Predicting the degradation of COVID-19 mRNA vaccine with graph convolutional networks

Wang, Y.

The COVID-19 pandemic has generated a great demand for effective vaccines to control SARS-CoV-2 that causes the disease. In particular, mRNA vaccines against that virus have taken the lead as the fastest vaccine candidates. However, RNA molecules tend to degrade spontaneously, leading to a key challenge to design super stable mRNA molecules for a refrigerator-stable vaccine. In this paper, to accelerate the research of mRNA vaccines, we develop a machine learning method to predict likely degradation rates at each base of RNA molecules in the Eterna dataset, which has been devised for COVID-19 mRNA vaccines. By employing graph convolutional networks with a multi-head attention mechanism, we successfully achieve the goal to predict the degradation of mRNA molecules based on their structure information. After applying K-Fold cross validation and pseudo-label technique, our model reaches a mean columnwise root mean squared error 0.3512 on the Private Test set at Kaggle platform. Such result ranks top 4% on the leaderboard, and can get a silver medal prize in the Kaggle competition. © 2021 ACM.