AI helps better understand how genetic mutations affect our health
As language models learn to interpret words in a sentence, protein Language Models learn how amino acids work together within a protein
Konstantina Tzavella, who used artificial intelligence in her research to better understand how genetic mutations affect the functioning of proteins in our bodies, defends her PhD at the Vrije Universiteit Brussel (VUB). Her work, carried out within the interdisciplinary VUB/UZ Brussel TumorScope project and the VUB/ULB (IB)2 Interuniversity Institute of Brussels, sheds new light on how modern AI models can contribute to predicting disease-related changes in our DNA.
Mutations, small changes in our genetic material, form the basis of evolution but can also lead to diseases such as cancer. Yet the effect of nearly 98 percent of human mutations remains unknown. “We know that mutations play a crucial role in health and disease”, says Tzavella, “but predicting their impact remains a huge challenge.” And that is precisely what her research focuses on.
In her PhD, Tzavella compared existing top-performing methods with a new generation of models inspired by language model technology, similar to how ChatGPT learns to understand language. These so-called “protein Language Models” (pLMs) learn in a similar way the relationships between amino acids, the building blocks of proteins, and can therefore predict how mutations alter a protein’s structure and function.
“Just as language models learn to interpret words in a sentence”, she explains, “pLMs learn how amino acids work together within a protein.” They open a new path to unravel complex genetic interactions.
One of the main challenges is understanding epistasis, the interaction between multiple mutations that often produces an unexpected effect. Most existing methods can barely predict these interactions, but pLMs appear to do better. By constraining pLMs with evolutionary information, Tzavella developed a new computational model that not only performs better in predicting mutation effects but is also applicable in clinical contexts, such as identifying mutations that drive cancer growth.
“Our results show that pLM-based methods are not only powerful but also more flexible”, says Tzavella. “They are less dependent on existing biological knowledge and can yield new insights into unknown genes.”
Her research thus represents an important step toward more reliable predictions of genetic risks and a better understanding of how AI can strengthen biomedical research.
Tzavella, originally from the Greek coastal village of Akrata, studied Electrical and Computer Engineering in Athens and obtained her master’s degree in Biomedical Engineering in Paris. After several years of experience in the pharmaceutical industry, she chose to return to science. “Returning to academia was a leap into the unknown”, she says. “I wanted to build a bridge between technology and medicine, between data and human health.”
With her PhD at the Vrije Universiteit Brussel, Konstantina Tzavella makes a valuable contribution to the future of personalized medicine—a future in which AI and biology are increasingly intertwined.
More info
Konstantina Tzavella: Konstantina.Tzavella@vub.be +33 768 352 522 (En)
Wim Vranken wim.vranken@vub.be +32 488 11 37 27 (Nl)
Frans Steenhoudt
