fishFingers: Prediction of bioconcentration factor (BCF) for chemicals detected by HRMS in fish by species using recommender system machine learning

Abstract Exposure to novel entities in the environment now exceeds our capacity to evaluate the risk of each chemical under traditional frameworks. Detecting, prioritizing and identifying thousands of anthropogenically-derived chemicals already present in the environment is a key challenge to understand this planetary boundary. The application of machine learning to chemical risk management has already demonstrated quick and accurate ecotoxicity characteristics of chemicals with low confidence structural annotation inferred from high-resolution mass spectrometry (HRMS). The rate at which a chemical is absorbed into an organism compared to the rate it is excreted, known as the bioconcentration factor (BCF), is also an important factor for evaluating chemical risk. Since toxicity and bioaccumulation are often used to regulate chemicals, accurate and precise in silico techniques can greatly enhance our capacity for chemical risk assessment. Unfortunately, all known methods of reliable BCF prediction require a confident chemical structure. The prediction of BCF from HRMS features without a priori structural annotation will be demonstrated, using high-quality tandem mass spectrometry (MS2) to elucidate key structural fingerprints. This offline prioritisation method for non-targeted screening can quickly rank features in order of decreasing risk of bioaccumulation. A dataset containing measured BCF values for more than 1000 chemicals was obtained and curated for training and testing. For each chemical, a matrix of 6701 structural fingerprints were calculated using the rcdk package, incorporating MACCS, PubChem, Klekota-Roth, and a number of custom SMARTS, excluding those derived from the chemical formula. A recommender system approach was applied to model y = f(x), where x = (c,s), with c representing the structural fingerprints, and s a one-hot encoding for over 100 species of fish. The performance of linear regression, decision tree and dropout tree algorithms were compared, and the best performing model was optimised and tested to achieve the highest balanced accuracy, calculated as (TPR+TNR)/2. Chemicals with high-quality MS2 in the MassBank spectral library were then used to validate the model using SIRIUS CSI:FingerID v5.8 to generate the structural fingerprints using the averaged peak lists. The performance of the model was compared with methods described in the literature, including hydrophobicity-based models (log Kow) and more recent machine learning models that use known structural annotations.