A conversation with Paolo Savoca of NexaBiome Life Sciences on the promise of AI in phage therapy.
Phage therapy is gaining attention as a promising treatment for infections that antibiotics are no longer responsive to. Because phage therapy is highly targeted, matching the right phage to the right bacteria to ensure a tailored and effective treatment for the patient, can be time intensive.
However, ongoing developments in phage therapy are helping to streamline and speed up this process, in part driven by advances in artificial intelligence (AI). At NexaBiome Life Sciences, bioinformatician Paolo Savoca is working at the intersection of genomics, machine learning and phage biology to help transform phage selection from a manual, experience-driven process into a reproducible, data-led workflow. We sat down with Paolo to discuss the potential of AI in phage therapy, and how he expects it to amplify the potential of the phage community.
Paolo Savoca, Bioinformatician at NexaBiome Life Sciences Ltd.
What are some of the modern challenges surrounding phage selection?
At first glance, phage therapy sounds simple enough – find a virus that kills the bacteria causing the infection. In reality, it’s far more complicated. Even within the same bacterial species, different strains can behave in completely different ways, and a phage that works well in one context may fail in another. Beyond that, it’s important to remember that infections don’t exist in isolation – they’re shaped by surrounding microbes, host immune responses, and local environmental conditions.
These complexities mean phage selection has traditionally relied heavily on specialist expertise and laboratory testing which, naturally, takes time. While broad phage “cocktails” can be designed to cover many cases, there will always be patients for whom a generic solution isn’t enough. Designing effective, reliable therapies therefore requires a way to navigate enormous biological variability – and this is where we are focusing our attention to develop new tools and methods.
When people talk about AI in this space, what does this mean?
In practical terms, AI – and machine learning in particular – is about finding patterns in data that are too complex or time consuming for humans to interpret reliably. In phage research, we generate vast amounts of data such as whole-genome sequences, phenotypic measurements, host-range assays, growth dynamics and more. This means each phage ends up with a kind of multidimensional fingerprint.
Machine learning allows us to connect those fingerprints in a systematic way. By training models on known phage/bacteria interactions, algorithms can learn which genetic or phenotypic features are associated with successful infection. Instead of testing every phage against every bacterium manually, AI can help us predict which candidates are most likely to work – guiding experimental effort and supporting scientific expertise, rather than replacing it.
How is AI already being used in phage research and medicine more broadly?
Across phage research, AI is being explored as a tool to speed up discovery, improve reproducibility, and reduce the trial-and-error aspect of selection. Researchers are using machine learning to predict host range, infer virulence, and even suggest phage combinations that may work synergistically. Similar approaches are also gaining traction in other areas of medicine, from drug discovery to diagnostics.
What’s particularly impactful about AI in phages is the ability to move from descriptive biology to predictive biology. Instead of only describing how a phage behaves after testing it, models can estimate behaviour based on genomic and phenotypic data alone. While experimental validation remains essential, these predictions dramatically narrow the search space, saving time and resources while increasing the chances of success.
How is NexaBiome applying these ideas in practice?
At NexaBiome, we’re building an integrated workflow that combines high-throughput sequencing, structured data management and predictive modelling. Using Oxford Nanopore Technologies, we carry out in-house whole-genome sequencing of phages and bacteria, generating large volumes of consistent, high-resolution genomic data. These data are processed through standardised bioinformatic pipelines to ensure uniform assembly, annotation and phylogenetic analysis.
In parallel, we collect detailed phenotypic data through standard operating procedures developed in-house. Together, these datasets feed machine learning models designed to link genotype to phenotype and predict phage activity against specific bacterial hosts. This work is closely supported by a NexaBiome-funded PhD at the University of Strathclyde, where we’re developing and validating these models in collaboration with academic experts in computational biology. Importantly, predictions are always tested experimentally against new strains to assess how well the models generalise beyond their training data.
What are the biggest benefits this approach could unlock for patients – and what do you predict the future to look like?
The most immediate benefit is speed. Phage selection can currently take weeks, which is time many patients with drug-resistant infections simply don’t have. AI-guided workflows could reduce this backlog by prioritising the most promising phages early on. Beyond speed, these approaches also improve consistency and reproducibility, helping to move phage therapy towards a fully scalable medical technology.
Looking ahead, the long-term vision is personalised phage therapy informed by microbiome and genomic data. As national phage services and clinical collaborations expand, clinician-led reporting of real-world outcomes can help further refine predictive models. By no means will AI replace the wet lab or clinical judgement, a human expertise and oversight are pivotal to this, but it can help turn phage selection into a data-driven, repeatable process, bringing us closer to delivering effective phage therapies to patients who may not otherwise have any options.
