Abstract
Early prediction of chronic diseases from routine blood tests has potential to transform public health prevention strategies. Here, we developed MetaboLM, a transformer-based language model pre-trained on plasma metabolomics data from 83,744 relatively healthy UK Biobank participants. After fine-tuning with metabolomics data from individuals diagnosed with 16 common chronic diseases, MetaboLM demonstrated excellent performance in disease prediction and stratification, and generated a metabolomic risk score (MetaboRS) capable of predicting disease onset more than 10 years in advance. MetaboRS outperformed established demographic predictors in 16 diseases, and outperformed atherosclerotic cardiovascular disease (ASCVD) risk equations in 13 diseases. Furthermore, interpretability analysis of the attention mechanism identified key metabolites related to disease prediction. These findings underscore the potential of metabolomic language models and derived risk scores for predicting the risk of multiple diseases and for other potential downstream applications.