Phagos is releasing CAPHARD, its proprietary phage-bacteria interactions model, openly on Hugging Face. Researchers and hospitals can query a bacterial genome and get a ranked list of phages likely to infect it, or rank the phages in their own collection.

Phages are natural viruses that infect bacteria and offer a promising approach to tackling antimicrobial resistance. Their specificity presents a practical challenge: a phage that infects one strain may fail against another. Testing every possible combination is practically impossible. CAPHARD is Phagos’ specialized genomic model for predicting phage-bacteria interactions. Given the genomes of a phage and a bacterial strain, it estimates how likely they are to interact, helping researchers prioritize the candidates most worth testing in the lab.

We are releasing CAPHARD on Hugging Face to give researchers, microbiologists and phage banks access to this capability, and to help accelerate progress in phage research.

Phagos collects no data.
The genomes you query with CAPHARD are not sent to us, stored or reused. Nothing leaves your hands unless you choose to share it.

What CAPHARD does

CAPHARD is one component of Phagos’ broader scientific AI platform for phage discovery and design. Our approach combines proprietary phage–bacteria interaction data, specialized predictive models, experimental tools and scientific expertise.

At Phagos, we connect computation to biology: we isolate phages, characterize bacterial strains, sequence both, predict their interactions, validate candidates in the lab and deploy them in the field. Experimental results generate new data that can improve subsequent predictions. This closed loop matters.

CAPHARD is a 900-million-parameter language model that predicts whether a phage infects a bacterial strain, from the genomes of both. It reads entire coding genomes. Most prediction tools rely on a small set of hand-picked proteins, selected with methods specific to one host species. CAPHARD considers all the protein-coding genes of the phage and of the bacterium, with no feature engineering. As a result, it is not tied to a single species or biased toward genes of known function.

How to use CAPHARD

CAPHARD helps researchers prioritize which phage-bacteria pairs to test first. How you use CAPHARD depends on where you start: from phages and bacterial strains you already have, or from a strain you need a phage for.

1. Rank the phages in your own collection

If you already hold sequenced phages and bacterial strains, query CAPHARD with your genomes to obtain an infectivity score for each pair. Use the scores to prioritize candidates for experimental testing.

2. Identify phage candidates for a new bacterial strain

If you are facing a bacterial strain with no phage at hand, query CAPHARD with its genome. It scores the strain against a reference collection of 832 selected phages and ranks them by infectivity score. You know quickly whether a candidate phage is likely to exist, and you can contact the owner of the phage.

In both cases, CAPHARD returns an infectivity score for each phage-host pair: higher scores indicate candidates the model ranks as more likely to interact with. The application accepts FASTA files or FFN files. CAPHARD prioritises and the lab confirms: predictions help reduce the number of pairs that need to be tested experimentally, but they do not replace laboratory validation.

Help us make the next version better

This is our first release, and we want to build on it with the research community. A model is only as good as the interactions it learns from, and the strains that matter most in hospitals are often underrepresented in public datasets.

We want to hear from you:

Share your feedback. Did CAPHARD’s rankings match your bench results? Where did it get it wrong? Which host species should we cover next?

Contribute data, if you wish. Phage-host interaction results with the matching genomes improve the model most. Negative results count as much as positive ones: in public data, only about 3% of tested pairs are infections. Contributions are entirely voluntary, and we will agree on attribution and permitted uses with you before any data is shared.

Every contribution can help inform future training rounds. We plan to release updated versions on Hugging Face and acknowledge contributors in the model card.

Put your phages on the map

A prediction is only useful if the phage can be found. Yet many promising phages remain in lab freezers, unknown to the researchers and clinicians who could benefit from them.

We invite labs and biobanks to deposit their phage genomes in CAPHARD’s reference collection. When CAPHARD ranks one of your phages as a promising match for a bacterial strain, researchers can see who holds it and contact you directly. You keep the physical phage and decide who receives it.

By making phage collections more visible and discoverable, we hope to connect researchers and clinicians with promising candidates they might otherwise never find, and help accelerate the search for phages that could make a difference for patients.

Get started

Try CAPHARD: phagos-biotech-phagos.hf.space

Share feedback or data: caphard@phagos.org

Deposit phage genomes: caphard@phagos.org