Fig 1.
Step 1: We collected phage genomes, along with their proteomes (see Section Data collection and preprocessing for more details), and performed preprocessing to obtain the host information (see Section Preprocessing of host information) and select annotated receptor-binding proteins or RBPs (see Section Identification of receptor-binding proteins). Step 2: We fed the RBP sequences to pretrained protein language models to generate meaningful dense embeddings (see Section Representation via protein language models). Step 3: We built a random forest model with the RBP embeddings as the input and the host genus as the predicted output (see Section Classifier building). Step 4: We evaluated our model’s performance (see Section Performance evaluation). Flat icons used in this figure are taken from [43–45].
Fig 2.
Distribution of the lengths of the receptor-binding proteins (RBPs).
(A) Distribution of the lengths of the annotated RBPs selected based on annotation in GenBank and the functional annotation obtained using PHROG [53]. (B) Distribution of the lengths of the annotated RBPs after excluding those longer than 1,587 amino acids. This cutoff was set by defining outlying lengths as those outside the interval [Q1 − 1.5 ⋅ IQR, Q3 + 1.5 ⋅ IQR], where Q1 is the first quartile, Q3 is the third quartile, and IQR is the interquartile range of the RBP lengths. (C) Distribution of the lengths of all the RBPs in our dataset, including those computationally predicted via the approach proposed by Boeckaerts et al. [50].
Table 1.
Statistics on the identification of receptor-binding proteins (RBPs).
GenBank refers to the RBPs annotated in GenBank, PHROG refers to those selected based on the functional annotation obtained using PHROG [53]. Predicted refers to those computationally predicted via the approach proposed by Boeckaerts et al. [50].
Table 2.
Protein language models for generating receptor-binding protein embeddings.
Table 3.
Number of training and test samples for the top 10 hosts associated with the most receptor-binding proteins (RBPs).
The RBPs associated with the top 10 hosts comprise 64.66% of our dataset. Four of the top 10 hosts (Escherichia, Salmonella, Klebsiella, and Erwinia) belong to the same order: Enterobacterales.
Table 4.
Model performance in terms of weighted F1.
The header row refers to the confidence thresholds at which we evaluated model performance; these confidence thresholds range from k = 60% to 100% in steps of 10%.
Fig 3.
Weighted precision-recall curves showing the model performance.
The curves plot the weighted precision against the weighted recall at different confidence thresholds ranging from k = 0% to 100% in steps of 10%.
Table 5.
Weighted F1 scores after integrating handcrafted sequence properties to the vector representations of the receptor-binding proteins.
The selected sequence properties are those with the highest Gini importance after training the phage-host interaction prediction tool by Boeckaerts et al. [15] on our dataset. The header row refers to the confidence thresholds at which we evaluated model performance.
Table 6.
Weighted F1 scores after integrating handcrafted protein sequence properties to the vector representations of the receptor-binding proteins.
The selected protein sequence properties are those with the highest Gini importance after training the phage-host interaction prediction tool by Boeckaerts et al. [15] on our dataset. The header row refers to the confidence thresholds at which we evaluated model performance.
Table 7.
Weighted F1 scores after integrating the top n handcrafted sequence properties to the vector representations of the receptor-binding proteins.
The selected sequence properties are those with the highest Gini importance after training the phage-host interaction prediction tool by Boeckaerts et al. [15] on our dataset. These properties (in order of decreasing importance) are the A nucleotide frequency, GC content, C nucleotide frequency, TTA codon frequency, and TTA codon usage bias. The header row refers to the confidence thresholds at which we evaluated model performance.
Table 8.
Weighted F1 scores after integrating the top n handcrafted protein sequence properties to the vector representations of the receptor-binding proteins.
The selected protein sequence properties are those with the highest Gini importance after training the phage-host interaction prediction tool by Boeckaerts et al. [15] on our dataset. These properties (in order of decreasing importance) are the K (lysine) frequency, isoelectric point, fourth protein Z-scale [60] (which is related to the heat of formation, hardness, electronegativity, and electrophilicity), percentage of residues with exposed solvent accessibility, and molecular weight. The header row refers to the confidence thresholds at which we evaluated model performance.
Fig 4.
t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP) plots of the ProtT5 embeddings, colored based on handcrafted sequence properties.
This figure shows the t-SNE (left of each subfigure) and UMAP (right of each subfigure) projections. Each point corresponds to the two-dimensional projection of a subvector of a receptor-binding protein’s ProtT5 embedding, the components of which are the ℓ components with the highest Gini importance after training our phage-host interaction prediction model (in this figure, ℓ = 100). The points were colored based on the sequence properties with the highest Gini importance after training the model by Boeckaerts et al. [15] on our dataset; these properties are as follows: (A) A nucleotide frequency, (B) GC content, (C) C nucleotide frequency, (D) TTA codon frequency, and (E) TTA codon usage bias (TTA codes for the amino acid leucine).
Fig 5.
t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP) plots of the ProtT5 embeddings, colored based on handcrafted protein sequence properties.
This figure shows the t-SNE (left of each subfigure) and UMAP (right of each subfigure) projections. Each point corresponds to the two-dimensional projection of a subvector of a receptor-binding protein’s ProtT5 embedding, the components of which are the ℓ components with the highest Gini importance after training our phage-host interaction prediction model (in this figure, ℓ = 100). The points were colored based on the protein sequence properties with the highest Gini importance after training the model by Boeckaerts et al. [15] on our dataset; these properties are as follows: (A) lysine frequency, (B) isoelectric point, (C) fourth protein Z-scale [60] (which is related to the heat of formation, hardness, electronegativity, and electrophilicity), (D) percentage of residues with exposed solvent accessibility, and (E) molecular weight.
Table 9.
Comparison of existing machine learning and deep learning tools for predicting phage-host interaction.