PBIP: a deep learning framework for predicting phage-bacterium interactions at the strain level

Abstract

Phage therapy has received great attention as a promising antimicrobial treatment, and its core technique, namely predicting phage-bacterium interactions (PBIs), is crucial for understanding infection mechanisms and optimizing therapeutic strategies. However, existing computational methods mainly focus on the species or higher taxonomic levels, and usually neglect the potential of deep embedding representations, limiting their ability to capture complex biological patterns inherent in sequences. This hinders the discovery of rich sequence features, and restricts the clinical application of phage therapy. To address these limitations, we propose a novel deep learning framework (called PBIP) for strain-level PBI prediction. In PBIP, we first identify strain-level interactions through biological infection experiments and sequencing of Klebsiella pneumoniae isolated from the clinical environment of Xiangya Hospital. Then, we utilize a pretrained unified representation model to convert protein sequences of phages and bacteria into deep embeddings. Next, we apply the synthetic minority oversampling technique to generate positive interactions in the embedding space to address the data imbalance issue. Subsequently, we design a deep neural network that uses a convolutional neural network to extract local features, a bi-directional gated recurrent unit to capture global features, and an attention module to highlight significant features. Finally, a fully connected layer integrates this information for PBI prediction. Experimental results show the superiority of PBIP over the state-of-the-art methods in predicting PBIs. The code and datasets are available at https://github.com/a1678019300/PBIP.

Keywords: attention mechanism; deep learning; phage–bacterium interactions; protein representation learning.

Figure 1

The strain-level interaction dataset, showing (A) the phages and bacteria in the training and test sets, respectively, and (B) the similarity between phages in the training and test sets.

Figure 2

The species-level interaction dataset, showing (A) the phages and bacteria in the training and test sets, respectively, and (B) the similarity between phages in the training and test sets.

Figure 3

Overview of PBIP framework for predicting strain-level PBIs. (A) Collecting used strain-level data. (B) Generating protein sequence embeddings using deep representation learning. (C) Utilizing SMOTE to alleviate data imbalance in the embedding space. (D) Developing a deep learning model for predicting PBIs.

Figure 4

Overview of generating a protein sequence embedding using UniRep.

Figure 5

Overview of generating a synthetic sample using SMOTE in the embedding space.

Figure 6

Overview of the proposed deep learning model architecture.

Figure 7

The architecture of the GRU cell.

Figure 8

Performance comparison between PBIP and baseline methods based on ROC and PR curves on strain-level and species-level test sets: (A) strain-level ROC, (B) strain-level PR, (C) species-level ROC, and (D) species-level PR.

Figure 9

Performance comparison between PBIP and baseline methods at different test set imbalance ratios on strain-level dataset.

Figure 10

Performance comparison between PBIP and baseline methods at different similarity intervals on (A) strain-level and (B) species-level datasets.