As the most ubiquitous and abundant viral community, bacteriophages (phages for short) play a vital role in regulating the ecological balance by infecting bacteria and archaea. Phages can be classified into two types based on their lifestyles: virulent phages and temperate phages, which are closely related to their functional characteristics and influence their interaction patterns with hosts. Therefore, identifying phage lifestyle is critical for understanding the mechanisms by which phages infect bacteria and represents a key step in mastering their functions and potential applications. In this paper, we propose a novel method for phage lifestyle identification by considering two perspectives. One perspective is based on the genomic sequences of phages, in which both local and global semantic features are integrated. The other perspective focuses on the host quorum sensing phenomena that influence phage lysogen-lysis decisions. Specifically, we first capture local sequence variation patterns by extracting the relative positional information of nucleotide fragments at different intervals, which enables robust representation of local genomic semantics. Secondly, the pretrained nucleotide language model DNABERT is applied to capture the semantics of genome sequences by considering the global contextual information. Finally, combined with quorum sensing signals from the bacterial host, a final fusion representation is obtained, which is fed into a predictive model to identify the phage's lifestyle. Experimental results show that our method has excellent and stable performance in both phage complete genome and short contigs from metagenomic data. We also investigate early-life viral colonization in the human gut metagenome, further validating the model's generalizability and real-world applicability.
Xiao et al. (Thu,) studied this question.