From experimental toxicity to computational prediction: a critical assessment of reliability, uncertainty, and applicability domains for graphene-family nanomaterials

Graphene-family nanomaterials (GFNs), including graphene, graphene oxide, and reduced graphene oxide, exhibit diverse physicochemical properties that can produce different biological responses. Reported toxicity outcomes vary with lateral dimensions, layer number, oxidation state, defect density, surface chemistry, aggregation, dispersion conditions, dose metrics, exposure duration, and assay design. Recent analyses show that adsorption, optical interference, fluorescence quenching, catalytic effects, residual impurities, and endotoxin contamination can distort conventional toxicity measurements, making apparently conflicting results difficult to interpret. These limitations become more consequential when experimental observations are converted into computational datasets, because heterogeneous characterization methods, endpoint definitions, units, dose descriptors, and binary toxicity labels can introduce uncertainty before model development begins. Computational approaches, including quantitative structure-activity relationship modelling, nano-QSAR, density functional theory-derived descriptors, pharmacokinetic descriptors, and machine learning (ML), can relate GFN characteristics to biological outcomes. Yet predictive performance depends strongly on data quality, descriptor selection, sample size, class balance, feature redundancy, validation strategy, and the similarity between training materials and the GFN being predicted. High accuracy or a high area under the receiver operating characteristic curve therefore cannot, by itself, establish confidence in a prediction. Uncertainty may arise from experimental measurements, descriptor generation, data curation, model structure, biological variability, and extrapolation beyond the training domain. This review critically examines the progression from experimental GFN toxicity assessment to computational prediction, with particular attention to how uncertainty is introduced and propagated across this workflow. A descriptor-centered perspective is developed that links structural, physicochemical, electronic, surface, colloidal, pharmacokinetic, and toxicological information with their measurement context. Applicability-domain assessment is considered an essential step for distinguishing interpolation from prediction of previously unseen GFNs. To address this gap, we propose a five-level reliability framework that distinguishes overall model performance from prediction-specific confidence by integrating experimental data quality, descriptor reliability, model reliability, applicability-domain evidence, and independent experimental confirmation. The framework separates model accuracy from prediction confidence and provides a structured basis for assigning confidence according to the strength and consistency of evidence across these levels. Such an approach can support computational nanotoxicology and help prioritize experiments, identify uncertainty-driving material properties, and guide material design. By linking characterization quality with model validation and confidence assessment, computational prediction can be positioned as complementary evidence rather than a substitute for experimental toxicity testing.