Artificial intelligence for predictive mixture toxicology

Identificador:  imarina:9625991

Handle: https://hdl.handle.net/20.500.11797/imarina9625991

Autors:  Domingo, Jose L; Souza, Marilia Cristina Oliveira; Barbosa Jr, Fernando

Resum:
Human populations and ecosystems are continuously exposed to complex mixtures of environmental contaminants rather than to individual chemicals in isolation. These mixtures include pesticides, metals and metalloids, persistent organic pollutants, endocrine-disrupting chemicals, per- and polyfluoroalkyl substances, pharmaceuticals, plastic-associated compounds, air pollutants, nanomaterials, and numerous poorly characterized substances. Their combined effects may be additive, synergistic, or antagonistic, and are strongly influenced by dose, component ratio, timing, exposure sequence, and biological susceptibility. Experimental evaluation of all environmentally relevant mixtures is infeasible because the number of possible combinations increases combinatorially. Artificial intelligence (AI) offers a possible way to address this limitation. Machine learning, deep learning, graph neural networks, Bayesian approaches, and natural-language processing can integrate heterogeneous data, including chemical structures, molecular descriptors, toxicokinetics, high-throughput screening results, omics profiles, adverse outcome pathways, biomonitoring data, and epidemiological findings. However, the evidence base is uneven. Relatively few studies have applied AI directly to experimentally characterized mixtures, and much of the current optimism is extrapolated from single-chemical toxicology. This review distinguishes explicitly between applications demonstrated in mixtures, proof-of-concept mixture applications, and approaches whose mixture use remains prospective. It further examines the methodological requirements for mixture prediction, including dose and ratio representation, additivity reference models, applicability domains, external validation, and mechanistic interpretability