Predictive Nanotoxicology of Graphene-Family Nanomaterials: Standardised Descriptors, Explainable Machine Learning and Safe-by-Design

Graphene-family nanomaterials (GFNs) comprise structurally and chemically diverse two-dimensional materials with broad applications in biomedicine, environmental technologies, energy, sensing, and advanced materials. Their expanding use has increased the need for reliable approaches to characterize and predict potential biological and environmental hazards. However, meaningful comparison of GFN toxicity data remains difficult because biological responses depend on interacting material characteristics, including lateral dimensions, layer number, defects, oxidation, surface functionalization, charge, morphology, aggregation, and dispersion state, together with exposure conditions, dosimetry, biological models, and endpoint definitions. This review examines how standardized descriptors can provide a quantitative basis for connecting GFN characteristics with toxicological responses and constructing reproducible predictive datasets. Structural, chemical, surface, colloidal, electronic, computational, exposure and biological descriptors are considered within an integrated framework that distinguishes intrinsic material properties from exposure-state characteristics. Particular emphasis is placed on descriptor harmonization, dataset curation, data provenance, and consistent reporting of measurement and dose information. The review also discusses machine-learning approaches for modelling nonlinear and multidimensional relationships between GFN descriptors and toxicity, emphasising the significance of proper feature selection, validation, applicability-domain assessment and uncertainty quantification. Because GFN datasets may contain limited numbers of independently characterized materials and substantial variation in experimental protocols, predictive performance should be interpreted in the context of data quality, validation design, and applicability-domain limitations. Explainable machine learning is considered as a means of identifying influential descriptors and generating mechanistically testable hypotheses rather than causal conclusions. Building on these principles, a standardized predictive framework is proposed that links characterization, descriptor harmonization, toxicity datasets, validated and explainable machine learning, experimental confirmation, and iterative model refinement. The framework provides a foundation for using predictive insights to guide material modification and support Safe-by-Design and Safe-and-Sustainable-by-Design strategies. Overall, standardized descriptors and traceable data practices can help advance GFN nanotoxicology from heterogeneous retrospective comparisons toward reproducible, interpretable, and experimentally testable prediction.