Structure-activity relationship of graphene-related materials: A meta-analysis based on mammalian in vitro toxicity data

Abstract

To support a safe application of graphene-related materials (GRMs) it is necessary to understand the potential negative impacts they could have on human health, in particular on the lung - one of the most sensitive exposure routes. Machine learning (ML) approaches can help analyse the results of multiple toxicity studies to understand the structure-activity relationship and the effect of experimental conditions, thus supporting predictive nanotoxicology. In this work we collected in vitro cytotoxicity data obtained from studies using lung cells; we then fitted multiple regression models to predict this endpoint based on the material properties and experimental conditions. Moreover, the data set was used to calculate the Benchmark Dose Lower Confidence Interval (BMDL), a dose descriptor widely used in risk assessment. Regression and classification models were applied for the prediction of the BMDL value and BMDL range. The analyses show that both cytotoxicity and the BMDL range can be predicted well (Q2 = 0.77 and accuracy = 0.71, respectively). Both physico-chemical characteristics such as the lateral size, number of layers, and functionalization, and experimental conditions such as the assay and media used were important predicting features, confirming the need for thorough characterization and reporting of these parameters.