Proteomic signatures improve risk prediction for common and rare diseases

Summary

• affinity-based assay targeting ~3,000 pre-selected plasma proteins (from here on I’ll just say proteomics) for disease prediction
• proteomics is in many cases more effective for predicting disease risk compared to current methods — both compared to models using only basic clinical information and models using clinical info + 37 common blood assays
• predicted 10-year incidence of 218 diseases; the proteomics approach outperformed in 67 (vs. clinical info only) and 52 (vs. clinical info + lab assays) respectively
• the proteomics models were sparse: only 5-20 proteins per model
• they detected both disease-specific proteins and proteins shared across multiple diseases
• 147 of 501 predictor proteins were selected for 2+ diseases, and 5 predicted across more than 10 diseases (4 of those 5 mainly bc they correlate with age)
• the biggest diagnostic improvement was achieved for rarer conditions and those where blood plays a more important role (e.g. leukemia)

Questions

• they tested ~3k proteins (including all isoforms, humans have >100k). Based on what criteria did they pre-select these? How can we make this work for proteins that occur at lower levels? Might these actually be even more informative for predicting disease?
• can we also predict common diseases better via proteomics (heart disease, diabetes)? this paper suggests that currently the evidence is much stronger for rarer ones.
• how long does it take to develop absolute quantification for a single protein?