As part of this project, the Institute of Medical Informatics is developing innovative machine learning methods to better understand the evolution of SARS-CoV-2 and reliably predict emerging viral variants. Although AI-based models already achieve high accuracy in classifying viral variants, they often reach their limits when dealing with the ongoing evolution of viruses. New variants can emerge within a short period of time, causing existing models to lose predictive power.

The aim of this project is therefore to develop prediction models that remain reliable even as viral populations change over time, while also providing interpretable insights into their decision-making processes.

A key focus is the combination of modern deep learning techniques with genomic sequence representations known as Frequency Chaos Game Representations (FCGRs). These methods are used to analyze viral genome sequences and identify characteristic features of Variants of Concern (VOCs). By explicitly accounting for temporal developments, we investigate how model performance changes over the course of viral evolution and which approaches enable robust classification over the long term.

In addition, we employ Explainable Artificial Intelligence (XAI) methods to make the decision-making processes of the developed models more transparent and improve our understanding of the underlying genetic features. Combining reliable predictions with interpretable models is intended to help contextualize changes in viral genetic material in the context of evolution and further advance the application of machine learning in bioinformatics research. In the long term, the approaches developed in this project are also intended to be transferable to other pathogens, opening up new perspectives for the genomic analysis and monitoring of viral evolution.

Contact: Dr. Hannah Franziska Löchel, Moritz Grunert

Funding reference number:  LÖ 12 25 05