Dose-Response Curves: Why 'Safe Doses' Aren't Always Safe

For most chemicals the model holds. Industrial poisons, heavy metals, classical solvents — the more you have in your blood, the worse the outcome, and the dose-response curve tells you where to set the limit. The trouble is the chemicals the consumer audience actually asks about. Hormones, and the synthetic molecules built to mimic them, don't follow the line. Some of them produce their largest biological effect at their lowest tested dose, then taper off — or reverse — as the dose climbs into the range a regulator usually tests. The signal hides below the floor of the experiment that was supposed to find it.

This is the methodology question that the combination-effects article in this series flagged at the close. Cocktail effects describe how chemicals add up across exposures. Non-monotonic dose-response — NMDR — describes how a single chemical's curve can bend the wrong way below the dose the regulator measured. The two arguments together explain why the 'tested for safety' line on the back of a bottle promises less than most readers assume. Part of our endocrine-disruptors guide. The dose makes the poison — except when it doesn't.

The dose-response relationship is the connection between the amount of a chemical an organism is exposed to and the size of the biological effect that chemical produces. It is the foundational assumption of modern toxicology, traceable to a sentence the Swiss physician Paracelsus wrote in 1538: 'the dose alone makes the poison.' Every modern safety standard — every TDITolerable Daily Intake — the daily dose of a substance regulators consider acceptable for lifelong human exposure, every RfDReference Dose — the US EPA's equivalent of a TDI, every NOAELNo Observed Adverse Effect Level — the highest dose at which a chemical produces no measurable effect in a standard animal test, every ADIAcceptable Daily Intake — the FAO/WHO terminology for food additives, equivalent to a TDI — is calculated from a dose-response curve drawn through three or four data points in a regulated animal test.

The architecture of those three-point tests is older than most chemical-safety claims you'll read on a label. A two-page paper called 'The 100-Fold Margin of Safety' in the Quarterly Bulletin of the Association of Food and Drug Officials by A J Lehmann and O G Fitzhugh — both staffers at what is now the FDA's Center for Food Safety and Applied Nutrition — set the framework out in 1954. Find the highest dose at which the test animal shows no measurable effect. Divide by 10 to account for the possibility that humans are more sensitive than the test species. Divide by 10 again to account for variation among humans — children, pregnant women, the elderly, the genetically unlucky. The 100-fold safety factor became the foundation of every TDI and RfD set by every food and chemical regulator in the world Lehmann & Fitzhugh 1954. The architecture has held for seventy years.

The framework was honest about its assumptions. Lehmann and Fitzhugh assumed the dose-response curve was monotonicA curve whose slope keeps the same sign across the tested range — either always increasing or always decreasing as dose increases. A monotonic curve never reverses direction. The opposite of monotonic, in this article, is non-monotonic. — a straight line, or at least a curve that goes one direction. They assumed there was a threshold below which no biological effect occurred. They assumed effects scale with dose in a predictable way. For the chemicals being evaluated in 1954 — pesticides, food preservatives, industrial solvents, food-additive synthetics — those assumptions held well enough to be useful. The chemicals on this article's list are different chemicals.

A regulatory dose-response study tests three or four doses of a chemical in a small group of animals, measures a panel of standard endpoints — body weight, organ weight, blood chemistry, reproductive output — and identifies the highest dose that doesn't produce a statistically significant effect. That dose becomes the NOAEL. The TDI is the NOAEL divided by an uncertainty factor, usually 100. Practically, this means the regulator's chart has three or four data points on it, drawn from animals exposed to doses chosen by the experimenter to span an effect range — often six or seven orders of magnitude higher than the dose any consumer would ever encounter.

In standard practice, the doses span four or more orders of magnitude. They're chosen to find the threshold, not to measure it precisely. The OECDOrganisation for Economic Co-operation and Development — sets internationally harmonised chemical-test guidelines that EU and US regulators adopt by reference test guidelines that govern this kind of work — TG 408 for 90-day rodent studies, TG 416 for two-generation reproduction studies — recommend three dose levels plus a control, with the highest selected to produce 'clear toxic effects' and the lowest set as a presumed-NOAEL. The effects you'd see between those points are extrapolated, not observed. JECFAJoint FAO/WHO Expert Committee on Food Additives — sets ADIs for food additives globally adopts this framework. So does EFSAEuropean Food Safety Authority — sets TDIs for food contact materials in the EU. So does the US EPA's reference-dose programme. The same three-point geometry, the same 100-fold uncertainty factor, the same architecture across the regulators that govern most of what reaches a consumer's mouth.

