AI Just Ranked the Most Dangerous Chemicals in Your Body

Every year, hundreds of new chemicals enter the water supply, the food chain, and human tissue before regulators know whether they are dangerous. Traditional toxicology tests each substance individually — a process that takes years and costs millions. Now, a January 2026 study published in the International Journal of Applied Resilience and Sustainability has demonstrated that an AI-driven approach can compress that timeline dramatically, correctly ranking known hazardous compounds with 83% accuracy using only their measurable physical and chemical properties.

We're Surrounded by Chemicals That Science Hasn't Caught Up With Yet

Emerging contaminants (ECs) are substances that appear in the environment — often through industrial processes or consumer products — before toxicology data exists to assess their danger. The category spans a wide range: PFAS compounds found in non-stick cookware and firefighting foam, microplastics smaller than 5 mm now detected in human blood, endocrine-disrupting chemicals such as bisphenol A, and novel pesticides like the neonicotinoid imidacloprid. Most have been commercially active for years before any formal health risk assessment begins. Regulators face tens of thousands of chemicals and the conventional toolkit — animal studies, epidemiology — cannot keep pace.

Inside the Model: How Four Algorithms Competed to Rank Chemical Risk

Joseph Ozigis Akomodi (New York City Department of Education) and Birupaksha Biswas (Burdwan Medical College & Hospital, India) assembled a feature matrix for six representative contaminants, combining physicochemical descriptors — including lipophilicity (log K_ow), molecular weight, and environmental persistence — with toxicological bioassay indicators such as estrogen and androgen receptor activity, plus exposure proxies like detection frequency in water samples and annual production volume. They trained four models across 5-fold cross-validation: logistic regression, Random Forest, XGBoost, and a multi-layer perceptron neural network. Every model was evaluated on accuracy, precision, recall, F1 score, and area under the ROC curve.

The Single Property That Predicts Whether a Chemical Will Harm You

The most striking result from the feature importance analysis is how much a single property — a chemical's affinity for fat over water, measured as log K_ow — drives the model's decisions, accounting for 20% of its predictive power. Lipophilic chemicals accumulate in fatty tissue, extending a person's internal exposure far beyond what ambient environmental concentrations suggest. PFOA carries a log K_ow of 2.7 and BPA scores 3.4; both received high-risk probabilities exceeding 0.90 from the model. Environmental persistence ranked second at 18%: a compound that degrades slowly keeps exposing populations long after its source is controlled. Endocrine activity — whether a chemical triggers estrogen or androgen receptors — ranked third at 17%, correctly reflecting decades of evidence that hormonal disruption occurs at doses orders of magnitude lower than those needed to cause direct toxicity.

What This Means for the EPA — and Every Regulator Playing Chemical Catch-Up

The practical payoff is speed. A conventional risk evaluation requires years of animal studies per chemical; a trained model scores a new compound in real time, provided it has the necessary input features. Regulatory agencies could deploy such a tool as a front-line filter — rapidly identifying the top 5% of untested chemicals that demand urgent investigation and directing limited testing resources accordingly. The study's authors note that the model flagged silver nanoparticles as a potential concern, a finding that could prompt targeted chronic toxicity studies long before population-level exposure reaches critical levels. The approach also bridges disciplines: a high AI risk score for a compound can trigger epidemiologists to study exposed communities, while emerging epidemiological findings can feed back into future model training.

This Study Tested Six Chemicals. It Needs to Test Thousands. Here's What's Missing.

The authors are candid about the study's limits. The dataset covers only six contaminants — sufficient as a proof of concept, but far too small for a production-ready regulatory tool. Risk labels were assigned by expert judgment, meaning the model cannot surpass the boundaries of current scientific knowledge; a genuinely novel mechanism of harm that no expert has yet characterized will remain invisible to it. Exposure proxies — primarily detection frequency in water samples — are crude substitutes for actual human dose estimation. Future versions of this framework need environmental fate modeling to translate production volumes and degradation rates into realistic human intake figures. The authors also identify mixture risk assessment as a critical frontier: people encounter dozens of contaminants simultaneously, and single-chemical models cannot capture synergistic effects. Closing those gaps will determine whether this AI framework moves from promising demonstration to essential regulatory infrastructure.

• Scale the dataset first — expanding from 6 to hundreds of chemicals would allow robust multi-class risk ranking and reduce reliance on expert-assigned labels.
• Add exposure modeling — replacing detection-frequency proxies with environmental fate models would translate chemical properties directly into estimated human intake doses.
• Build living models — continuous retraining as new toxicology data arrives prevents the framework from becoming a static snapshot of today's incomplete knowledge.

"The AI model on a selection of representative emerging contaminants performed encouragingly and was able to correctly identify known high-risk substances, including PFAS and endocrine disruptors, and provide plausible predictions on additional less-studied contaminants." — Akomodi & Biswas, International Journal of Applied Resilience and Sustainability, 2026.