Machine learning can predict the effect of genetic mutations

Essentially, machine learning is concerned with developing algorithms that can teach a machine to solve problems based on examples, says Associate Professor Jes Frellsen, whose research shows, among other things, that machine learning can be used to predict the effect of genetic mutations.

Jes FrellsenComputer Science DepartmentResearchartificial intelligence

Written 3 April, 2018 08:09 by Louise Eltard-Larsen

What is your current research about?

All of my research is about artificial intelligence, and more specifically machine learning, where we enable computers to learn something, based on examples. I am interested in fundamental, theoretical issues in the field, but also in how machine learning can be applied - and here, one of my most important focus areas is bioinformatics. For example, you can use machine learning to say something about the effect a mutation in your genome might have. Will it cause a disease, or will it be harmless? In collaboration with the University of Copenhagen, I have looked into how substitutions of amino acids affect the stability of proteins. That is, more specifically, whether a substitution has a destabilizing effect on the protein. If a genetic mutation causes a destabilizing substitution in the protein, it might cause a lower concentration of this protein in your cells; and because the protein has an important function in your body, this will make you ill. By looking at numerous examples of protein structures, a computer program can learn to predict, which substitutions will have a destabilizing effect, and potentially cause illness or disease.

What have you discovered so far?

Normally, when you make these kinds of predictions, you run a computer simulation based on the laws of physics - and these simulations are often very time-consuming. But what we have found out is that, based on examples of protein structures, we can teach so-called neural networks to predict the effect of substitutions. Neural networks, in this context, is a type of mathematical model, inspired by the structure of the brain, which is used in machine learning. And the most interesting result what that these neural networks could actually make these predictions as well as the simulations - only faster.

What do you find most exciting about your field?

It is a fast-moving field, at the moment, which is getting a lot of attention, both en terms of research and from the industry. For example, various health systems are showing a great interest in machine learning. I am actually on an expert committee in the Capital Region of Denmark, because they want to use artificial intelligence in parts of the health system.

This could for example be in terms of automating some processes, so that the decision of a course of treatment for a patient does not depend solely on the individual doctor's knowledge. Rather, through machine learning, you would be able to support the doctor's decision based on a much larger set of data. Another example could be in terms of personalized medicine, where you might use information about people's genomes, and about genetic mutations, to give more targeted medicine. Machine learning can be applied in many important and relevant areas.

More information
Jes Frellsen, Associate Professor, email: jefr@itu.dk

Louise Eltard-Larsen, Research Communications Officer, phone: 7218 5304, email: loel@itu.dk