Phylogenetic-Informed Generative Adversarial Network for Predicting Mutations in SARS-CoV-2

In the pursuit of developing more effective vaccines against viruses like SARS-CoV-2, understanding their potential mutations is crucial. In order to generate protein sequences, this study presents a unique framework that synergistically blends phylogenetic data with Generative Adversarial Networks (GANs). The proposed method combines the strengths of two models: TEMPO, a transformer-based model proficient in leveraging phylogenetic data for mutation detection in SARS-CoV-2, and MutaGAN, a GAN-based model adept in generating synthetic protein sequences. The phylogenetic data from TEMPO is used in this novel design as the input for the MutaGAN generator, replacing the conventional random noise input used in GANs. The generator may now create synthetic protein sequences that are impacted by the evolutionary past included in the phylogenetic data resulting in the creation of artificial protein sequences that are more biologically significant and lifelike. The discriminator portion of the model, similar to MutaGAN, is still in charge of determining whether or not the input sequences comprise a valid parent-child pair and uses sequences produced by the first GAN as extra negative cases. This prevented the model from going too far off its planned route. Additional investigation into this fascinating confluence of bioinformatics and AI may be possible as a result of this study’s conclusions. This approach has tremendous promise in many applications, such as the production of more powerful vaccinations, protein engineering, medication design, and improving our knowledge of biological systems, by forecasting how SARS-CoV-2 could change. A precision of 0.662 and an accuracy of around 0.671, together with a Matthews Correlation Coefficient (MCC) score of 0.313, show that the framework performs very well. These scores show that the framework performs better than a number of cutting-edge baseline techniques. The synthetic protein sequences generated by the framework not only bear higher biological relevance but also exhibit more realistic characteristics. Furthermore, the discriminator component effectively determines the parent-child relationship of input sequences, contributing to the model’s overall performance.