A foundational model encodes deep phenotyping data and enables diverse downstream applications.

Area of Science:

• Computational biology
• Genomics
• Precision medicine

Background:

• Deep phenotyping data presents analytical challenges due to scale and complexity.
• Conventional approaches struggle to address these challenges effectively.

Purpose of the Study:

• Introduce ukbFound, a foundation model for analyzing deep phenotyping data.
• Demonstrate ukbFound's capabilities in disease stratification, multimorbidity analysis, and prediction.

Main Methods:

• Developed ukbFound, a foundation model encoding individual traits into language-like sequences.
• Incorporated domain-specific tokenization, position-free embedding, and interpretable reasoning.
• Applied ukbFound to 502,118 UK Biobank individuals for analysis.

Main Results:

• Identified distinct patient subgroups in 289 diseases, with prognostic differences in 18.3%.
• Uncovered novel associations between conditions and disease communities.
• Outperformed benchmark models in disease prediction using lifestyle and dietary data, identifying high-risk individuals.

Conclusions:

• ukbFound offers a scalable and interpretable framework for deep phenotyping data analysis.
• The model advances precision medicine by revealing latent disease-trait relationships.
• Identified potential novel indicators for disease progression, such as basophil counts in COPD.