Optimizing Human Exposome Biomonitoring: A Machine Learning Approach to Predict Optimal Biofluid Matrices

Abstract

Biomarker identification is crucial for exposomic studies, yet few have been established relative to the vast number of chemicals human encounter. While biomarkers can be detected in blood or urine, the optimal biological matrix for each chemical remains unclear. We curated data on biomarker identification in urine or blood for 526 chemicals from 4797 biomonitoring entities, sourced from 89 distinct cohorts across 43 countries, and developed a machine learning model named Biomarker Matrix Identifier (BMI) to predict the most suitable biological fluid for biomarker identification. Our model achieves over 94% accuracy using circular fingerprints as the input. Applying this method to the Human Exposomic Metabolome Database (HExPMetDB) containing over 20,000 chemicals revealed that approximately 67% of compounds are predicted to be more effectively monitored using urine as the optimal biomonitoring matrix. This predictive model enhances the accuracy of the exposure assessment in human exposomic analysis, facilitating more efficient biomarker identification strategies. In sum, we have established an effective prediction model in facilitating the prediction of whether the identified chemicals in the biological fluids can represent exposure for human exposomic analysis.