Multi-Instance Contrastive Learning with Binomial K-Mers for Phage Host Interaction Prediction

Understanding the interaction between prokaryotic viruses (or phages) and their hosts provides prospects for applications in pharmaceutical, agricultural, and food industries. Since the advent of high-throughput sequencing technologies, an increasing number of sequenced data has emerged and accelerated the study of phage-host interaction through computational tools. Although their interactions are mostly strain-specific, it is difficult to obtain enough genomic sequences at the strain level to train machine learning models, which could make the models in previous studies overfit training data. In this study, we propose a novel k-mer augmentation technique, called binomial k-mer frequency, in the framework of multi-instance contrastive learning. Our method was evaluated based on the precision recall area under the curve (PR-AUC), outperformed state-of-the-art benchmark methods with statistical significance, and achieved an 11.4 % improvement in prediction performance at the species levels.