Artificial intelligence for PFAS toxicology and risk assessment

Abstract

Per- and polyfluoroalkyl substances (PFAS) pose a major environmental toxicology and public-health risk-assessment challenge because their chemical space is large, persistent, structurally diverse, mixture-dominated and unevenly characterized. The difficulty is practical as well as scientific: risk assessors must often make decisions about monitoring, testing, exposure prevention and remediation before complete toxicological evidence is available for the many legacy, emerging, short-chain, ether-based, precursor and transformation-product PFAS. Artificial intelligence (AI), machine learning (ML), quantitative structure–activity/property relationships (QSAR/QSPR), read-across, molecular simulation, omics analytics and probabilistic models are increasingly used to prioritize PFAS, predict bioactivity, estimate exposure, interpret biomonitoring data and support risk assessment. However, predictive performance alone is insufficient for decision-making; model outputs intended to inform toxicology, testing priorities, read-across, public-health communication or regulatory action must be validated, uncertainty-aware, reproducible and constrained by clearly defined applicability domains. This review presents a PRISMA-ScR-informed, two-tier evidence map of AI applications relevant to PFAS toxicology and risk assessment. Searches of databases, publisher platforms, Google Scholar, PDF batches and citation-chasing sources identified 134 corpus entries. After removing 8 duplicate records, 126 unique records were retained in the broad AI-PFAS evidence map. Of 46 records initially coded as toxicity or human/ecological health, endpoint-based reclassification identified 39 core AI-PFAS toxicology records; 35 underwent detailed full-text extraction and 4 were retained at abstract-level coding. The core evidence base covered QSAR/read-across, bioactivity prediction, protein binding, toxicokinetics, receptor activity, endocrine disruption, neurotoxicity, reproductive toxicity, immunotoxicity, omics, mixture effects and ecological toxicity. Stronger studies combined interpretable or mechanism-linked models with external, biological or field validation, explicit applicability-domain statements, uncertainty analysis and reproducible reporting. Recurring limitations included optimistic validation designs, sparse toxicity data for emerging PFAS, inconsistent chemical identifiers, limited uncertainty quantification, weak applicability-domain reporting and incomplete code availability. In response, this Review expands the regulatory-readiness framework to incorporate OECD QSAR validation principles, chemical data curation, descriptor generation, applicability-domain assessment, scaffold-aware and prospective validation, explainability beyond SHAP, calibrated uncertainty analysis and AOP-linked mechanistic interpretation. The framework separates exploratory screening, prioritization, supporting evidence and decision-critical use. AI can accelerate PFAS hazard identification and exposure prevention, but it should support, not replace, measurement, biological validation, expert review and transparent environmental-health decision-making. A PRISMA-ScR-informed two-tier evidence map separated 126 broad AI-PFAS records from 39 core AI-PFAS toxicology records. The strict core subset retained studies with direct toxicological, bioactivity, toxicokinetic, mechanistic, omics or ecotoxicological endpoints. Mechanistic anchors include reproductive AOPs, transporter and protein-binding biology, early-life effects, hepatotoxicity, proteomics, neurotoxicity, mixture risk and NAMs. Reliable AI studies combine interpretable models, external or biological validation, uncertainty analysis, applicability domains, reproducible reporting and explicit regulatory boundaries. Regulatory use should distinguish exploratory screening, prioritization, supporting evidence and decision-critical read-across.