The review, conducted by Abir Hamze and Emran Alotaibi of Canadian University Dubai, is structured as a PRISMA-ScR-informed evidence map rather than a conventional narrative survey. The authors searched bibliographic databases, publisher platforms, Google Scholar and citation-chasing sources, identifying 134 corpus entries before duplicate removal. After screening, 126 unique records formed a broad map of AI applications to PFAS, spanning detection, environmental fate, treatment, exposure and toxicity. The team then applied a stricter, endpoint-based reclassification to isolate the toxicological core of the literature. Of 46 records initially coded as toxicity or human and ecological health, only 39 survived as genuine AI-PFAS toxicology studies, and detailed full-text extraction was possible for 35 of them. The distinction matters: many studies that mention PFAS and health in the same breath are actually modeling environmental occurrence or exposure classification, not biological hazard.

Within that core subset, the methodological landscape is strikingly heterogeneous. Molecular simulation combined with machine learning emerged as the family producing the strongest-validated examples, while QSAR, QSPR and read-across studies, the traditional workhorses of computational toxicology, often clustered at intermediate validation levels. Tree ensembles, boosted-tree algorithms, Bayesian probabilistic models, neural networks and graph-based approaches all appear in the literature, each suited to different tasks. The authors emphasize that no algorithm is inherently decision-ready: a random forest may serve tabular biomonitoring data well if externally validated, a graph neural network may excel at structure-based prediction when scaffold-split validation is reported, and Bayesian models are particularly attractive for sparse exposure data because they make uncertainty explicit rather than hiding it behind a single deterministic number.

One of the review’s most valuable contributions is its grounding of AI prediction in mechanistic toxicology. Drawing on recent specialist reviews, the authors identify several biological anchors that any credible PFAS model must respect. Reproductive toxicity work points to reactive oxygen species as a molecular initiating event that cascades through oxidative stress, mitochondrial dysfunction, steroidogenic disruption and thyroid-hormone changes, ultimately impairing gametogenesis and fertility. Transporter-focused research highlights serum albumin, solute carrier and ATP-binding cassette transporters, and nuclear receptors as critical but underdeveloped determinants of where PFAS go in the body. Early-life studies identify the placenta as a sensitive target that may mediate latent disease decades after exposure. Omics research, mixture toxicology and underrepresented endpoints such as neurotoxicity and ocular toxicity round out the picture. An AI model whose predictions contradict these established pathways, or whose feature-importance scores ignore chain length, functional group or protein-binding chemistry, should raise immediate red flags.

The clearest practical role for AI, the authors argue, is prioritization. It is simply impossible to test thousands of PFAS comprehensively in animal studies or in vitro batteries, so QSAR models, read-across frameworks, molecular descriptors and generative-AI workflows can triage which compounds and endpoints deserve earlier experimental attention. But the review insists this only works as a triage tool, never as a substitute for evidence. Before any such model is built, chemical data curation must come first: structure standardization, salt stripping, explicit handling of acids, salts, branched isomers and tautomers, duplicate removal, harmonization of chemical identifiers and careful treatment of missing values. Because the same PFAS substance may appear in databases as an acid, a salt, a precursor or a transformation product, inconsistent representation can silently distort descriptor values and inflate apparent model performance.

Validation design receives particularly sharp criticism. The authors show that random train-test splitting, the default in much of the machine learning literature, can seriously overstate performance for PFAS because structurally similar homologues, samples from the same laboratory or geographically clustered cohorts can end up in both training and test sets. More informative alternatives include scaffold splits, leave-family-out or leave-subclass-out validation, temporal and spatial validation, and independent-cohort testing for biomonitoring models. Applicability-domain assessment should likewise move beyond vague disclaimers, using tools such as Williams plots, distance-to-model measures, conformal prediction and ensemble disagreement to specify whether a prediction actually applies to long-chain legacy compounds, short-chain acids, ether PFAS, polymers or specific biological matrices. These requirements align directly with the OECD’s long-standing QSAR validation principles, which the authors argue apply with full force to modern deep learning and graph-based models.