Three doses works for industrial poisons. They produce a curve where response scales monotonically with dose, the curve has a clear threshold, the threshold sits comfortably above environmental exposures, and the safety factor of 100 covers most of the residual uncertainty. The model has saved millions of lives by setting limits on substances like aflatoxin, dioxin, methylmercury, and lead. The trouble shows up when the chemical being tested isn't an industrial poison.

For chemicals that act on hormone receptors — endocrine disruptors — the dose-response curve frequently isn't monotonic at all. The biggest measurable effect can occur at the lowest tested dose, with the response taper or reverse as dose increases. Non-monotonic dose-responseA relationship in which the slope of the curve changes sign across the tested dose range — typically inverted-U or U-shaped. NMDR is most common in chemicals that act through hormone receptors, where receptor saturation, dimerisation kinetics, and selective cell apoptosis at high doses can each reverse the response direction. (NMDR) is the formal name. Note that NMDR is distinct from hormesisA specific NMDR sub-pattern in which low doses of a chemical produce a beneficial or stimulatory response, with adverse effects only at higher doses. The classic example is alcohol; the mechanism is generally adaptive cellular stress response. Hormesis is one cause of inverted-U dose-response, but NMDR is a broader category that includes adverse-at-low-dose curves that hormesis does not describe., which describes the specific case of low-dose stimulation followed by high-dose inhibition; NMDR is the broader category.

The mechanism explanation that became standard came out of the University of Missouri-Columbia, in a paper titled 'Large effects from small exposures.' Welshons, Thayer, Judy, Taylor, Curran, and vom Saal walked through the receptor biology in EHPEnvironmental Health Perspectives — peer-reviewed open-access journal published by the US National Institute of Environmental Health Sciences in 2003: at low doses, every hormone-mimicking molecule that binds the receptor produces an effect; at high doses, most of the receptors are already occupied, so additional molecules add nothing — the response saturates. Worse, at very high doses, hormone-mimicking molecules can trigger compensatory desensitisation: the cell downregulates receptor expression, the response shrinks, and additional molecules produce less effect than the saturating dose just before Welshons et al. 2003. The receptor-saturation argument is mechanistically uncontroversial; it is taught in every endocrinology textbook. What the field hadn't fully digested by 2003 was that this same machinery, applied to a synthetic chemical that mimics the hormone, predicts an inverted-U dose-response curve as the default — not the exception.

There are at least three other mechanisms that can reverse a hormone-active chemical's curve. Cell-selective apoptosis: high doses kill the cells that would otherwise produce the effect. Receptor cross-talk: at different doses, the molecule activates different receptors with opposing biological consequences. Mixed activation profiles: the same chemical agonises one receptor and antagonises another, with the balance shifting across the dose range. The published mechanistic literature catalogues each. The framework was honest about its assumption. The chemicals just don't share it.

In 1997, a team led by Frederick vom Saal at the University of Missouri-Columbia published the cleanest mammalian demonstration of non-monotonic dose-response in the modern toxicology literature. Pregnant CD-1 mice were given diethylstilbestrol — the canonical synthetic estrogen, prescribed to millions of pregnant women between 1940 and 1971 before being withdrawn for causing rare reproductive cancers — through their drinking water during the developmental window when fetal prostate organogenesis happens. The male offspring were grown to adulthood and their prostates dissected and weighed.

Four DES doses were tested, spanning a ten-thousand-fold range: 0.02 ng per gram of body weight per day, 0.2 ng/g/day, 2.0 ng/g/day, and 200 ng/g/day. The first three doses produced increased adult prostate weight; the fourth produced decreased adult prostate weight. The paper's verbatim sentence is the canonical statement of mammalian NMDR: 'Relative to controls, DES doses of 0.02, 0.2, and 2.0 ng per g of body weight per day increased adult prostate weight, whereas a 200-ng-per-g dose decreased adult prostate weight in male offspring' vom Saal et al. 1997. A parallel arm of the same study ran the same doses with 17β-estradiolThe body's own primary estrogen — chemically structured very similarly to the synthetic estrogens DES, BPA, and the major xenoestrogens that mimic it, the body's own primary estrogen, and saw the same pattern: prostate weight first increased, then decreased with dose. The inverted-U held for both compounds.

