Phagos Launches CAPHARD, a 900-Million-Parameter AI Model to Predict Phage–Bacteria Interactions

On October 20, 2026, French biotechnology company Phagos announced the public release of CAPHARD, a genomic artificial intelligence model designed to predict interactions between bacteriophages and bacterial strains. Accessible through Hugging Face, the platform enables researchers to evaluate the compatibility of sequenced phages with bacterial isolates or search a reference collection of 832 bacteriophages for candidates potentially capable of infecting a particular strain. The release follows several years of development supported by European supercomputing resources and represents a new application of large language models to phage discovery, microbiology and antimicrobial resistance research.

The identification of suitable bacteriophages remains one of the principal technical challenges in developing personalized phage therapies. Unlike conventional broad-spectrum antibiotics, many phages exhibit highly restricted host ranges, sometimes distinguishing between closely related bacterial strains belonging to the same species. This specificity reflects complex interactions between viral attachment proteins, bacterial receptors, intracellular defense systems and mechanisms controlling viral replication. Consequently, identifying an effective phage generally requires experimental screening against the target bacterial isolate, a process that becomes increasingly demanding as the number of available phages and bacterial strains increases.

Phagos developed CAPHARD to reduce this experimental burden by predicting potential interactions directly from genomic information. Rather than attempting to establish infectivity without laboratory evidence, the model provides a ranking of phages that can guide subsequent testing, allowing researchers to concentrate experimental resources on candidates considered more promising by the computational analysis.

A 900-Million-Parameter Model Analyzing Complete Coding Genomes

CAPHARD, whose full name is Codon-Assisted Phage-Host Automatic inteRaction Discovery, is a specialized language model containing approximately 900 million parameters. Its central objective is to predict whether a bacteriophage can infect a bacterial strain using the genomic sequences of both organisms.

The approach differs from prediction methods that rely primarily on a restricted set of previously characterized proteins or biological features. Many conventional strategies use receptor-binding proteins, phage tail structures, sequence similarities or other molecular characteristics associated with host recognition. Although these approaches can be effective, their predictive performance may depend on the availability of functional annotations and on the bacterial species being investigated.

CAPHARD instead considers the complete protein-coding gene repertoire of both the phage and its bacterial host, without requiring researchers to manually select specific predictive features. By incorporating information distributed throughout both genomes, the model is designed to identify genomic patterns associated with infection, including those involving proteins whose biological functions are not yet understood.

This broader genomic representation is particularly relevant because phage susceptibility depends on considerably more than the initial attachment of a virus to a bacterial receptor. Successful infection also requires the introduction of viral genetic material, evasion or inhibition of bacterial defense mechanisms, expression of phage genes and production of infectious viral progeny. Restriction-modification systems, CRISPR-associated immunity, abortive infection mechanisms and other antiphage defenses can prevent productive infection even when initial adsorption occurs successfully.

By analyzing both genomic partners, CAPHARD attempts to capture combinations of biological characteristics that influence these interactions. Its architecture is not intrinsically restricted to a single bacterial species, although actual predictive coverage remains dependent on the available training data and the confidence associated with individual bacterial hosts.

The model forms part of Phagos' broader computational and experimental platform, which combines genomic sequencing, proprietary phage–bacteria interaction datasets, artificial intelligence and laboratory validation. Within this system, experimentally observed interactions can be incorporated into subsequent training cycles, progressively refining the predictions generated by computational models.

Identifying Compatible Phages Through Two Prediction Modes

The public CAPHARD application offers two complementary approaches depending on whether researchers already possess a phage collection or are searching for candidates against a newly characterized bacterial strain.

The first mode, called Pair Prediction, allows users to submit the genome of a bacterial host together with the genomic sequences of one or several bacteriophages. The model evaluates these combinations and provides an infectivity score for each phage–host pair. Researchers can subsequently compare the scores to prioritize phages for experimental investigation, potentially reducing the number of combinations requiring conventional host-range assays.

This approach could be useful for laboratories maintaining sequenced phage collections, particularly when screening multiple bacterial isolates associated with an infection or evaluating possible components of a therapeutic phage cocktail. Computational ranking may help organize these experiments, although interactions predicted by the model must still be confirmed using appropriate microbiological methods.

