High Data Quality Enhances Microplastic Toxicity Prediction

Ana L. Antonio Vital, Scott Coffin, Andrea Bonisoli-Alquati, Maaike Vercauteren, Luan de Souza Leite, Maximilian Pichler, Magdalena M. Mair 

Unlike chemicals, microplastic (MP) particles lack standardized identifiers, limiting the applicability of traditional predictive ecotoxicology methods such as quantitative structure-activity relationship (QSAR) models. By leveraging an extensive database of MP effects, this study aimed to predict MP toxicity using MP properties, MP concentration, organismal traits, endpoints, and experimental design as features, and to evaluate how data pre-processing, dataset size, and data quality influence model performance. Specifically, we applied the Boosted Regression Tree (BRT) machine learning algorithm to four datasets derived from the Toxicity of Microplastics Explorer database (ToMEx 2.0): (i) imputed missing values, (ii) complete-case (missing values removed), (iii) high-quality data, and (iv) low-quality data. The high-quality dataset yielded the best final predictions with highest area under the curve (AUC) values for both random cross-validation (AUC = 0.93) and blocked cross-validation by particle identifier schemes (AUC = 0.87). Explainable artificial intelligence (xAI) analyses showed that predictive performance was primarily determined by endpoints and concentration, with contributions by MP properties despite limited reporting. Our findings demonstrate the feasibility of predicting MP toxicity and identifying key drivers for its predictive performance. They further indicate that high-quality data improve predictive performance, reducing the need for extensive data mining efforts and computational costs. Standardized experimental design, detailed characterization of MP properties, and rigorous reporting practices will enhance toxicity predictions, support more reliable risk assessment frameworks, and inform the design of safer materials in the future.