Proformer: a multimodal proteomics transformer model for multidisease early risk assessment

Area of Science:

• Biomedical informatics
• Computational biology
• Proteomics

Background:

• Current chronic disease risk assessment is disease-specific and data-intensive.
• Existing protein large language models are not clinically applicable for disease risk prediction.

Purpose of the Study:

• To develop and validate a multimodal proteomics Transformer (Proformer) for multidisease risk assessment.
• To integrate protein expression, sequence, and function for enhanced predictive capabilities.

Main Methods:

• Trained Proformer on UK Biobank proteomics data from 47,124 individuals.
• Evaluated Proformer's performance in discriminating risk for 20 common chronic diseases.
• Compared Proformer against machine learning, deep learning, and clinical risk predictors.

Main Results:

• Proformer achieved state-of-the-art performance across all 20 chronic diseases.
• Outperformed Age+Sex, ASCVD risk score, and a 35-variable clinical panel in 10-year risk discrimination.
• Results were successfully replicated in independent Scotland and Wales cohorts.

Conclusions:

• Proformer offers a novel, single-measurement approach for multidisease risk profiling.
• Enables direct generation of 10-year chronic disease risk reports from proteomics data.
• Represents a significant advancement in clinical proteomics for personalized disease prevention.