The second mode, called Compatible Phages, addresses situations in which a bacterial strain has been isolated but no suitable bacteriophage is immediately available. Researchers can upload the bacterial genome and compare it against CAPHARD's reference collection of 832 selected bacteriophages. The model assigns scores to the possible interactions and returns a ranked selection of candidates.

This function potentially extends the usefulness of genomic prediction beyond laboratories possessing extensive phage libraries. When a promising phage belongs to an external collection, researchers may be able to identify the institution holding the corresponding biological material and contact its owner to request access for further characterization.

Both prediction modes accept genomic sequences in FASTA or FFN format. Higher infectivity scores indicate interactions ranked more favorably by the model, but these scores should not automatically be interpreted as experimentally established infection probabilities. Furthermore, identifying a phage capable of infecting a bacterium does not necessarily establish its therapeutic suitability, which also depends on properties such as lytic activity, genomic safety, resistance development, formulation stability and the biological environment in which it will be administered.

Phagos also specifies that the initial version of CAPHARD does not provide predictions for every possible bacterial host. The public application restricts predictions to organisms for which the model is considered sufficiently reliable, with broader coverage expected as additional genomic and experimental data become available.

Predictive Performance Evaluated Across 1,920 Phage–Host Pairs

Alongside its release, Phagos disclosed performance measurements obtained from an internal evaluation dataset containing 1,920 phage–host combinations. The reported results demonstrate how prediction thresholds influence the balance between precision and recall, two important parameters when selecting candidates for experimental screening.

At an infectivity score cutoff of 0.5, CAPHARD achieved a precision of 45.1% and a recall of 63.7%. Precision represents the proportion of interactions classified as positive that were actually positive in the evaluation dataset, while recall indicates the proportion of experimentally positive interactions successfully identified by the model.

Increasing the cutoff to 0.8 improved precision to 55.8%, while recall decreased to 48.0%. At the highest reported threshold of 0.9, precision reached 64.5%, but recall declined to 31.1%.

These results illustrate a trade-off between selecting fewer candidates with greater confidence and identifying a larger proportion of potentially compatible phages. A relatively permissive threshold may be appropriate when researchers wish to avoid excluding promising bacteriophages, whereas a stricter threshold may be useful when experimental capacity is limited and testing must be concentrated on a smaller number of candidates.

At a threshold of 0.9, for example, approximately two-thirds of the interactions predicted as positive were confirmed as positive in the internal dataset. However, the model identified only about one-third of the total positive interactions, meaning that many potentially useful phages would be excluded if this threshold were used as the sole criterion for candidate selection.

These measurements therefore require careful interpretation. Precision and recall are influenced by the composition of the evaluation dataset, including the proportion of positive and negative interactions, the diversity of bacterial strains and the characteristics of the tested bacteriophages. The reported internal results do not establish equivalent performance for every bacterial species or demonstrate that the model has undergone independent clinical validation.

The figures also differ from the broader performance claims previously associated with Phagos' artificial intelligence platform. In January 2026, the World Economic Forum reported that the company's AI-based approach had achieved 95% accuracy and accelerated phage discovery by up to tenfold compared with manual methods. However, the publicly available description does not provide sufficient methodological details to equate this accuracy figure with CAPHARD's precision or recall on the 1,920-pair evaluation dataset. The two sets of results should consequently be considered separately rather than presented as directly comparable measures.

The practical contribution of CAPHARD will ultimately depend on its ability to reduce experimental screening while maintaining acceptable sensitivity across diverse bacterial strains. External benchmarking, prospective laboratory validation and comparisons with established phage–host prediction methods will be important for determining its performance in routine research applications.

European Supercomputing Resources Supported CAPHARD Development

The development of CAPHARD has been closely associated with the European High Performance Computing Joint Undertaking, or EuroHPC JU, an initiative supporting advanced scientific and industrial computing across Europe.

The CAPHARD project received an allocation of 50,000 node-hours on the Leonardo supercomputer, specifically its Booster partition. The allocation covered the period from July 2024 to July 2025 and was awarded through the EuroHPC programme supporting artificial intelligence and data-intensive applications.

