postdoc and PhD student in the Aerts lab, developed and implemented machine learning models to characterize and compare different types of cells across human, mouse, and chicken brains, covering approximately 320 million years of evolution. But before they could truly compare, they first had to better understand the cell type composition of the chicken brain, so they created a comprehensive transcriptomic atlas.
"Our study demonstrates how we can use deep learning to characterize and compare different cell types based on their regulatory codes," explains Hecker. "We can use these codes to compare genomes of different species, identify which regulatory codes have been evolutionarily preserved, and gain insights into how cell types have evolved."
The team found that while some regulatory cell type codes are highly conserved between birds and mammals, others have evolved differently. Notably, the regulatory codes for certain bird neurons resemble those of deep-layer neurons in the mammalian neocortex.
"Looking directly at the regulatory code presents a significant advantage," adds Kempynck, "It can tell us which regulatory principles are shared across species, even if the DNA sequence itself has changed."
Tool to study disease
This regulatory information is useful beyond understanding evolution. In previous work, Aerts and his team already verified that regulatory codes for melanoma (skin cancer) cell states are conserved between mammals and zebrafish. They also identified variants in the genomes of melanoma patients. The models presented in the current study on brain cell types provide useful tools to study the impact of genomic variants and their association with mental or cognitive traits and disorders.
Aerts: "Ultimately, models that learn the genomic regulatory code hold the potential to screen genomes and investigate the presence or absence of specific cell types or cell states in any species. This would be a powerful tool to study and better understand disease."
To the zoo
Aerts and his team are already applying their models on both fronts, he says:
"In collaboration with Zoo Science and Wildlife Rescue Center, we are now expanding our evolutionary modeling to many more animal brains: different types of fish to dear, hedgehogs and capibaras. At the same time, we're also exploring how these AI models can help to unravel genetic variation linked to Parkinson's disease."
