PhaTYP: Predicting the lifestyle for bacteriophages using BERT

1 Introduction

• Virulent phages support therapy, while temperate phages can engineer host genomes and functions; costly, slow culturing motivates computational annotation of metagenomic phages.Source: §1 · pp. 1–2
• Only 1,290 virulent and 577 temperate phages had known labels, while metagenomic genomes reach nearly 2 million; incomplete, ambiguous fragments compound prediction difficulty.Source: §1 · p. 2

1.1 Related work

• Marker genes risk low recall on fragments; PHACTS reaches ~60% below 2 kbp, BACPHLIP is for complete genomes, while DeePhage handles 100-bp contigs but tops ~80%.PhagePred instead uses learned k-mer-frequency Markov distances.Source: §1.1 · p. 2

1.2 Overview

• PhaTYP achieves the highest performance on complete genomes and improves short-contig accuracy by over 10% against state-of-the-art methods.Source: §1.2 · p. 2
• PhaTYP uses BERT contextualized embeddings to learn protein composition and associations from phage genomes, evaluating contigs of different lengths and real metagenomic assemblies.Source: §1.2 · p. 2

2.1 Sequence embedding

• Phage proteins are clustered into 63,855 vocabulary tokens, and contigs are translated, assigned their best-alignment tokens, truncated or padded to 298 tokens, then combined with learned token and position embeddings.Prodigal identifies genes; DIAMOND BLASTP matches query proteins to clusters. The 512-dimensional embeddings are trained end-to-end.Source: §2.1 · pp. 3–4

2.2 Model structure

• PhaTYP stacks eight Transformer blocks, each using eight-head attention to learn associations among protein tokens, with feed-forward layers, residual connections, and layer normalization.The Transformer output retains the embedding dimensions, allowing each position to represent latent features for its protein token.Source: §2.2 · p. 4

2.3 Self-supervised training for Masked LM

• Masked-language pretraining uses all available RefSeq phage genomes to learn protein associations by predicting masked tokens, including genomes without lifestyle annotations.The dataset contains 3,474 genomes and 142,434 contigs; 5% of sentence tokens are randomly masked for cross-entropy training.Source: Figure 1 · p. 3, §2.3 · p. 5

2.4 Fine-tuning the model for lifestyle classification

• For lifestyle fine-tuning, PhaTYP classifies phages as virulent or temperate from the [CLS] output, using augmented fragments from 100 bp to 20 kbp.The dataset comprises 1,290 virulent and 577 temperate phages; augmentation yields 160,000 contigs, and a fully connected layer predicts the two classes.Source: Figure 1 · p. 3, §2.4 · p. 6

2.5 Model training

• Both training tasks use ten-fold cross-validation and Adam at a 0.001 learning rate; the lowest-loss pretraining model is fine-tuned, and the best-AUCROC model is retained.Training used four GTX 2080Ti GPUs.Source: §2.5 · p. 6

3.1 Classification performance comparison using ten-fold cross-validation

• PhaTYP had the best ROC performance on complete genomes and higher specificity for identifying temperate phages than the other tools.The ablation study also found that self-supervised learning improved classification accuracy.Source: Figure 4 · p. 7, §3.1 · p. 7, Table 2 · p. 7
• On the low-similarity test set, PhaTYP still had the best results, while test-set similarity affected specificity more than sensitivity and AUCROC decreased for all methods except BACPHLIP.The authors suggest BACPHLIP’s exception may reflect overlap between the test phages and its provided-model training data.Source: Figure 5 · p. 8, §3.1 · p. 8, Table 4 · p. 8
• PhaTYP outperformed the other tools on short contigs, and its accuracy improved with fragment length while self-supervised learning helped especially on short sequences.Training one model with contigs of different lengths and their complete genomes did not affect classification performance, allowing one model to handle multiple length ranges and save computational resources.Source: Figure 6 · p. 9, §3.1 · pp. 9–10

3.2 Predicting the lifestyle of crAssphages

• PhaTYP correctly classified all 33 tested crAssphages as virulent, while PhagePred ranked second and improved relative to earlier experiments.The authors suggest PhagePred’s k-mer frequency feature may be important for crAssphages.Source: Figure 7 · p. 10, §3.2 · p. 10

3.3 Case study: phage lifestyle analysis for infants

• Among 2,291 metagenomic phage contigs, 23 with integrase homology (identity >90%, coverage >90%, e-value <1e-10) were labeled temperate; PhaTYP and DeePhage correctly classified all 23, while integrase classified only ~1%.Source: §3.3 · p. 11, Figure 8 · p. 11
• Predicted temperate-phage proportions decreased as infants grew, and the same trend among newly colonized phages had a smaller p-value; these findings were consistent with the original study.The authors interpret the pattern as a shift from prophages induced from pioneer bacteria toward more environmentally acquired phages and bacteria as infants grow.Source: §3.3 · p. 11, Figure 9 · p. 12
• Delivery type was associated with phage lifestyle composition at month 0, and the difference among delivery types gradually diminished with age.The authors suggest, following, that environmental contamination differences between delivery types may affect the gut microbiome at month 0.Source: Figure 10 · p. 12, §3.3 · p. 12
• Feeding type showed little difference at month 1, but the predicted temperate-phage proportion declined more rapidly by month 4 among formula-fed infants than among mixed-fed infants.The authors propose that breastfeeding may reduce microbial infection and contribute to relatively stable environments for infants with access to breast milk.Source: §3.3 · p. 12

• PhaTYP combines self-supervised protein-association learning with fine-tuning for lifestyle prediction, achieving more stable performance on short contigs than other methods.Source: §4 · pp. 12–13
• In human neonates, PhaTYP’s lifestyle predictions yielded insights mostly consistent with recently published studies, supporting its use to analyze phage composition in metagenomic data.Source: §4 · p. 13
• Future work could distill PhaTYP to reduce its parameter size and training-resource demands, which currently include four 24Gb GPU units.Source: §4 · p. 13
• Another planned extension is incorporating phage identification [35] so users can analyze phage lifestyles directly in metagenomic data.Source: §4 · p. 13
• PhaTYP achieved the most accurate lifestyle prediction among available tools on simulated and real sequencing data.Source: §4 · p. 13