In silico assessment of nanoparticle toxicity powered by the Enalos Cloud Platform: Integrating automated machine learning and synthetic data for enhanced nanosafety evaluation

Abstract

The rapid advance of nanotechnology has led to the development and widespread application of nanomaterials, raising concerns regarding their potential adverse effects on human health and the environment. Traditional (experimental) methods for assessing the nanoparticles (NPs) safety are time-consuming, expensive, and resource-intensive, and raise ethical concerns due to their reliance on animals. To address these challenges, we propose an in silico workflow that serves as an alternative or complementary approach to conventional hazard and risk assessment strategies, which incorporates state-of-the-art computational methodologies. In this study we present an automated machine learning (autoML) scheme that employs dose-response toxicity data for silver (Ag), titanium dioxide (TiO₂), and copper oxide (CuO) NPs. This model is further enriched with atomistic descriptors to capture the NPs' underlying structural properties. To overcome the issue of limited data availability, synthetic data generation techniques are used. These techniques help in broadening the dataset, thus improving the representation of different NP classes. A key aspect of this approach is a novel three-step applicability domain method (which includes the development of a local similarity approach) that enhances user confidence in the results by evaluating the prediction's reliability. We anticipate that this approach will significantly expedite the nanosafety assessment process enabling regulation to keep pace with innovation, and will provide valuable insights for the design and development of safe and sustainable NPs. The ML model developed in this study is made available to the scientific community as an easy-to-use web-service through the Enalos Cloud Platform (www.enaloscloud.novamechanics.com/sabydoma/safenanoscope/), facilitating broader access and collaborative advancements in nanosafety.

Keywords: Automated machine learning; Nanoinformatics; Safety and sustainability by design; Synthetic data.

Graphical abstract

Fig. 1

Schematic workflow of the data analysis, modelling of the toxicity endpoints, and release of the final model.

Fig. 2

Accuracy statistics of the random forest model when applied on the test and blind datasets.

Fig. 3

Left: qualitative sketch of a TiO₂ NP (of 5 nm diameter) depicting the average potential energy per atom. Ti atoms are depicted in pink colour and O atoms are depicted in red colour. Right: the number of neighbouring atoms for a Ti atom in the shell (upper NPs) and in the core (lower NPs) at 3, 4, and 5 Å from the Ti atom. The neighbouring atoms are highlighted in the graphic with increased thickness.

Fig. 4

SafeNanoScope web application interface. Users can either input the required descriptors via the provided form or by uploading a CSV file with all the required properties. The output is a prediction of each NPs’ toxicity class and a comment on the reliability of the prediction based on the model’s domain of applicability.