Using Explainable AI and Genetic Algorithms to Drive the Discovery of Novel Antiviral Molecules

Abstract

The COVID-19 pandemic has prompted the scientific community to expedite therapeutic drug discovery. The partnership between biomedical scientists and artificial intelligence (AI) specialists has resulted in the creation of several computational tools for the preliminary assessment of antiviral drugs. One such category of drugs comprises antiviral peptides (AVPs) that are part of an organism's initial immune response to viral invasion and infection. The models developed to identify AVPs do not offer any insight into the characteristics that significantly contribute to their antiviral nature. So, a domain expert needs a thorough analysis to choose the most potent AVPs among the ones found by such models in antiviral proteins before going for synthesis. This work aims to accelerate this process by proposing a fully automated in silico framework that identifies AVPs in antiviral protein chains and optimizes them based on some desirable characteristics identified by analyzing the existing AVPs through an explainable machine learning model. This model has been built using extreme gradient boosting and has an accuracy of 90%, which is better than the existing classifiers. The two characteristics found by this explainable model conflict with each other's optimization; hence, non-dominated sorting genetic algorithms are used to find the pareto-optimal AVPs by establishing a trade-off between them. To evaluate the efficacy of the suggested framework, we found and optimized the AVPs present in some well-known antiviral proteins. The pareto-optimal AVPs were identified and proposed for synthesis and validation of antiviral activity. Lastly, a free online app has been deployed at https://avpdesign.anvil.app. © 2023 IEEE.