Multimodal feature fusion machine learning for predicting chronic injury induced by engineered nanomaterials

Concerns regarding chronic injuries (e.g., fibrosis and carcinogenesis) induced by nanoparticles raised public health concerns and need to be rapidly assessed in hazard identification. Although in silico analysis is commonly used for risk assessment of chemicals, predicting chronic in vivo nanotoxicity remains challenging due to the intricate interactions at multiple interfaces like nano-biofluids and nano-subcellular organelles. Herein, we develop a multimodal feature fusion analysis framework to predict the fibrogenic potential of metal oxide nanoparticles (MeONPs) in female mice. Treating each nano-bio interface as an independent entity, eighty-seven features derived from MeONP-lung interactions are used to develop a machine learning-based predictive framework for lung fibrosis. We identify cell damage and cytokine (IL-1β and TGF-β1) production in macrophages and epithelial cells as key events closely associated with particle size, surface charge, and lysosome interactions. Experimental validations show that the developed in silico model has 85% accuracy. Our findings demonstrate the potential usefulness of this predictive model for risk assessment of nanomaterials and in assisting regulatory decision-making. While the model is developed based on 52 MeONPs, further validation using a larger nanoparticle library is necessary to confirm its broader applicability. The prediction of chronic toxicity is a major challenge in nanotoxicity studies. Here, the authors present an in silico framework for predicting ENM-induced lung fibrosis, displaying 85% accuracy in experimental validation and leading to identification of key events at nano-bio interfaces that allows mechanism interpretation of ENM-induced lung fibrosis.