AIRR-BRIDGE is exploring how Trusted Research Environments (TREs) can securely connect to the UK’s national artificial intelligence supercomputing infrastructure, to safely make large-scale health data available for research.
Large scale supercomputers in the UK have historically been unable to support TREs, because TREs require strict user access controls, separation of projects, need-to-know access and detailed audit trails, making it difficult to scale research using sensitive data. The recent growth in the scale of machine learning (ML) and Artificial Intelligence (AI) models and the data used to train them has exacerbated this problem, as training and evaluating these large models requires access to very large-scale supercomputers with specialised compute capabilities such as Graphics Processing Units (GPUs).
AIRR-BRIDGE (Bridging Research Infrastructure for Data Governance Extension on AIRR) is enabling three real-world research projects running in the TREs at Alan Turing Institute and King’s College London (KCL) to use the UK’s national AI Research Resource supercomputers (AIRR).
Providing a TRE capability for AIRR is critical to support large-scale AI research with sensitive data at a national scale. This work will build generalised access pathways that can easily be reused by other projects wanting to use these datasets on AIRR, as well as a clear pathway to supporting the use of sensitive data as a standard service offering across current and future AIRR systems.
This project is led by the Alan Turing Institute alongside Bristol University (Isambard-AI), Cambridge (Zenith), KCL, University College London, and UK Biobank. A dedicated Patient and Public Involvement Expert (PPIE) consultant leads the project’s PPIE workstream.
We will adopt and adapt the federated “satellite” TRE model demonstrated in FRIDGE to support the use of three large scale health datasets on AIRR – one national research database (UK Biobank), and two NHS datasets from London hospitals hosted within the KCL TRE.
In addition, we will validate several other TREvolution components, including SATRE (a standard for TRE requirements), K8TRE (a reference TRE implementation providing a “standalone” alternative to the satellite TRE model), and SACRO-ML (a toolkit for semi-automated disclosure control for machine learning models and their outputs).
What AIRR-BRIDGE aims to achieve
By the end of the project, AIRR-BRIDGE will demonstrate how sensitive health data can be securely accessed and analysed on national AI supercomputing infrastructure. The teams’ goals include:
- Advancing federation – Demonstrating secure federation between TREs and AIRR supercomputers
- Expanding the use of national AI infrastructure – Enabling analysis of UK Biobank and KCL datasets on the national AI infrastructure
- Extending TRE capabilities – Evaluating satellite TRE extensions to supercomputing environments
- Aligning with SATRE – Assessing governance alignment with the Standard Architecture for Trusted Research Environments (SATRE) TRE standard
- Scaling access – Developing a reusable blueprint for scaling secure access to AI supercomputing
Progress so far
Within the first six months of delivery, the AIRR-BRIDGE team has:
- Successfully tested FRIDGE deployments on the AIRR supercomputers
- Successfully tested with non-sensitive data workloads
- Developed work for persistent storage at AIRR sites
- Met with our Public Advisory Group (PAG) in April, May and June to co-produce “explainers” on relevant TRE topics aimed at public audience on the topics of “shared responsibility”, and “security controls”. Meeting and consulting with the PAG has improved communication from the project team to the public.
In the final six months of the project the team aims to complete testing sensitive workloads from the three data sources, evidencing proof of concept, alongside open software, documentation and papers. Additionally, the team will share their work in 2 planned workshops; one Hackathon and an adopter’s workshop with academic, private and healthcare researchers. We will share the explainers publicly to increase awareness and understanding of our work, creating opportunities for ongoing discussion and feedback.
How AIRR-BRIDGE supports TREvolution
AIRR-BRIDGE exemplifies the DARE UK Research Exemplars programme by exploring innovative ways to enable secure access to sensitive health data on national AI supercomputing infrastructure, while maintaining strong governance, privacy and public trust. Through technical experimentation, multi-sector collaboration and public involvement, it is generating evidence, prototypes and reusable approaches that could shape the future of secure data research across the UK.
About the DARE UK Research Exemplars
The DARE UK Research Exemplars are short, impactful research projects demonstrating how emerging Trusted Research Environment (TRE) capabilities can support real-world sensitive data research across the UK. Working across multiple TREs and research domains, these projects bring together researchers, TRE operators, public contributors and other partners to apply and evaluate innovations developed through TREvolution. Their role is to test new approaches in live research settings, generating practical evidence on what works well, what challenges remain, and how sensitive data research can be strengthened for the public good.