Leonardo, hosted by the Italian supercomputing consortium CINECA, provides the computational infrastructure required for demanding scientific workloads, including the training of large-scale artificial intelligence models. For CAPHARD, these resources supported the development of genomic language modeling approaches incorporating information from bacteriophage and bacterial proteins.

The project was led by Andrea Di Gioacchino, who headed the data team at Phagos. Its scientific objective was to develop computational methods capable of predicting which bacteriophages from a reference collection were most likely to infect and eliminate a particular bacterial pathogen, using genomic sequences as the principal input.

The initiative brought together expertise in microbiology, artificial intelligence and bioinformatics, alongside experimental datasets describing interactions between characterized bacteriophages and bacterial strains. These data are essential because genomic prediction models require experimentally established relationships between sequence information and observed infection phenotypes.

CAPHARD subsequently received recognition at EuroHPC User Day 2024, held in Amsterdam on October 22 and 23. During the event, Andrea Di Gioacchino received the Best AI Project Award for the development of the platform and its application to the identification of bacteriophages targeting antibiotic-resistant bacterial pathogens.

The distinction highlighted the potential contribution of large language models to microbiological research and illustrated how European computing infrastructure can support the development of specialized biotechnology tools. The public release in October 2026 represents a further stage in the project, making a prediction capability developed through this research programme accessible to the wider scientific community.

Improving Prediction Through Experimental Data and International Collaboration

One of the principal limitations of machine learning approaches to phage–host prediction is the availability and representativeness of experimental datasets. Although genomic databases contain extensive information about bacteriophage diversity, comparatively few viral genomes are associated with sufficiently characterized infection profiles across large collections of bacterial strains.

This problem is particularly important for pathogenic bacteria displaying substantial intraspecies diversity. Two isolates classified within the same bacterial species may differ considerably in their susceptibility to a given phage because of variations in surface structures, defense systems or other genetic determinants.

Phagos acknowledges that clinically relevant bacterial strains are frequently underrepresented in publicly available interaction datasets, creating difficulties when extending predictive models beyond the organisms used during training.

The company also emphasizes the importance of negative experimental results. According to its release, only approximately 3% of tested phage–host combinations in the public interaction data it references correspond to successful infections. Although this proportion depends on the composition and experimental design of the underlying datasets, it illustrates the rarity of compatible interactions within many screening collections.

Negative results can provide essential information about the boundaries of host specificity and help distinguish genomic characteristics associated with successful infection from those present in incompatible combinations. Their inclusion may therefore improve model training, particularly when negative interactions are experimentally verified rather than assumed from missing observations.

To expand the available evidence, Phagos is inviting researchers to provide feedback on CAPHARD's rankings and contribute characterized phage–host interaction datasets. Contributions may include both successful and unsuccessful infections, accompanied by the corresponding genomic sequences.

Participation is voluntary, and the company states that permitted data uses and attribution arrangements will be agreed upon before information is shared. Contributors are expected to receive acknowledgment in future model documentation, while subsequent versions of CAPHARD are planned for release through Hugging Face.

These collaborations could help increase the diversity of bacterial species represented in the training data, improve performance against previously underrepresented strains and identify systematic prediction errors that may not become apparent from internal evaluation alone.

Connecting Phage Collections With Researchers and Hospitals

Beyond computational prediction, the CAPHARD initiative also aims to improve access to existing bacteriophage resources by making independently maintained collections more discoverable.

Many research laboratories and biobanks possess sequenced bacteriophages that remain difficult for external researchers to identify. These collections may contain viruses capable of infecting clinically relevant bacterial strains, but their practical usefulness can be limited when information about their host ranges or institutional availability is incomplete.

Phagos is therefore encouraging laboratories to deposit phage genomic sequences into CAPHARD's reference collection. When a deposited bacteriophage is ranked as a promising candidate against a bacterial target, the application may allow researchers to identify the institution maintaining the phage and establish direct contact.

The participating laboratory retains ownership and physical possession of its biological material, while decisions concerning distribution remain under its control. This approach does not require institutions to transfer their phage stocks to Phagos or automatically make them available for therapeutic use.