The audit of the broader 126-record corpus quantifies how far the field still has to go. Using the full map as denominator, the authors found that 73 records, or 57.9 percent, showed no or limited external validation; 49 records, or 38.9 percent, lacked clear experimental or field validation; 41 records, or 32.5 percent, reported no formal uncertainty or applicability-domain analysis; and 28 records, or 22.2 percent, had no clearly available code. Bias compounds these gaps in specific directions: training data concentrate heavily on PFOA, PFOS and a few legacy analogues; available assays cover hepatotoxicity and endocrine activity far better than immune, developmental, neurological or reproductive endpoints; and monitoring data cluster in known hotspots and wealthy jurisdictions, leaving children, pregnant people, occupational groups, private-well users and under-monitored communities underrepresented.

On the exposure side, the review acknowledges genuine promise alongside real peril. Machine learning models that predict which adults carry high PFAS burdens, or that integrate biomonitoring with environmental data, can help target confirmatory sampling, drinking-water treatment, dietary guidance and public-health follow-up. Yet serum concentrations reflect a tangle of drinking water, diet, dust, occupation, geography, renal function, pregnancy and protein binding, and a model may classify high exposure accurately while learning proxies for monitoring intensity rather than true exposure determinants. Interpretable association studies linking PFAS to diabetes and cardiovascular-kidney-metabolic outcomes illustrate both the analytical potential and the danger of causal overinterpretation. Mixture context adds another layer: because humans encounter shifting PFAS cocktails rather than isolated compounds, the authors argue that regulatory-grade models must support cumulative assessment and propagate uncertainty, with Bayesian frameworks offering a natural fit.

To translate these findings into practice, the review proposes a tiered regulatory-readiness framework that scales evidence requirements to decision consequence. Exploratory screening can tolerate uncertainty as long as it triggers further testing. Prioritization demands fit-for-purpose validation, descriptor transparency and domain checks. Supporting evidence requires external or biological validation, mechanistic plausibility and reproducible workflows. Decision-critical uses, such as regulatory read-across, testing waivers, restrictions, cleanup priorities or public-health communication, demand independent validation, calibrated uncertainty, explicit applicability domains, audit trails, expert review and stakeholder transparency. The authors also propose a minimum reporting checklist covering data provenance, chemical identifiers, validation, uncertainty, interpretability and code availability, and they sketch a closed-loop workflow in which AI generates hypotheses, human-relevant new approach methodologies such as organoids, receptor assays and omics platforms test them, and results feed back to update models and shrink uncertainty.

The review’s bottom line is neither technophobic nor uncritically enthusiastic. AI can genuinely accelerate PFAS hazard identification, bioactivity prediction, toxicokinetic inference, protein-binding assessment and ecological screening at a pace conventional testing cannot match. But prediction should never be mistaken for proof: a docking result suggests binding without demonstrating harm, an omics signature shows association rather than mechanism, and a low predicted risk is not evidence of safety. The authors are explicit about what AI must never be used to do, including declaring a compound safe because it falls below a predicted threshold, waiving testing outside a model’s domain, or treating absent monitoring as absent exposure. With transparent reporting, mechanistic anchoring, equity-sensitive validation and human oversight, they conclude, artificial intelligence can make PFAS toxicology faster, broader and more trustworthy, supporting rather than replacing the measurement, biological validation and expert judgment on which environmental-health protection ultimately depends.

Subject of Research: Artificial intelligence applications in PFAS toxicology and environmental-health risk assessment

Article Title: Artificial intelligence for PFAS toxicology and risk assessment

Article References: Hamze, A., & Alotaibi, E. (2026). Artificial intelligence for PFAS toxicology and risk assessment. Discover Artificial Intelligence, 6(1), Article 1291. https://doi.org/10.1007/s44163-026-02266-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02266-0

Keywords: PFAS, artificial intelligence, machine learning, toxicology, QSAR, read-across, risk assessment, biomonitoring, adverse outcome pathways, uncertainty quantification, regulatory science, environmental health