Molecular Toxicity Prediction and Machine Learning

Abstract

With the rapid growth in the number of chemicals, conventional in vivo and in vitro toxicity tests face challenges including high cost, long experimental cycles, and ethical concerns, creating an urgent need for efficient computational toxicology approaches. This review systematically summarizes the application of machine learning to molecular toxicity prediction, with particular emphasis on two core components: algorithm selection and descriptor construction. At the algorithmic level, the principles and applicable scenarios of conventional machine-learning models, including random forest (RF), extreme gradient boosting (XGBoost), and k-nearest neighbors (KNN), as well as deep-learning methods, are introduced in detail, together with the classification framework of supervised, unsupervised, and semi-supervised learning. At the descriptor level, in view of the dependence of conventional computational descriptors on accurate molecular structures, the advantages and recent applications of mass spectrometry (MS) data as a new type of descriptor are highlighted. These applications include molecular-structure recognition (CANOPUS and SIFTER), quantitative compound analysis (an internal-standard replacement strategy based on natural organic matter background peaks and virtual molecular projection CNNs for nanoparticles), and spectrum prediction. Current studies indicate that combining mass-spectrometric descriptors with machine learning can reduce dependence on structural databases and enable rapid classification and quantitative prediction of unknown compounds, showing considerable potential for environmental monitoring and toxicity screening. Nevertheless, current models remain strongly data-dependent and are limited by insufficient interpretability and restricted applicability domains. Future work should therefore focus on expanding datasets, improving model interpretability, and strengthening cross-domain generalization.