The UK Cancer Vaccine AI & Supercomputing Project

Cancer vaccines are a form of treatment that trains the immune system to recognise and attack tumour cells. Unlike traditional vaccines that prevent infection, these treat existing cancers by targeting mutations unique to each patient's tumour. Early results are promising. Moderna's personalised vaccine reduced melanoma disease recurrence by 44% when combined with immunotherapy. BioNTech has shown similar success. The challenge is choosing which mutations to target.

A typical tumour contains thousands of mutations. Most are invisible to the immune system. Current selection tools predict whether a mutant protein fragment can stick to a molecule called HLA on the cell surface. This is necessary but not sufficient. At best, only about 20% of predicted targets generate measurable T-cell activity in patients. This means most of the antigens included in a vaccine may contribute little to its effect and potentially dilute its effectiveness.

We need better prediction methods that consider the full antigen processing pathway. A mutation must be expressed by the tumour, processed inside the cell, transported to the right location, loaded onto HLA, and presented on the cell surface - all while the immune system is capable of responding. The tumour's biology matters. Expression levels vary. Processing machinery can be faulty. The local environment may suppress immune activity. Current tools ignore most of this context because they were designed when computing power was limited.

We have designed a different kind of predictor using the UK's Dawn supercomputing infrastructure. It is a foundation model trained on 120 million peptides from 12,000 tumours. Rather than predicting binding alone, it predicts functional immune responses by integrating gene expression, cellular processing, immune environment, and T-cell recognition. In retrospective analysis across multiple cancer types, it outperformed existing tools by approximately 40%.

The major question is whether computational...