Scientists develop AI tool to read the “code” of bacterial viruses and design personalised phage treatments

Antibiotic resistance is making the treatment of bacterial infections increasingly difficult. Phages, which attack bacteria, are emerging as an alternative to conventional antibiotic treatment. However, identifying which phage is effective against a given bacterium is highly complex. This study, led by Robby Concha-Eloko, Beatriz Beamud and professors of Genetics at the UV Pilar Domingo-Calap and Rafael Sanjuán, proposes the use of artificial intelligence to facilitate this prediction process.

To develop the model, the researchers used Klebsiella, a bacterium included in the World Health Organization’s list of priority bacterial pathogens, responsible for severe hospital-acquired infections and noted for its high resistance to antibiotics. Klebsiella bacteria are protected by polysaccharide capsules that block the action of antibiotics as well as the entry of phages. To overcome this barrier, many phages produce depolymerases, enzymes that degrade these capsules and allow the bacteriophage to enter and infect the bacterium, thereby aiding treatment.

However, the enormous genetic diversity of these capsules — more than 100 serotypes have been recorded in Klebsiella — has hindered prediction of which phage may penetrate the capsule and infect the bacterium. At the same time, this great variety of capsule serotypes makes Klebsiella an ideal model for studying the interaction between phages and capsules.

With this aim, the research team has developed a pioneering tool that exploits genetic information from thousands of Klebsiella bacteria and their “dormant” viruses (prophages) integrated into their genomes. By analysing more than 74,000 prophages and almost 20,000 depolymerase sequences, the researchers have created a database that associates each enzyme with the type of bacterial capsule it can degrade.

Using advanced machine learning techniques and artificial intelligence models inspired by natural language processing (similar to those used in automatic translators), they have succeeded in accurately predicting the “tropism” or specificity of each depolymerase — that is, what types of bacterial capsules it can recognise and destroy.

A solution against biofilms

This study represents a key advance for phage-based biotechnology and its components, as it allows their specificity to be predicted. This is crucial for designing future applications, such as solutions against biofilms — the protective structures formed by some bacteria to adhere to surfaces and resist treatment.

Biofilms are increasingly recognised as a major obstacle in the treatment of infections. In fact, they have been shown to be involved in the chronicity of diseases such as cystic fibrosis, chronic wounds, prosthesis-related infections and urinary tract infections.

“The use of depolymerases, either in combination with current treatments (antibiotics or antimicrobial peptides) or potentially as enhancers of the immune system, can address issues related to biofilm formation, leading to a reduced risk of treatment failure”, explains Robby Concha.

“Compared with the traditional method, which relies on a tedious search-and-test process with phages to find an effective depolymerase, artificial intelligence models allow us to predict specificity in silico”, the researcher states. In this sense, the method demonstrated in the study generates depolymerase libraries that can be used to identify the most effective enzyme, optimising capsule degradation and, subsequently, biofilm disruption.

According to Concha, one of the developers of this pioneering tool, although Klebsiella was used as a model, the methodology can be applied against any other capsule-producing bacterium. This includes most of the priority pathogens listed by the WHO.

In summary, this study addresses predictions of phage–host interactions in two ways. First, by exploiting data contained within bacterial genomes (prophages), which provides essential training data; and second, by proposing an architecture that enables the model to be trained on all bacterial species simultaneously, in an integrative way.

Reference: Concha-Eloko, R., Beamud, B., Domingo-Calap, P. et al. Unlocking data in Klebsiella lysogens to predict capsular type-specificity of phage depolymerases. Nat Commun 16, 8798 (2025). https://doi.org/10.1038/s41467-025-63861-w