ViPER: Virus-Prokaryote intEraction prediction using foundation model with hieraRchical contrastive learning

Prokaryotic viruses, also known as phages, play a crucial role in regulating bacterial populations and influencing microbial ecology. Predicting virus-prokaryote interactions is critical for advancing phage therapy, and enhancing our understanding of microbial ecosystem dynamics. The advent of metagenomic sequencing technologies has spurred the development of numerous computational methods to infer the hosts of novel viruses. However, these methods face serious obstacles, such as limited and imbalanced data, and difficulty capturing complex sequence patterns with long sequences. In this study, we introduce ViPER, a framework integrating a large language foundation model with byte pair encoding and hierarchical contrastive learning to enhance the accuracy and robustness of virus-host association predictions. In order to extract the variable-granularity feature and meaningful subsequence patterns, ViPER leverages byte pair encoding to tokenize viral genome sequences. Then Hyena, a large language model, is applied to extract the intrinsic characteristics of viral genomes. Meanwhile, for host genomes of considerable length, multi-scale k-mer frequencies are embedded into a host feature space by a shared-weight convolutional gated linear unit, which selectively emphasizes important features while suppressing less relevant ones. Aiming to map viral and prokaryotic genomes into a unified feature space, ViPER utilizes hierarchical contrastive learning to draw related pairs closer and push unrelated pairs apart, leveraging hierarchical taxonomic information to ensure robustness across both genus and species levels. Evaluation of benchmark datasets demonstrates that ViPER significantly outperforms existing methods. In real metagenomic applications, ViPER showcases its potential to reveal a comprehensive interaction map.