ninety-five percent of chemicals, never tested

Problem statement

More than 350,000 chemicals and mixtures of chemicals are registered for production and use worldwide — up to three times as many as previously estimated (Wang et al. 2020). Only a small fraction carry a completed human health assessment: among chemicals on the U.S. TSCA active inventory, just 1.8% have a published human health assessment from EPA or ATSDR, and developing one "typically takes at least 4 years" (EPA ETAP value-of-information case study, 2024). The conventional route is not scalable to the backlog — current approaches "rely on extensive animal testing, cost millions of dollars, and can take 2–3 years per chemical" (Judson et al. 2009). Newer high-throughput screening methods (in vitro assays, computational toxicology) flag molecular-level bioactivity but have not been shown to reliably predict chronic organism-level effects like endocrine disruption, developmental neurotoxicity, or carcinogenicity from low-dose, long-duration exposure — the exposure pattern most relevant to human health.

Why this matters

Pollution of all kinds remains responsible for an estimated 9 million premature deaths per year — one in six deaths worldwide (Fuller et al. 2022). Deaths from "modern" pollution risk factors — ambient air pollution and toxic chemical pollution such as lead — rose 7% between 2015 and 2019 and by more than 66% since 2000, while deaths from "traditional" pollution (household air pollution, unsafe water) fell. Data coverage runs in the opposite direction from where the gap is widest: because REACH registration deadlines took the highest-tonnage substances first, the largest published analysis of registered safety data is "clearly biased toward high-production volume substances" (Luechtefeld et al. 2016) — lower-volume chemicals get the least scrutiny of all. PFAS compounds circulated for decades before their persistence and bioaccumulation were characterized; similar latent risks likely exist among thousands of newer compounds.

What’s been tried and why it hasn’t worked

The EPA's ToxCast/Tox21 program has screened almost 10,000 substances using high-throughput in vitro assays aggregated from more than 20 assay sources, but these assays measure molecular-level endpoints (receptor binding, enzyme inhibition) rather than organism-level chronic effects. The most systematic test of whether they bridge that gap came back negative: a cross-validation of more than 600 ToxCast phase I assays against 60 in vivo endpoints using 84 different statistical classification methods found that, with the exception of chronic in vivo cholinesterase inhibition, "the overall predictive power of both the in vitro assays and the chemical descriptors was relatively low," the assays did no better than chemical structure descriptors alone, and prefiltering assay data outside the cross-validation loop — as earlier studies had done — "significantly biased estimates of model performance." The authors concluded the assays have "limited applicability for predicting in vivo chemical hazards using standard statistical classification methods" (Thomas et al. 2012). QSAR (quantitative structure-activity relationship) models trained directly on the Tox21 assay data perform better but vary widely and unpredictably by endpoint — AUC-ROC 0.55–0.87 across the Tox21 Challenge assays (Capuzzi et al. 2016) — and what they predict is an in vitro assay result, not an in vivo chronic effect. Organ-on-chip and organoid models offer better physiological relevance than cell-based assays but face the same challenge as all in vitro methods: no validated mapping exists from in vitro response to in vivo chronic effect at population scale. The European REACH regulation requires toxicity information for substances manufactured or imported above 1 tonne per year, with requirements rising by tonnage band — but a review of publicly available data for 9,801 registered substances found 5,551 of them (56.6%) "were missing explicit in vivo key experimental skin sensitization studies," registrants instead relying on read-across and study waivers.

What would unlock progress

An integrated approach combining high-throughput in vitro screening, QSAR prediction, and human biomonitoring data (exposome studies) could enable tiered screening that prioritizes the most concerning chemicals for deeper assessment. The key missing piece is a validated adverse outcome pathway (AOP) framework that connects molecular initiating events (measurable in vitro) to adverse outcomes in organisms. If even 10 well-characterized AOPs were validated for the most common toxicity endpoints, they could serve as bridges from high-throughput data to regulatory decisions.

Entry points for student teams

A student team could take an existing ToxCast dataset for a specific chemical class (e.g., organophosphate flame retardants) and attempt to build a predictive model linking in vitro assay results to known in vivo outcomes from the literature, evaluating which assay combinations are most predictive. Alternatively, teams could develop visualization tools for the ToxCast/Tox21 database that help regulators identify chemicals with concerning activity profiles that lack in vivo studies. Relevant disciplines: computational biology, toxicology, data science, environmental engineering, public health.