phold - Phage Annotation using Protein Structures

phold is a sensitive annotation tool for bacteriophage genomes and metagenomes using protein structural homology.

To learn more about phold, please read our manuscript:
https://academic.oup.com/nar/article/54/1/gkaf1448/8415830

Bouras G., Grigson S.R., Mirdita M., Heinzinger M., Papudeshi B.,
Mallawaarachchi V., Green R., Kim S.R., Mihalia V., Psaltis A.J.,
Wormald P-J., Vreugde S., Steinegger M., Edwards R.A.

Protein Structure Informed Bacteriophage Genome Annotation with Phold
Nucleic Acids Research, Volume 54, Issue 1, 13 January 2026
https://doi.org/10.1093/nar/gkaf1448

phold uses the ProstT5 protein language model to rapidly translate protein amino acid sequences to the 3Di token alphabet used by Foldseek. Foldseek is then used to search these against a database of over 1.36 million phage protein structures mostly predicted using Colabfold.

Alternatively, you can specify protein structures that you have pre-computed for your phage(s) instead of using ProstT5 using the parameters --structures and --structure_dir with phold compare.

phold strongly outperforms sequence-based homology phage annotation tools like Pharokka, particularly for less characterised phages such as those from metagenomic datasets.

If you have already annotated your phage(s) with Pharokka, phold takes the Genbank output of Pharokka as an input option, so you can easily update the annotation with more functional predictions!

Tutorial

Check out the phold tutorial at https://phold.readthedocs.io/en/latest/tutorial/.

Google Colab Notebooks

If you don't want to install phold locally, you can run it without any code using one of the following Google Colab notebooks:

To run pharokka + phold + phynteny use this link
- phynteny uses phage synteny (the conserved gene order across phages) to assign hypothetical phage proteins to a PHROG category - it might help you add extra PHROG category annotations to hypothetical genes remaining after you run phold.
Pharokka, Phold and Phynteny are complimentary tools and when used together, they substantially increase the annotation rate of your phage genome
The below plot shows the annotation rate of different tools across 4 benchmarked datasets ((a) INPHARED 1419, (b) Cook, (c) Crass and (d) Tara - see the Phold preprint for more information)
The final Phynteny plots combine the benefits of annotation with Pharokka (with HMM, the second violin) followed by Phold (with structures, the fourth violin) followed by Phynteny

Phold plot Wasm App

We recommending running the web app to generate phold plot genomic maps using WebAssembly (Wasm) in your broswer - no data ever leaves your machine!
Please go to https://gbouras13.github.io/phold-plot-wasm-app/ to use it
You will need to first run Phold and upload the GenBank file via the button
This was built during the WebAssembly workshop at ABACBS2025 - for more, you can find the website here

Recent Updates

v1.2.0 Update (8 January 2026)

Improved ProstT5 3Di prediction throughput for phold run, phold predict and phold proteins-predict due to smarter batching implmentations
Addition of phold autotune subcommand to detect an appropriate --batch_size for your hardware
You can also use --autotune with phold run, phold predict and phold proteins-predict to automatically detect and use the optimal --batch_size (only recommended for large datasets with thousands of proteins)

phold - Phage Annotation using Protein Structures
Tutorial
Google Colab Notebooks
Phold plot Wasm App
Recent Updates
- v1.2.0 Update (8 January 2026)
Table of Contents
Documentation
Installation
Quick Start
Output
Usage
Plotting
Citation

Documentation

Check out the full documentation at https://phold.readthedocs.io.

Installation

For more details (particularly if you are using a non-NVIDIA GPU), check out the installation documentation.

The best way to install phold is using conda via miniforge, as this will install Foldseek (the only non-Python dependency) along with the Python dependencies.

To install phold using conda:

conda create -n pholdENV -c conda-forge -c bioconda phold

To utilise phold with GPU, a GPU compatible version of pytorch must be installed. By default conda will install a CPU-only version.

If you have an NVIDIA GPU, please try:

conda create -n pholdENV -c conda-forge -c bioconda phold pytorch=*=cuda*

If you have a Mac running an Apple Silicon chip (M1/M2/M3/M4), phold should be able to use the GPU. Please try:

conda create -n pholdENV python==3.13  
conda activate pholdENV
conda install pytorch::pytorch torchvision torchaudio -c pytorch 
conda install -c conda-forge -c bioconda phold

If you are have a different non-NVIDIA GPU, or have trouble with pytorch, see this link for more instructions. If you have an older version of CUDA installed, then you might find this link useful.

Once phold is installed, to download and install the database run:

phold install -t 8

If you have an NVIDIA GPU and can take advantage of Foldseek's GPU acceleration, instead run

phold install -t 8 --foldseek_gpu

Note: You will need at least 8GB of free space (the phold databases including ProstT5 are just over 8GB uncompressed).

Quick Start

phold takes a GenBank format file output from pharokka or from NCBI Genbank as its input by default.
If you are running phold on a local work station with GPU available, using phold run is recommended. It runs both phold predict and phold compare

phold run -i tests/test_data/NC_043029.gbk  -o test_output_phold -t 8

If you have an NVIDIA GPU available, add --foldseek_gpu
If you do not have any GPU available, add --cpu.
phold run will run in a reasonable time for small datasets with CPU only (e.g. <5 minutes for a 50kbp phage). With GPU it should complete in under 1 minute.
phold predict will complete much faster if a GPU is available, and is necessary for large metagenomic datasets to run in a reasonable time.
In a cluster environment where GPUs are scarce, for large datasets it may be most efficient to run phold in 2 steps for optimal resource usage.

Predict the 3Di sequences with ProstT5 using phold predict. This is massively accelerated if a GPU available.

phold predict -i tests/test_data/NC_043029.gbk -o test_predictions

Compare the the 3Di sequences to the phold structure database with Foldseek using phold compare. This does not utilise a GPU.

phold compare -i tests/test_data/NC_043029.gbk --predictions_dir test_predictions -o test_output_phold -t 8

Output

The primary outputs are:
- phold_3di.fasta containing the 3Di sequences for each CDS
- phold_per_cds_predictions.tsv containing detailed annotation information on every CDS
- phold_all_cds_functions.tsv containing counts per contig of CDS in each PHROGs category, VFDB, CARD, ACRDB and Defensefinder databases (similar to the pharokka_cds_functions.tsv from Pharokka)
- phold.gbk, which contains a GenBank format file including these annotations, and keeps any other genomic features (tRNA, CRISPR repeats, tmRNAs) included from the pharokka Genbank input file if provided

Usage

Usage: phold [OPTIONS] COMMAND [ARGS]...

Options:
  -h, --help     Show this message and exit.
  -V, --version  Show the version and exit.

Commands:
  autotune          Determines optimal batch size for 3Di prediction with
  citation          Print the citation(s) for this tool
  compare           Runs Foldseek vs phold db
  createdb          Creates foldseek DB from AA FASTA and 3Di FASTA input...
  insta

Phold

Install / Use

README