Artificial Intelligence Improves Risk Prediction of Complex Diseases
Neural network models developed at the HSE AI Research Centre have significantly improved the prediction of risks for obesity, type 1 diabetes, psoriasis, and other complex diseases. A joint study with Genotek Ltd showed that deep learning algorithms outperform traditional methods, particularly in cases involving complex gene interactions (epistasis). The findings have been published in Frontiers in Medicine.
Traditional methods of assessing genetic risk, such as linear regression, often fail to account for intricate gene interactions that influence disease development. These epistatic effects are difficult to detect using classical models, which reduces the accuracy of predictions.
To overcome these limitations, the researchers simulated data incorporating various types of epistasis—additive, multiplicative, and threshold—and trained neural network models on genetic data from over 58,000 individuals of European descent. The study modelled a range of gene interaction scenarios and assessed their impact on disease risk.
The use of deep learning methods, particularly recurrent neural networks (RNN), significantly improved prediction accuracy. The most notable improvement was observed in predicting the risk of type 1 diabetes: the area under the ROC curve (AUC) for the RNN models reached 0.823.
Maria Poptsova
‘Our research demonstrates new possibilities for personalised medicine and prevention. If we can more accurately identify individual risks, it will help doctors to develop more effective treatment and prevention strategies,’ noted Maria Poptsova, Head of the International Laboratory of Bioinformatics.
The study thus confirms the high effectiveness of non-linear machine learning models in predicting genetic risk, paving the way for more accurate personalisation of medical recommendations and therapies.
Alexander Rakitko
‘The genetic passport is becoming an integral part of modern personalised medicine. It is not enough to simply decode a person's genome—we must interpret the results in the most informative way. That is why we are constantly working on training new models to assess the risks of multifactorial diseases. Our joint research shows that neural networks can be effective in this area,’ said Alexander Rakitko, Director of Science at Genotek.
Based on the study, the team at the HSE AI Research Centre developed specialised software—Deep Learning Models for Polygenic Risk Assessment. The programme predicts the likelihood of disease development by analysing individual genomic variations. This tool has already been licensed by Genotek for use in practical genetic research.
