henhao
Sign in
MicrobiologybioRxiv
Preprint

PhageTAILor leverages machine learning for phage tail-like elements detection and classification in plant-associated bacteria

Heejung ChoiD, Socheata Hour, Simon RouxiD, Clément CocletiD, Oluwamayowa AmusatiD, Vivek K. MutalikiD, Alexey E. KazakoviD, Asaf LevyiD, Nimrod Nachmias, Lorenzo Aureli, Tyrome SweetiD, Axel ViseliD, Ruben Michael CeballosiD, Jonelle T. R. BassoiD

Abstract

Phage tail-like elements (PTEs) -- tailocins, bacterial type VI secretion systems (T6SS), and extracellular contractile injection systems (eCIS) -- are contractile nanomachines that bacteria use to kill their neighbors and compete within their micro-ecosystems. PTEs help shape microbial community composition. Most PTE detection tools only detect a single PTE class. Moreover, most tailocin detection methods are largely restricted to Pseudomonas, leaving a key part of tailocin diversity uncharacterized. In this work, we present PhageTAILor (https://github.com/hjcho-bio/PhageTAILor), an integrative and fully automated pipeline that detects and classifies prophages and 3 PTE classes from bacterial genomes. PhageTAILor combines a 6-detector homology-based candidate search (geNomad, tail-gene, PHROGs-tail, SecReT6, eCIStem, and a divergence-tolerant tail-HMM detector) with a LightGBM classifier comprising 1 multiclass and 3 binary heads, trained on 6,501 bacterial genomes carrying 13,082 prophages and PTEs. A phylogeny-free feature matrix used in our model keeps predictions reproducible between model construction and user inference. PhageTAILor performs strongly at the genome level and generalizes beyond its Pseudomonas-rich training set. On a 76-strain cross-clade benchmark, PhageTAILor detected tailocins at F1 = 0.955. Furthermore, it identified 12 of 13 experimentally validated tailocins spanning five genera versus 2 of 13 for a Pseudomonas-restricted tool TattleTail. PhageTAILor also demonstrated sensitivity equivalent to viral detection tool geNomad while avoiding its higher false-positive rate. Applied to 7,925 plant- and soil-associated bacterial isolates, PhageTAILor showed that prophages in the phyllosphere and tailocins in plant-associated bacteria, whereas eCIS are enriched in soil. PhageTAILor is distributed as an open-source, modular pipeline with a command-line interface.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

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

PhageTAILor uses LightGBM on 74 detector features to flag prophages and bacterial phage tail-like elements (tailocins, T6SS, eCIS), recovering 12 of 13 validated tailocins under leave-one-genus-out.

Phage tail-like elements (PTEs), including tailocins, type VI secretion systems (T6SS) and extracellular contractile injection systems (eCIS), mediate bacterial competition and host interactions, but existing detectors cover a single element class or, like the tailocin tool TattleTail, only Pseudomonas. PhageTAILor merges prophage and three-class PTE detection into one model and is explicitly benchmarked on lineages distant from its training data.

Key results

  1. ShownWith each producer's whole genus removed from training, the tailocin head still generalized across genera and missed only the Bacillus subtilis PBSX element, whereas TattleTail detected only Pseudomonas aeruginosa producers.

    PhageTAILor recovered 12 of 13 experimentally validated producers across five genera (0.92), covering every one of the seven non-aeruginosa Pseudomonas strains plus producers in three of the four other genera; TattleTail recovered 2 of 13 (0.15). · Tier 4. Clean-negative precision panel and…; Figure legends

  2. ShownOn the literature-labeled strains of the held-out benchmark, PhageTAILor had the best tailocin F1 of the tools compared. It matched geNomad's recall with fewer false positives and beat TattleTail's recall.

    41 strains (33 tailocin producers): PhageTAILor precision 0.941, recall 0.970, F1 0.955; geNomad recall 0.970, precision 0.865; TattleTail precision 1.000, recall 0.879. One false negative (Aeromonas dhakensis SSU); both false positives were prophage-bearing controls (E. coli K-12 MG1655, Thermus thermophilus HB8). · Comparison with Other PTE-Prediction Tools; Comparison with other PTE prediction tools; Tables

  3. ShownA leakage analysis showed that dropping phylogenetic placement features costs more accuracy than within-genome leakage, and a large gap remains when whole genera are held out. Gradient-boosted trees beat the other classifier families on the same features.

    Per-region macro F1 0.886 (random folds, phylogenetic features), 0.861 (genome-grouped), 0.775 (deployed phylo-free); per-class F1 0.99 negative, 0.35 prophage, 0.84 tailocin, 0.85 T6SS, 0.84 eCIS. Leave-one-genus-out macro F1 0.577. Macro AUPRC: LightGBM 0.765, XGBoost 0.746, Random Forest 0.709, MLP 0.618, logistic regression 0.501. · Model evaluation; Tier 2. Cross-genus generalization…; Algorithm Comparison; Figure legends

  4. ShownEach class was scored on its own benchmark. T6SS performed well in-distribution. Tailocin recall was perfect on a clean-negative panel, but prophages caused some false positives. Recall for held-out eCIS was moderate.

    T6SS AUPRC 0.901, F1 0.92 (genome-grouped cross-validation). Tailocin precision 0.81 at recall 1.00, with no false positives on hard same-clade or structural negatives; the three false positives all carried prophages. eCIS precision 0.83, recall 0.62, F1 0.71, AUPRC 0.77 across eight positives. · Tier 1. In-distribution genome-level…; Tier 3. Held-out cross-clade benchmark; Tier 4. Clean-negative precision panel and…

  5. SuggestedAcross plant- and soil-associated isolates, the raw niche differences in PTE and prophage carriage were mostly explained by taxonomic composition. A few associations held after adjusting for taxonomic order.

