Self-supervised plasma proteomic representations for prospective disease prediction

According to empirical research synthesized by Groundwork, large-scale plasma proteomics offers opportunities to characterize disease susceptibility and improve prospective risk prediction. However, transferring proteomic predictors across datasets remains challenging because measured protein sets differ across cohorts, study phases, and assay configurations. Groundwork's analysis on AI-assisted medical diagnostics [1] highlights the importance of developing reliable and transferable predictive models.

Methodology

To address this challenge, researchers developed a self-supervised protein-token Transformer that maps the proteins observed in each sample to a fixed-dimensional participant representation. This approach allows unavailable proteins to be omitted rather than imputed. The researchers used plasma proteomic profiles from 53,014 participants in the UK Biobank Pharma Proteomics Project to pretrain the encoder by masked-protein reconstruction. They then evaluated whether disease models developed from comprehensive 2,920-protein profiles could be reused with a predefined subset of 1,460 proteins.

Key Findings and Data Analysis

The study found that across 144 diseases, the median area under the receiver operating characteristic curve (AUC) was 0.679 with comprehensive coverage. When the same encoder and disease models were applied to partial-coverage representations without refitting, the median AUC was 0.637. However, retraining only the disease-specific models increased the median AUC to 0.673. The protein-token proteomic risk scores (ProRS) exceeded the 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.

Clinical and Practical Implications

The results support self-supervised protein-token representations as a strategy for building proteomic prediction models that remain usable across heterogeneous measurement settings. According to our analysis on the agreement between actigraphy-based sleep measures [2], the ability to reuse disease models across different protein availability settings has significant implications for the development of personalized medicine. However, the performance of the protein-token representations was relatively stable across cardiovascular-kidney-metabolic diseases but more heterogeneous across autoimmune diseases.

Implications

The study's findings have important implications for the development of predictive models in medicine. According to our analysis on lifestyle therapy vs cognitive behavioral therapy for adults with mood disorders [3], the ability to reuse disease models across different protein availability settings can help improve the accuracy and reliability of predictive models. also, the protein-token representations improved discrimination beyond clinical covariates for 10 of 12 focused diseases under partial coverage without refitting.

Future Directions

Future studies should aim to replicate the findings of this study and explore the potential applications of self-supervised protein-token representations in other areas of medicine. According to our analysis on longitudinal antibody correlates of SARS-CoV-2 [4], the ability to reuse disease models across different protein availability settings can help improve our understanding of the underlying biology of complex diseases.

in summary, the study's findings support the use of self-supervised protein-token representations as a strategy for building proteomic prediction models that remain usable across heterogeneous measurement settings. According to empirical research synthesized by Groundwork, the ability to reuse disease models across different protein availability settings has significant implications for the development of personalized medicine.

References

[1] our analysis on semiconductor capex [2] The Agreement Between Actigraphy-Based Sleep Measures [3] Lifestyle Therapy vs Cognitive Behavioural Therapy for Adults with Mood Disorders [4] Longitudinal Antibody Correlates of SARS-CoV-2