henhaobio
Sign in

bioRxivBioinformaticsPreprint

PhageAI - Bacteriophage Life Cycle Recognition with Machine Learning and Natural Language Processing

Piotr TyneckiiD, Arkadiusz GuzińskiiD, Joanna KazimierczakiD, Michał Jadczuk, Jarosław DastychiD, Agnieszka OniśkoiD

PDFDOIData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (12 Jul 2020) · CC BY 4.0

A linear SVM on Word2Vec 6-mer embeddings of phage nucleotide sequences calls phages virulent or temperate with 98.90% cross-validated accuracy, needing no ORF annotation.

Phage therapy needs strictly lytic (virulent) phages, because temperate phages can integrate into bacterial genomes and pass on virulence or antibiotic-resistance genes. Per the paper, the only automatic predictor, PHACTS, depends on imperfect protein annotation and often returns probabilities around 0.5 to 0.6; PhageAI instead classifies from the nucleotide sequence alone.

Key results

  1. ShownA linear-kernel SVM was the best of 11 tuned supervised classifiers trained on averaged Word2Vec 6-mer embeddings of phage genomes.

    Average accuracy 98.90% on the 10-fold stratified validation splits (80%/20%). Training used 278 virulent and 174 temperate genomes plus their reverse complements. RFECV kept 150 of 300 embedding dimensions, and Bayesian optimization tuned hyperparameters on F1-weighted score. The per-classifier figures are in Table 1, which the text does not reproduce. · Results; Train-validation-test split; Reverse complement augmentation; Supervised learning

  2. SuggestedOn a held-out test set of 84 phages from different species and families, the best SVM assigned every phage's life cycle correctly.

    54 virulent and 30 temperate test phages were kept out of training, and the SVM predicted all of them correctly. The abstract also gives 97.18% as average test accuracy, which the Results section describes as averaged across all classifiers. · Abstract; Results; Datasets; Train-validation-test split

  3. SuggestedThe SVM also matched the manual life-cycle calls for every phage in a second, non-public holdout set from Proteon Pharmaceuticals.

    61 phages (49 virulent, 12 temperate) not seen during training; all predictions agreed with the manual life-cycle assignment. · Results; Train-validation-test split

  4. ClaimedExisting tools are described as unreliable for life-cycle calls, and PhageAI is claimed to be more accurate than any existing method.

    PHACTS (a Random Forest ensemble on PHANTOME proteins) reportedly called the same phages virulent and temperate on different runs, often with probabilities around 0.5 to 0.6. VIBRANT v1.2.1 labelled Enterobacteria phage Mu (NC_000929.1), which integrates by transposition, as virulent. No head-to-head accuracy figures are given. · Abstract; Background; Review of existing solutions

  5. SuggestedA UMAP projection of the genome embeddings separates virulent from temperate phages, with much of each class grouped into same-life-cycle clusters and subgroups; it is offered as an aid for interpreting the classifier.

    Shown as an interactive 3D visualization in the tool (Figure 7); no clustering or separation metric is reported. · Limitations and future directions; Figure legends

Methods, data and reproducibility

Phage genomes and their reverse complements (added as augmentation) are split into overlapping 6-mers, embedded with a Word2Vec skip-gram model (300 dimensions) and averaged per genome. RFECV keeps 150 features. 11 classifiers are tuned by Bayesian optimization with 10-fold cross-validation, and a linear SVM performs best.

