fishFingers: Prioritization of High-Resolution Mass Spectrometry Features by Prediction of Bioconcentration Factors in Fish Species

Abstract The number of novel entities produced around the world exceeds our capacity to evaluate the risk of each chemical under traditional frameworks. Detecting, prioritizing, and identifying thousands of chemicals present in the environment is a challenge to understanding this planetary boundary. Machine learning has demonstrated quick and accurate predictions for ecotoxicological end points for chemicals with structural fingerprints derived from high-resolution mass spectrometry (HRMS). This study provides a new nontargeted screening strategy to prioritize HRMS features by the prediction of bioconcentration factors (BCF) without a priori structural annotation. An XGBoost model was trained on a database with structural fingerprints from 1285 unique substances and log10 BCF measurements from 111 species of fish. Validation was performed using structural fingerprints derived from high-quality tandem mass spectrometry (MS2), with RMSE = 1.02, MAE = 0.81, and ρ = 0.50, demonstrating its capability in predicting log10 BCF and to rank HRMS features by predicted bioconcentration from a wide chemical space. Variations in BCF were observed in species such as common carp (Cyprinus carpio), which exhibits lower average BCF compared with zebrafish (Danio rerio). Halogenation, aliphatic and aromatic carbon, and heteroatomaticity were among the most important structural fingerprints for BCF. The model and package are available from https://github.com/drewszabo/fishFingers.