Transformer model predicts human disease across partial proteomes

In 53,014 participants from the UK Biobank Pharma Proteomics Project, researchers evaluated a self-supervised protein-token Transformer designed to handle varying protein measurements without imputation. The encoder was pretrained by masked-protein reconstruction to map observed plasma proteins into fixed-dimensional representations. The authors tested whether risk models trained on 2,920 proteins could predict 144 diseases when restricted to a subset of 1,460 proteins. Across all 144 conditions, median AUC reached 0.679 with full coverage and 0.637 under partial coverage without refitting. Retraining only the disease-specific heads raised the median AUC to 0.673. Under partial coverage, the model exceeded coefficient-truncated LASSO scores by a median paired AUC difference of 0.027. Discrimination improved beyond clinical covariates for 10 of 12 focused diseases under partial coverage without refitting, with performance remaining relatively stable across cardiovascular-kidney-metabolic conditions but more variable in autoimmune diseases.