Building the UK's largest federated multimodal biomedical dataset.

Reach the same outcome

Two paths. Same first-class status. Both end with you running it independently — like PharosAI is now scaling.

The build

Four years ago, Greg and colleagues at King's College London published a breast cancer recurrence model that no clinician ever used. They couldn't validate it across other patient cohorts — the data access agreements were too hard to negotiate. By the time they'd containerized the model and shipped it manually, the field had moved on to multimodal foundation models. PharosAI is the answer to that failure. The team is building a 'data refinery' that digitizes tissue from UK biobanks into harmonized multimodal datasets, plus a federated platform that lets researchers train on the aggregate without any data leaving the biobank's trusted research environment. Each biobank runs its own TRE. Each research project gets an isolated workspace with a Flower SuperNode. SuperGrid orchestrates the federation centrally, with manual egress review on every model update that comes out. The first federated training run — a breast cancer classifier across the King's Health Partners and Breast Cancer Now biobanks — is now live.

Not a single clinician has used this model to help treat a patient. The blocker was not model performance, but validation: collaborators had relevant patient cohorts, yet governance and data-access agreements made moving data impractical.

Greg Verghese · on the breast cancer model that became the reason for PharosAI