We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Predicting the toxicity of microplastic particles through machine learning models
Summary
Researchers developed machine learning models to predict microplastic particle toxicity from physical and chemical descriptors, addressing the classification challenge posed by the enormous diversity of particle types that cannot be characterized using conventional chemical hazard methods. The models provided accurate toxicity predictions across diverse microplastic types, offering a practical screening tool for the field.
Unlike chemicals, which can be identified based on their well-defined molecular structure and stable properties (e.g. CAS numbers, SMILES), micro- and nanoplastic particles (MNP) lack such straightforward classification. Each MNP has its own combination of traits, including characteristics such as polymer composition, particle dimensions (length and width), and shape, as well as physicochemical properties like surface charge, surface chemistry, and plastic-associated chemicals. Furthermore, these traits may change over time, particularly due to degradation processes when MNPs are exposed to natural environments. To achieve reliable hazard and risk assessments for MNPs, it is thus necessary to predict the toxicity of MNPs with trait combinations that have not been tested directly in the lab. Similar to Quantitative Structure-Activity Relationship (QSAR) models that link molecular structures of chemicals to toxic outcomes, models are needed to link MNPs traits to their toxicities. The recently compiled Toxicity of Microplastics Explorer (ToMEx) 2.0 database consisting of 13,412 data points from 290 published studies of MNP effects on aquatic species offers a unique opportunity to approach this task. Using ToMEx 2.0 data, we trained machine learning models on tasks to predict the toxicity (presence/absence of effects, effect direction, effective concentrations) of untested MNP. We compare the predictive performance of two machine learning algorithms (boosted regression trees and deep neural networks) and use methods of explainable AI (average marginal effects) to gain insights into the relationships between toxic outcomes and MNP traits, experimental parameters, and species traits. We also compared whether the predictions change when filtering the dataset according to study quality based on assigned quality scores. Finally, we discuss how such models can be used to predict the toxicity of environmentally relevant mixtures of MNPs and how they contribute to the development of less toxic and more environmentally friendly plastic materials in the future. Also see: https://micro2024.sciencesconf.org/559370/document