ProkBERT PhaStyle

ProkBERT PhaStyle predicts whether a bacteriophage follows a virulent or a temperate lifestyle directly from nucleotide sequence. It was built by Balázs Ligeti's group at Pázmány Péter Catholic University, with collaborators at Semmelweis University and the HUN-REN Hungarian Research Centre for Linguistics, and posted to bioRxiv in December 2024. The distinction it predicts is practical: virulent phages lyse their host and are the candidates of interest for phage therapy, while temperate phages integrate into the bacterial genome and can carry toxin, integrase, and antibiotic resistance genes between strains.

Existing lifestyle predictors — BACPHLIP, PhaTYP, DeePhage — depend on curated protein-domain databases or on translating contigs and searching for lysogeny-associated markers. Those pipelines degrade on short, fragmented sequences, which is exactly what metavirome and metagenome assemblies produce, and they need expert effort to maintain as the reference databases grow. PhaStyle replaces that machinery with a fine-tuned genomic language model that reads nucleotides and emits a binary call.

The model is a classification head on ProkBERT, a genomic language model the same group pretrained on microbial sequence. That choice is the paper's central claim: a compact backbone pretrained on the right domain beats much larger general-purpose genomic language models on this task. The authors fine-tuned DNABERT-2, the Nucleotide Transformer, and ProkBERT on identical fragmented datasets so the comparison isolates the backbone rather than the training recipe.

#Key Features

• Nucleotide input, no preprocessing pipeline: Sequences go in as DNA. There is no gene calling, no translation to protein, no HMM scan, and no curated marker database to keep current.
• Built for fragmented sequence: Training and evaluation both use short segments — 512 bp and 1,022 bp — matching the contig lengths that metavirome assemblies actually yield rather than complete phage genomes.
• Out-of-sample generalization as the headline test: Evaluation sets were chosen to be genuinely unfamiliar, including phages isolated from deep-sea, acidic, and arsenic-rich environments whose hosts are extremophiles absent from the training data.
• Small enough to run cheaply: The released checkpoint carries roughly 25 million parameters, up to twenty times fewer than the genomic language models it is benchmarked against, which keeps inference fast enough to screen whole assemblies.
• One-click inference: A Colab notebook and a command-line script take a FASTA file and emit per-sequence predictions, so using the model requires no training infrastructure.

#Technical Details

The backbone is ProkBERT, a BERT-architecture genomic language model of 21–26 million parameters pretrained on microbial genomic sequence, extended here with BertForBinaryClassificationWithPooling for the two-way lifestyle call. Fine-tuning used the BACPHLIP dataset with Escherichia sequences removed — 1,868 training and 246 validation sequences — processed into 512 bp segments to simulate fragmented assemblies. Two held-out collections test transfer: the Guelin collection of 96 taxonomically diverse Escherichia phages with experimentally characterized lytic behaviour, and an extremophile collection of 16 phages from extreme environments.

On the Escherichia test set at 512 bp segments, the ProkBERT-mini variant reaches 0.91 balanced accuracy and 0.83 MCC, matching PhaTYP (0.91 / 0.83) and ahead of DNABERT-2-117M (0.84 / 0.72) and Nucleotide Transformer-500m (0.88 / 0.78). The gap widens where it matters: on the extremophile set at 512 bp, ProkBERT-mini reaches 0.93 balanced accuracy and 0.83 MCC while PhaTYP drops to 0.76 and 0.52, and at 1,022 bp segments ProkBERT-mini reaches 0.96 balanced accuracy with 1.00 sensitivity. Three checkpoints are published, differing in training exclusions; PhaStyle-mini is the canonical release, and PhaStyle-mini-excluded_esch documents the Escherichia-excluded variant used for the generalization experiments.

#Applications

The immediate use is triage in phage biology: screening metavirome assemblies for virulent candidates in phage therapy pipelines, flagging temperate phages that may be shuttling toxin or resistance genes through a microbial community, and characterizing viral fractions of environmental samples where reference coverage is thin. Because the model consumes short fragments, it fits into assembly-first workflows without requiring complete genomes, and because it runs from a fixed checkpoint it can be dropped into existing annotation pipelines alongside conventional tools.

#Impact

PhaStyle is a narrow model — one binary decision — but it makes a general argument cleanly. Fine-tuning a domain-matched genomic language model outperformed models with up to twenty times more parameters trained on broader corpora, and did so most decisively on the sequences furthest from the training distribution. That is a useful counterweight to the assumption that scale is the dominant variable in genomic language modelling. The practical caveats are worth stating: the released checkpoints carry non-commercial licences (CC-BY-NC-SA-4.0 and CC-BY-NC-ND-4.0), the preprint has not been through peer review, and the base ProkBERT model on which everything rests is not itself cataloged here.