According to the World Health Organisation, the rise of bacteria resistant to antibiotics is one of the most critical challenges for global health and development today. One potential solution is the employment of bacteriophages, which are viruses that target and kill bacteria with remarkable precision. This same precision, while enabling the development of therapies with few side effects, also necessitates the rapid identification of the appropriate phages to combat specific pathogenic bacteria.
There are two main approaches to obtain the most effective phages against a given bacterial infection:
- by collecting natural phages from the environment,
- or by engineering the genome of existing phages to tune their host range
This second approach holds the key for a revolution in phage therapy as it potentially leads to a tailored phage for each pathogenic bacteria. However, this approach presents many challenges; engineering new, functional phage proteins or protein complexes demands a deep understanding of intricate inter-organism protein-protein interactions.
This project will capitalise on previous AI efforts utilising Large Language Models (LLMs) for analysing vast and diverse phage and bacterial genomic datasets, to train the first generative model able to adapt phage genomes to new hosts. This generative approach will set the basis for de novo design of novel phage sequences, overcoming limitations of alternative experiment-intensive methods and accelerating the discovery of effective therapies. This ambitious undertaking necessitates substantial computational power for model training on extensive genomic datasets, high-throughput inference, and iterative optimisation, requiring access to large GPU and CPU clusters. Our researchers will then use our facilities at Phagos to test the model predictions with in vitro experiments, and feed the resulting data back to the model in a continuous positive feedback loop, accelerating the development of urgently needed phage therapies.
Andrea Di Gioacchino, PHAGOS, Italy