Atom-level descriptors and explainable prediction of iron carbide nanoparticles' cytotoxicity via the Enalos Cloud platform

Abstract

Iron carbide nanoparticles (ICNPs), a distinct type of magnetic nanostructure, have been proposed as novel candidate therapeutic agents for a wide range of biomedical applications, yet their biocompatibility remains a critical concern for their safe implementation. To mitigate the need for excessive experiments that screen bio-related interactions through conventional pathways, in silico methodologies have been established as cost and time-efficient alternatives. This study presents the development of data-driven workflows for the risk assessment of ICNP-induced cytotoxicity. Two modelling strategies were established: an evidence-based approach relying on experimental features and exposure conditions, and an atomistic-based approach combining attributes that describe NPs at the atomistic scale. While the former methodology struggled to meet the requirements for robust quantitative predictions, the models implemented on the enriched dataset displayed improved accuracy. The resultant Random Forest model fully adheres to the principles outlined by the OECD for the development of quantitative structure–toxicity relationship (QSTR) models. Beyond deciphering the mechanistic influence of individual features, the effect of the selected nanodescriptors was reviewed through Shapley additive values and permutation importance analyses to reveal key input characteristics that drive cell viability predictions. The produced model is disseminated as a free to use web service hosted by the Enalos CHIASMA Cloud Platform (https://www.enaloscloud.novamechanics.com/chiasma/icnp/) and data are publicly available through the NanoPharos database for easy access to the nano-safety community.