A new step for artificial intelligence towards personalized medicine

Which of our genes are used at any given time is influenced by several factors. Examples include smoking, eating habits, and environmental pollution. This regulation of gene activity is called epigenetics and can be likened to a switch that controls which genes are turned on and off – without changing the genes themselves.

75,000 human samples

Researchers at Linköping University used epigenetic data from more than 75,000 human samples to train a large number of neural network-type artificial intelligence models. These models are called autoencoders that independently organize information and look for patterns.

Researchers hope that such AI-based models can eventually be used to develop treatments and prevention strategies tailored to individuals.

Artificial intelligence sees if you are a smoker

To test their model, the researchers compared it to existing models. For example, there are already models of the effects of smoking habits on the body. They are based on the fact that some epigenetic changes reverse the effects of smoking on lung function. These effects remain long after a person stops smoking. This type of model can therefore determine whether a person is an active smoker, an ex-smoker, or has never smoked.

Other existing models can estimate people’s chronological age via epigenetic effects or see whether a person can be described as healthy or sick.

The AI ​​models seem to work well

The researchers at Linköping trained their autoencoder and used the results to test age, smoking status and diagnosis of systemic lupus erythematosus.

According to the research results, the researchers’ autoencoder works as well or better than current models.

Our models not only give us the opportunity to classify individuals based on their epigenetic data. We also found that our models found genetic markers that were already known and used in other models, but they also found new markers associated with the condition we were studying, says David Martinez, a doctoral student at Linköping University.

One example is that our smoking model identifies markers that are related to lung cancer, respiratory disease, and DNA damage.

Data is self-organizing

Mika Gustafsson, professor of translational bioinformatics at Linköping University, says:

– We did not control the model and did not have hypotheses based on existing biological knowledge, but we let the data speak for themselves. When we next looked at what happened in the autoencoder, we saw that the data self-organized in a way similar to how it works in the body.

In the next step, researchers can use the most important features found by the autoencoder to create models that can classify for a large number of environment-specific and individual-specific factors, as there is a lack of training data large enough to train more complex AI models.

A look inside the black box

Some types of AI are sometimes likened to a black box that leaves answers, but where humans cannot see how the AI ​​arrived at that answer. Mika Gustafsson and his colleagues seek to create explainable AI models, allowing researchers to understand more.

– We want to be able to understand what the model tells us about the biological factors that cause disease and other conditions. “Then we can not only know whether a person is sick or not, but by interpreting the data we also have a chance to find out why,” says Mika Gustafsson.

Scientific material:

NCAE: Data-driven representations using Network Coherent DNA Methylation Autoencoder identifies robust signatures of diseases and risk factors., Abstracts in bioinformatics.

communication:

Mika Gustafsson is Professor of Translational Bioinformatics at Linköping University
[email protected]