Self-supervised plasma proteomic representations for prospective disease prediction across varying protein availability

Large-scale plasma proteomics offers opportunities to characterize disease susceptibility and improve prospective risk prediction, but transferring proteomic predictors across datasets remains challenging because measured protein sets differ across cohorts, study phases and assay configurations. Here we developed a self-supervised protein-token Transformer that maps the proteins observed in each sample to a fixed-dimensional participant representation, allowing unavailable proteins to be omitted rather than imputed. Using plasma proteomic profiles from 53,014 participants in the UK Biobank Pharma Proteomics Project, we pretrained the encoder by masked-protein reconstruction and evaluated whether disease models developed from comprehensive 2,920-protein profiles could be reused with a predefined subset of 1,460 proteins. Across 144 diseases, median AUC was 0.679 with comprehensive coverage and 0.637 when the same encoder and disease models were applied to partial-coverage representations without refitting; retraining only the disease-specific models increased median AUC to 0.673. Under partial coverage, protein-token proteomic risk scores (ProRS) exceeded coefficient-truncated LASSO ProRS by a median paired AUC difference of 0.027 and were comparable to LASSO ProRS refitted using outcome labels, with a median difference of 0.003. Performance was relatively stable across cardiovascular-kidney-metabolic diseases but more heterogeneous across autoimmune diseases, while protein-token representations improved discrimination beyond clinical covariates for 10 of 12 focused diseases under partial coverage without refitting. These results support self-supervised protein-token representations as a strategy for building proteomic prediction models that remain usable across heterogeneous measurement settings.