  • Phage genomes with life-cycle labels from ACLAME and PhagesDB, annotated in DNAMaster and manually checked for lysogeny factors (HMMscan, InterPro, HHpred); unclear phages discarded · bacteriophages · n = more than 600
  • Curated training set · bacteriophages · n = 278 virulent, 174 temperate
  • Held-out test set · bacteriophages from different species and families · n = 84 (54 virulent, 30 temperate)
  • Additional non-public holdout set provided by Proteon Pharmaceuticals S.A. · bacteriophages · n = 61 (49 virulent, 12 temperate)
  • Genome used to test VIBRANT's life-cycle call · Enterobacteria phage Mu · NC_000929.1
Methods & tools
Word2Vec skip-gram (gensim v3.8.1), scikit-learn v0.22.1 (SVC, RFECV), scikit-optimize v0.7.4 (Bayesian optimization), XGBoost v1.0.1, CatBoost v0.22, LightGBM v2.3.1, Biopython v1.76, DNAMaster v5.23.3 (Glimmer, GenMarkS), HMMscan, InterPro, HHpred, PHACTS, VIBRANT v1.2.1, UMAP
Reagents & environment
Biopython v1.76, gensim v3.8.1, scikit-learn v0.22.1, xgboost v1.0.1, catboost v0.22, lightgbm v2.3.1, scikit-optimize v0.7.4, matplotlib v3.2.0; full Word2Vec and SVM hyperparameters printed, random_state = 42.
Assay conditions & compute
Gradient-based classifiers were trained on multiple NVIDIA GPUs; no runtime or memory figures are given.
Reproducibility gaps
The phageai Python package is described as open source, but no repository URL is printed (https://phage.ai/ is the web service and REST API). The 61-phage Proteon Pharmaceuticals holdout set is not public, and Proteon owns the platform.
Limitations (4)
  • The SVM is hard to interpret because its inputs are averaged, multidimensional embedding features; an interactive 3D UMAP view is provided as a partial aid. · stated by the authors
  • Prediction is binary (virulent vs temperate). Chronic-infection prediction for filamentous phages and retraining with phages from more families and species are planned, not done. · stated by the authors
  • No quantitative comparison with PHACTS or VIBRANT on the same phages is reported, so the claim of higher accuracy than existing methods rests on anecdotes (inconsistent PHACTS runs, one VIBRANT miscall). · noted in this summary
  • The ground truth is the authors' own manual annotation, and unclear phages were discarded. Test reporting is ambiguous (all 84 test phages correct, yet 97.18% average accuracy). The 61-phage set is non-public and comes from Proteon, which owns the platform. · noted in this summary

Read this if you need to sort sequenced phages into virulent or temperate (for example, to pick candidates for phage therapy) and want a classifier that needs only FASTA nucleotide sequences, not annotated proteins.

Citations
166

Abstract

BackgroundAs antibiotic resistance is becoming a major problem nowadays in a treatment of infections, bacteriophages (also known as phages) seem to be an alternative. However, to be used in a therapy, their life cycle should be strictly lytic. With the growing popularity of Next Generation Sequencing (NGS) technology, it is possible to gain such information from the genome sequence. A number of tools are available which help to define phage life cycle. However, there is still no unanimous way to deal with this problem, especially in the absence of well-defined open reading frames. To overcome this limitation, a new tool is definitely needed. ResultsWe developed a novel tool, called PhageAI, that allows to access more than 10 000 publicly available bacteriophages and differentiate between their major types of life cycles: lytic and lysogenic. The tool included life cycle classifier which achieved 98.90% accuracy on a validation set and 97.18% average accuracy on a test set. We adopted nucleotide sequences embedding based on the Word2Vec with Ship-gram model and linear Support Vector Machine with 10-fold cross-validation for supervised classification. PhageAI is free of charge and it is available at https://phage.ai/. PhageAI is a REST web service and available as Python package. ConclusionsMachine learning and Natural Language Processing allows to extract information from bacteriophages nucleotide sequences for lifecycle prediction tasks. The PhageAI tool classifies phages into either virulent or temperate with a higher accuracy than any existing methods and shares interactive 3D visualization to help interpreting model classification results.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISUrbanowicz '25Baláž '20Hockenberry '21Marquet '22Thung '21Albin '23
79% semantic matchBioinformatics
Maria Urbanowicz, Arkadiusz Guziński, Zaneta Szulc·bioRxiv·Sep 2, 2025· 5 citations