Biobank proteomics and AI: from proteins to risk prediction

New and Comprehensive Protein Dataset Elevates Health Research

Technology Networks · 2023-10-04 · Source

In this English-language Technology Networks interview published on 4 October 2023, Chris Whelan discusses the UK Biobank protein dataset and research opportunities. It predates the 2024 and 2025 risk-model studies discussed here and does not validate their results or clinical benefit.

What is a protein signature?

Proteomics measures many proteins together. A protein signature is a combination of measurements used by a prediction model. UK Biobank studies link these measurements with diagnoses recorded during follow-up. A protein’s contribution to prediction does not by itself establish that it causes a disease.

What did the 2024 study show?

Carrasco-Zanini and colleagues analyzed approximately 3,000 plasma proteins alongside clinical information in 41,931 participants, modeling the 10-year incidence of 218 diseases. Panels containing 5–20 proteins improved prediction over basic clinical-information models for 67 diseases. This is a specific research comparison, not evidence that clinical diagnostic tests for 67 diseases are ready.

Is a neural network always better?

The 2025 study by Smith and colleagues compared linear ElasticNet and neural-network models for 27 outcomes in 53,030 UK Biobank participants over 5- and 15-year follow-up periods. Neural networks outperformed linear models for 11 outcomes. The same model is not superior for every task; because both studies use UK Biobank data, they are not independent population validations of one another.

Is risk ranking the same as an individual probability?

The C-index evaluates how well a model ranks people in relation to event timing and risk; it is not a person’s percentage probability of disease. External replication in the 2024 study covered six diseases in EPIC-Norfolk, not all models. Transfer to another population requires testing both ranking and agreement between predicted probabilities and observed outcomes.

What is still needed before clinical use?

These observational modeling studies do not themselves provide regulatory authorization or evidence that using a model improves patient outcomes. Our editorial assessment: local laboratory measurements, different age and population groups, false alarms, follow-up workload and the clinical pathway after a high-risk result should be evaluated together. This article offers no individual screening or treatment advice and makes no effectiveness claim for RAI products.