Transfer learning enables robust prediction of cellular toxicity from environmental micro- and nanoplastics.

Micro- and nanoplastics (MNPs) are emerging pollutants that accumulate in ecosystems, food chains, and the human body, raising concerns about human health risks. However, understanding their toxicity remains challenging due to limited experimental data and the poor performance of conventional machine learning models on small, fragmented datasets. To address this, we developed a transfer learning-based quantitative structure-activity relationship (QSAR) model to predict MNP-induced cytotoxicity. Our approach leverages a three-layer neural network pre-trained on a large nanoparticle dataset and fine-tuned on three small MNP datasets. By incorporating experimental annotations (e.g., dose, exposure time) and quantum chemistry descriptors, the model enhances predictive reliability despite data scarcity. The transfer learning QSAR model achieved high performance (ROC-AUC = 0.88, balanced accuracy = 0.83, MCC = 0.67, accuracy = 0.83, precision = 0.93, recall = 0.70, and F1 score = 0.80) based on the independent test set, outperforming traditional methods. Feature importance analysis identified surface functionalization and dose as critical predictors of cytotoxicity. The results also suggest that nanoplastics may induce cytotoxicity more rapidly than microplastics, even with shorter exposure durations. This study demonstrates the utility of transfer learning in environmental toxicology and supports its use for emerging contaminants where data are limited. Future work will focus on expanding dataset diversity and refining models to improve generalizability for regulatory and risk assessment applications.