ProteoKnight: phage virion protein classification with CNN and uncertainty quantification

Abstract

microbial ecosystems. This has led to their increased utilization in several research areas, such as bacterial genome engineering, phage therapy, disease diagnostics, and viral host identification. The structure of phages is made up of proteins called phage virion proteins (PVP). Classifying these proteins is important for genomic research, which in turn helps us understand the complex interactions between phages and their hosts in the context of making antibacterial drugs. Replacing the tedious traditional procedures, a growing number of computational strategies are being employed to annotate phage protein sequences acquired using high-throughput sequencing. Among these techniques, deep learning approaches demonstrate improved performance in classification outcomes. Such procedures require special sequence encodings for the model to perceive the protein sequences with their distinctive features. Numerous ways have been examined and assessed, while novel methods continue to emerge in order to optimize the task in terms of resource utilization and prediction accuracy. The objective of our work, ProteoKnight, is to explore and develop a unique encoding technique for phage proteins and demonstrate its effectiveness via classification. In our work, we make use of the time-separated PVP dataset that [47] introduced. Furthermore, this study aims to address the lack of research conducted on uncertainty analysis by exploring the domain of uncertainty in binary PVP classification using Monte Carlo Dropout (MCD) method. The experimental findings demonstrate the effectiveness of our strategy for binary classification, achieving a prediction accuracy of 90.2%. However, the accuracy for multi-class classification remains suboptimal. Furthermore, our uncertainty analysis reveals that the class and sequence length show variability in prediction confidence for our suggested classification approach.