By associating genomic prediction with information about the location of biological resources, CAPHARD could facilitate collaborations between phage banks, microbiology laboratories and clinical research teams. However, additional characterization, material transfer arrangements and relevant regulatory procedures would remain necessary before any candidate could be considered for therapeutic application.

The accessibility of genomic data also raises questions about confidentiality and research data management. Phagos states that genomes submitted through the public CAPHARD application are not collected, stored or reused by the company. The company presents this privacy policy as a way to allow researchers to analyze genomic sequences without contributing them to proprietary datasets unless they explicitly choose to do so.

Phagos Expands Its AI Platform Following Regulatory and Commercial Progress

Founded in France in 2021 by Alexandros Pantalis and Adèle James, Phagos has developed a biotechnology platform combining bacteriophage isolation, microbial genomics, artificial intelligence and experimental validation. Its initial commercial activities have concentrated on veterinary applications, with the longer-term objective of adapting its technological approach to human medicine.

The company's development strategy is based on an iterative relationship between computational predictions and biological experiments. Phages are isolated and sequenced, bacterial strains are characterized, interactions are predicted computationally and selected candidates undergo laboratory testing. Experimental observations and field results can subsequently generate additional information for improving predictive models.

This approach is particularly relevant for phage therapy because bacterial populations can evolve resistance to individual bacteriophages. The ability to identify alternative phages and update therapeutic formulations may therefore become important when maintaining treatment activity against changing bacterial populations.

In 2025, Phagos announced a regulatory authorization in France for its personalized veterinary phage therapy platform, which the company described as the first authorization of its kind. The development concerns veterinary medicines and should not be confused with authorization for human clinical use.

Its Alphagos platform was designed to identify bacterial pathogens, select appropriate bacteriophages and develop formulations adapted to the strains encountered in livestock production. Phagos indicates that its technology can support the development and deployment of customized veterinary formulations within approximately two months, depending on the relevant application and procedures.

The company has concentrated particularly on bacterial diseases affecting agricultural production, including infections associated with organisms such as Escherichia coli and Salmonella. Its wider research interests also include additional veterinary pathogens and applications beyond animal health.

The commercial expansion was supported by a €25 million Series A financing round announced in October 2025. The investment was co-led by Capagro, Hoxton Ventures, CapHorn and Demeter, with participation from Acurio Ventures, Citizen Capital, Entrepreneurs First, Founders Capital and STATION F.

The financing was intended to strengthen research and development, expand the company's artificial intelligence capabilities, increase deployment of veterinary phage products and support international expansion across Europe, Asia and the Americas.

Phagos has since reported substantial deployment of its phage-based formulations in animal agriculture. According to the company's published figures, more than 3 million animals in France have received phage cocktails selected through its AI-assisted discovery process since commercial field use began in April 2025.

This figure reflects the scale of reported deployment rather than a controlled measure of clinical or veterinary efficacy. Publicly available information does not establish treatment success rates, reductions in mortality or comparative antimicrobial consumption for all animals included in this total.

Nevertheless, the reported deployment provides an indication of the volume of practical application being pursued by the company and of the potential value of connecting computational prediction with repeated experimental and field observations.

Phagos has also gained international recognition for its artificial intelligence activities. In January 2026, the World Economic Forum selected the company for the second cohort of its MINDS programme, which recognizes applications of artificial intelligence demonstrating measurable operational impact. The selection followed the company's earlier recognition as a World Economic Forum Technology Pioneer in 2025.

These developments reflect the company's progression from a research-oriented biotechnology venture toward an organization seeking to apply genomic prediction and personalized phage selection in commercial settings.

From Veterinary Applications to Human Phage Therapy Research

Although Phagos has focused its initial deployments on veterinary medicine, CAPHARD could also have implications for laboratories investigating bacteriophages against human bacterial pathogens.

Personalized phage therapy depends on the identification of viruses capable of infecting the bacterial strains responsible for individual infections. This often requires the isolation and characterization of the causative organism, followed by screening against available phage libraries.

A genomic prediction model could help researchers organize this selection process, particularly when comparing large collections or searching for candidate phages maintained by external institutions. It could also contribute to the early stages of phage cocktail development by highlighting combinations worthy of further experimental characterization.

