Assessing comparable bioconcentration potentials for nanoparticles in aquatic organisms via combined utilization of machine learning and toxicokinetic models

Abstract

Abstract The toxicokinetic (TK) model‐derived kinetic bioconcentration factor (BCFk) provides a quantitatively comparable index to estimate the bioaccumulation potential of nanoparticles (NPs) that barely reach thermodynamic equilibrium in aquatic organisms, but experimental data are limited for various NPs. In the present study, a machine learning model was applied to offer reliable in silico predictions for the dynamic body burden of diverse NPs to derive corresponding parameters for the TK model. The developed eXtreme Gradient Boosting‐derived TK (XGB‐TK) model was applied to predict BCFk results for a broad range of metallic or carbonaceous NPs, with an appreciable prediction R2 of 0.96. The BCFk values were predicted based on a random combination of selected variable features, revealing that their bioaccumulation potential showed an overall negative correlation with NP density or organism size. By applying importance analysis and partial dependence plots, NP density and organism size were revealed to be the top essential features that impact the bioaccumulation potential. The conjunctively used XGB‐TK model enabled a prior comparison for diverse NPs and straightforward derivation on the dependency of features, which could also guide the bioaccumulation mechanism exploration and experimental condition formulation.

Keywords

• aquatic organism
• BCF
• bioaccumulation
• machine learning
• nanoparticles
• toxicokinetic models