Answers are grounded in this paper and cite their sources.
To ensure environmental safety, we need to predict aquatic toxicity, but experimental testing is expensive, time consuming and often raises ethical concerns. In this study, we used machine learning to predict the 96-hour lethal concentration (LC50) of 908 organic compounds to fathead minnow ( Pimephales promelas ). We compared four different ways to represent molecular structure: (A) six pre-calculated descriptors from the original dataset, (B) twelve physicochemical descriptors computed from SMILES using RDKit, (C) 2048-bit Morgan fingerprints, and (D) a combined set of original and RDKit descriptors (18 features total). Three regression models were used such as Linear Regression, Random Forest, and Gradient Boosting. Results showed that the combined descriptor set with Gradient Boosting gave the best prediction (cross-validation R² = 0.6278), which was about 5% better than using the original descriptors alone (R² = 0.5956). Feature importance analysis showed that lipophilicity (MolLogP) was the most important predictor, which makes chemical sense because fat-soluble molecules cross fish gill membranes more easily. Morgan fingerprints performed poorly, likely because the number of features (2048) was too large for the dataset size. These results show that combining different types of molecular descriptors can improve toxicity prediction
No takes yet. Share an insight, caveat, or question.
Rahim et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: