Antigenicity prediction and vaccine recommendation of human influenza virus A (H3N2) using convolutional neural networks

The rapid evolution of influenza A viruses poses a great challenge to vaccine development. Analytical and machine learning models have been applied to facilitate the process of antigenicity determination. In this study, we designed deep convolutional neural networks (CNNs) to predict Influenza antigenicity. Our model is the first that systematically analyzed 566 amino acid properties and 141 amino acid substitution matrices for their predictability. We then optimized the structure of the CNNs using particle swarm optimization. The optimal neural networks outperform other predictive models with a blind validation accuracy of 95.8%. Further, we applied our model to vaccine recommendations in the period 1997 to 2011 and contrasted the performance of previous vaccine recommendations using traditional experimental approaches. The results show that our model outperforms the WHO recommendation and other existing models and could potentially improve the vaccine recommendation process. Our results show that WHO often selects virus strains with small variation from year to year and learns slowly and recovers once coverage dips very low. In contrast, the influenza strains selected via our CNN model can differ quite drastically from year to year and exhibit consistently good coverage. In summary, we have designed a comprehensive computational pipeline for optimizing a CNN in the modeling of Influenza A antigenicity and vaccine recommendation. It is more cost and time-effective when compared to traditional hemagglutination inhibition assay analysis. The modeling framework is flexible and can be adopted to study other type of viruses.