Integrating Structure and Experimental Data Annotations with Machine Learning Approaches to Develop Micro-nanoplastics Toxicity Models

Micro-nanoplastics (MNPs) are increasingly released into the environment, raising concerns about their impact on human health. Traditional toxicity testing is costly, time-consuming, and lacks universally accepted protocols, making computational modeling using machine learning (ML) approaches a promising alternative. However, current MNP models are limited by insufficient high-quality data and inadequate representation of complex particle structures. To address this, we constructed three MNP datasets with common toxicity endpoints and generated virtual MNPs (vMNPs) using nanostructure annotation techniques. Geometrical descriptors were derived through Delaunay Tessellation, while key experimental factors were incorporated as additional variables. Partial least squares regression (PLSR) models were then developed and validated using leave-one-out cross-validation. The models showed good performance in predicting toxicity potentials of new MNPs. Moreover, a vMNP library with predicted properties and bioactivities was constructed to support future research. This study provides three novel ML models and a scalable strategy for assessing MNP toxicity and expanding prediction to other toxicity endpoints.