    7,925 genomes; T6SS 43%, tailocin 26%, prophage 21%, eCIS 6%. Crude plant vs soil tailocin OR 1.69 (p = 2 × 10⁻²⁴); after order-stratified Mantel-Haenszel adjustment, pooled ORs 1.0–1.1 (p > 0.1). Still significant after Bonferroni: tailocin in plant-associated OR 1.41 (p = 3 × 10⁻³), leaf prophage OR 1.53 (p = 2 × 10⁻³), plant eCIS OR 0.33 (p = 3 × 10⁻⁵). · Case study: PTE and prophage distribution across…; Case study (PTE and prophage distribution across…; Figure legends

Methods, data and reproducibility

Pyrodigal-gv gene calls feed six detectors (geNomad, four MMseqs2 searches against PHROGs-tail, SecReT6, eCIStem and tail-gene databases, and a tail HMM library) whose colocalized hits form candidate regions; each region's 74 features are scored by a four-head LightGBM model, and skani ANI to training genomes sets a confidence band.

  • Curated training genomes from NCBI RefSeq and JGI IMG, quality-filtered with CheckM2 and dereplicated with skani, spanning 24 phyla · bacteria (24 phyla; mostly Pseudomonadota) · n = 6,501 genomes
  • Annotated element intervals: literature-curated tailocins, Mageeney et al. prophages, SecReT6 T6SS and eCIStem eCIS coordinates · bacteria · n = 13,082 (10,284 T6SS; 1,022 eCIS; 891 prophage; 885 tailocin)
  • Held-out cross-clade benchmark withheld from training; tool comparison used its literature-labeled strains · bacteria (multiple genera) · n = 76 strains; 41 labeled (33 tailocin producers)
  • Tier 4 panel: curated tailocin-free genomes in three categories plus experimentally validated tailocin producers · bacteria (five producer genera) · n = 18 negatives; 13 producers
  • JGI IMG isolate genomes grouped by isolation niche (soil, root-, plant-, leaf-associated) for the ecological case study · bacteria (plant- and soil-associated isolates) · n = 7,925 (soil 4,018; root 2,511; plant 846; leaf 550)
  • PhageTAILor data files deposited on Zenodo · 10.5281/zenodo.21152308
Methods & tools
LightGBM v3.3 (four-head classifier), geNomad v1.7.4, MMseqs2 v14, HMMER v3.3 (hmmsearch, custom 597-profile tail HMM library), pyrodigal-gv v0.3.2 (Pyrodigal 3.6.3), skani v0.3.1, GTDB-Tk v2.6.1 with GTDB r226, CheckM2 v1.0.1, SecReT6, eCIStem and PHROGs reference databases, TattleTail (comparator), XGBoost, Random Forest, MLP and logistic regression baselines (scikit-learn), Snakemake with Conda; SciPy v1.11 (Fisher's exact test)
Code
https://github.com/hjcho-bio/PhageTAILor
Software & environment
pyrodigal-gv v0.3.2 (Pyrodigal 3.6.3), MMseqs2 v14, HMMER v3.3, geNomad v1.7.4, skani v0.3.1, GTDB-Tk v2.6.1 (GTDB r226), CheckM2 v1.0.1, LightGBM v3.3, SciPy v1.11; Snakemake with rule-specific Conda environments in the repository.
Assay conditions & compute
About 3 minutes for a typical 5-Mb genome on 8 CPU cores, mostly spent in the geNomad marker search; the bundled skani sketch of the training genomes is 4,164,667,441 bytes.
Reproducibility gaps
Case-study provirus calls came from an earlier Microbial Virus Pipeline run, not a fresh geNomad run; classification needs all six detectors (zero-filled feature blocks cause over-prediction); data are given only as a Zenodo DOI.
Limitations (6)
  • 83.5% of training genomes are Pseudomonadota, and 86% of tailocin producers are Pseudomonas. Sensitivity is therefore best on familiar lineages, and calls on distant phyla should be read alongside the nearest_other_ani confidence band. · stated by the authors
  • There are only eight held-out cross-clade eCIS positives, and the eCIS labels and the eCIStem detector come from the same reference catalog. eCIS metrics therefore measure recovery of reference-like systems, and truly novel element families are out of scope. · stated by the authors
  • Known failure modes: tailocin loci split across draft contigs are missed (the Aeromonas dhakensis SSU false negative had its HMM signal spread over 70 contigs), and the binary heads over-predict when detector feature blocks are zero-filled rather than computed. · stated by the authors
  • The niche conclusions rest on modest, taxonomy-adjusted presence/absence associations. Induction, competition and effector-target experiments are still needed, and the tool does not predict host range, which depends on fast-diversifying tail fibers. · stated by the authors
  • The benchmark is described inconsistently: 31 positives plus 5 negatives with F1 0.98 in one place, and 41 labeled strains with 33 producers and F1 0.955 in another. Cross-genus recall rests on only 13 producers, and no confidence intervals are given. · noted in this summary
  • The case study counts unvalidated calls. Prophage per-region F1 is only 0.35, and prophage tail genes cause tailocin false positives, so the leaf prophage and plant tailocin associations may partly reflect calls swapped between classes. · noted in this summary

Who should read:Read this if you need automated detection of tailocins, T6SS, eCIS and prophages beyond Pseudomonas, or want a worked example of leakage-aware evaluation for genome-level classifiers.

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
THISPais '26Ilchenko '26Albin '23Gauthier '25Campbell '26Hryckowian '20
70% semantic matchBioinformatics
Rayhaan Gerard Pais, W.S. Chen, Sebastián Leptihn+2 more·bioRxiv·Mar 27, 2026