However, predicted genomic compatibility represents only one component of therapeutic phage selection. Candidate phages must also undergo analyses of their replication characteristics, host range, genome composition, capacity to transfer undesirable genetic material and suitability for pharmaceutical development.

Furthermore, infection conditions encountered in vivo can differ considerably from those reproduced during routine laboratory assays. Bacterial physiology, biofilm organization, tissue accessibility, immune responses and interactions with other microorganisms can influence therapeutic outcomes independently of genomic susceptibility.

For these reasons, computational prediction cannot replace laboratory susceptibility testing, preclinical studies or clinical evaluation. The practical value of CAPHARD lies primarily in supporting the identification and prioritization of candidate phages rather than establishing their therapeutic effectiveness.

Phagos has identified human health as an area for future expansion, particularly through research collaborations and drug discovery partnerships. The public release of CAPHARD may facilitate contacts with academic institutions and clinical microbiology groups while generating additional opportunities to evaluate the model against medically relevant bacterial strains.

The broader importance of the project will depend on whether its predictions generalize beyond existing datasets and meaningfully reduce the experimental resources required to identify compatible bacteriophages.

Toward More Accessible Computational Phage Discovery

The release of CAPHARD makes a substantial genomic prediction system available to researchers who may not possess the computational infrastructure needed to develop comparable models independently. Its combination of whole coding genome analysis, a reference collection of 832 bacteriophages and a publicly accessible prediction interface provides a practical tool for exploring potential phage–bacteria interactions.

The project also reflects a wider evolution in microbiology, where increasingly sophisticated computational methods are being integrated with sequencing, experimental screening and the management of biological collections.

For CAPHARD, the central scientific questions now concern predictive generalization, experimental reproducibility and the extent to which the model can improve candidate selection across bacterial species and phage families. The incorporation of independently generated positive and negative interaction data will be particularly important for evaluating performance beyond the company's existing experimental datasets.

By opening access to its prediction interface and encouraging collaboration with external phage collections, Phagos is seeking to extend the usefulness of its genomic modeling technology beyond its internal discovery programmes. Whether this approach can substantially accelerate personalized phage therapy will require further evidence, but it provides an additional computational resource for researchers investigating bacterial host specificity and developing more systematic strategies for bacteriophage selection.

Researchers can access CAPHARD through the Phagos Hugging Face application, while scientific feedback, experimental interaction datasets and requests to include phage genomes in the reference collection can be directed to caphard@phagos.org.

Sources :

Phagos. (October 20, 2026). Introducing CAPHARD: a model to find the viruses most likely to infect your bacteria.
https://www.phagos.com/news/introducing-caphard-hugging-face

Phagos. CAPHARD: Phage-host interaction prediction application. Hugging Face. Includes the reported internal evaluation results for 1,920 phage–host pairs.
https://phagos-biotech-phagos.hf.space/

European High Performance Computing Joint Undertaking. CAPHARD: Codon-Assisted Phage-Host Automatic inteRaction Discovery. Project allocation and scientific objectives.
https://www.eurohpc-ju.europa.eu/caphard-codon-assisted-phage-host-automatic-interaction-discovery_en

European High Performance Computing Joint Undertaking. (October 23, 2024). EuroHPC User Day 2024: Awards Session to Celebrate Excellence in HPC and AI.
https://www.eurohpc-ju.europa.eu/eurohpc-user-day-2024-awards-session-celebrate-excellence-hpc-and-ai-2024-10-23_en

Phagos. (October 2025). Phagos Raises €25m to End Bacterial Disease.
https://www.phagos.com/news/series-a-fundraising

Phagos. (January 19, 2026). Phagos Joins the World Economic Forum’s Second MINDS Cohort.
https://www.phagos.com/news/wef-minds-second-cohort

World Economic Forum. (January 19, 2026). What Works in AI: The Leading Companies Turning AI Into Real-World Impact.
https://www.weforum.org/stories/artificial-intelligence/the-leading-companies-turning-ai-into-real-world-impact

Phagos. About Us. Company development, funding and reported animal health deployment.
https://www.phagos.com/fr/about-us

Phagos. Alphagos Platform. Personalized veterinary phage development and deployment process.
https://www.phagos.com/fr/platform