The result is exactly what Welshons' receptor-saturation mechanism predicted six years before the mechanism paper was published. The lowest doses, where receptor occupation was the rate-limiting step, produced the largest fractional effect. The highest dose — ten thousand times higher than the lowest — produced the opposite-sign effect, almost certainly through compensatory downregulation of estrogen receptor expression in the developing prostate. A regulator's three-point test, designed to find a threshold and bracket it with a hundred-fold safety factor, would have measured no effect at the lowest tested dose, a small effect at the middle dose, and no effect at the highest. The lowest dose was the worst dose. That was the data the regulator already had.

By the late 2000s the published evidence for low-dose effects and non-monotonic curves had grown into hundreds of studies. In 2012 a consortium led by Laura Vandenberg at the University of Massachusetts pulled the literature together into a single review in Endocrine Reviews — at 78 pages, the longest article the journal published that year and one of the most-cited papers in modern toxicology. They found, in the paper's own language, 'hundreds of examples from the cell culture, animal, and epidemiology literature' of low-dose effects and non-monotonic curves across BPA, DES, atrazine, dioxin, perchlorate, soy isoflavones, and dozens of other endocrine-active chemicals.

Vandenberg, Colborn, Hayes, Heindel, Jacobs, Lee, Shioda, Soto, vom Saal, Welshons, Zoeller, and Myers — a who's-who of the academic endocrine-disruption field — methodically catalogued the evidence in Vandenberg et al. 2012. The paper's central table listed 35 individual EDCs with documented low-dose effects in vivo. For each, it traced the experimental setup, the dose ranges, and the endpoints affected. Across the catalogue, two patterns recurred: effects appeared at doses below the regulatory NOAEL, and the dose-response shape was frequently non-monotonic — peaks and troughs in the curve that a three-point regulatory test could not have picked up. The conclusion: the existing chemical safety framework, with its three-dose architecture and 100× uncertainty factor, was systematically blind to the dose ranges where hormones produce their characteristic effects.

Linda Birnbaum, then the director of both the NIEHSNational Institute of Environmental Health Sciences — part of the US National Institutes of Health and the NTPNational Toxicology Program — the US federal interagency programme that conducts toxicity testing, the most senior toxicology figure in the US federal government, wrote the EHP editorial accompanying the Vandenberg review. Her position was unambiguous: the evidence was sufficient to justify revising the testing framework Birnbaum 2012. Six years later Hill, Myers, and Vandenberg followed with a more pointed paper showing that NMDR curves occur at doses relevant to regulatory decision-making — not just at exotically low doses, but at the doses where TDIs are actually being set Hill, Myers & Vandenberg 2018.

The geometry question — how many dose points it actually takes to detect a non-monotonic curve — got a quantitative answer in 2014, this time from Vandenberg as a sole author. Working through 250 published BPA experiments across 93 individual studies, she compared the experimental design of papers that detected NMDR against those that didn't. Experiments that detected non-monotonic responses examined an average of 6.9 doses. Experiments that failed to detect them examined 4.6. The difference was statistically significant (Student's t-test, p<0.05).

The analysis split each experiment into its design features — number of doses, number of replicates, choice of endpoint, dose range — and asked which features predicted whether NMDR would be visible in the result. Number of doses was the strongest single predictor. Of the 93 BPA studies analysed, 32 of 93 34.4% of BPA studies analysed in Vandenberg 2014 reported non-monotonic dose-response on at least one endpoint — 250 experiments total, 59 NMDRC + 191 monotonic reported NMDR on at least one endpoint. The cleanest analytic result was the dose-count comparison: 'Experiments reporting NMDRCs examined on average 6.9 doses, whereas experiments that failed to detect NMDRCs examined only 4.6 doses' Vandenberg 2014. The implication for the regulatory framework is direct. The three-dose-plus-control geometry that OECD TG 408 codifies sits below the threshold at which non-monotonic curves become visible. Three dose points draws a line; you need seven to draw a curve.

A separate 2015 EFSA-commissioned methodology paper by Lagarde, Beausoleil, Belcher, Belzunces, Emond, Guerbet, and Rousselle proposed a more conservative qualitative framework for evaluating NMDR claims. Their scoring system — five weighted criteria, with the minimum dose-level requirement set at 3 plus a control — was applied to 148 individual NMDR observations in the literature, and classified 82 as moderate-to-high plausibility Lagarde et al. 2015. Lagarde set the threshold for consideration of an NMDR claim. Vandenberg set the threshold for detection of one. The two papers don't conflict — they describe different parts of the same methodological problem, and both reach the same practical answer: standard regulatory testing is built on a dose-count geometry that systematically misses what hormone-active chemicals do.

Yes — for one chemical, after twenty years, by a factor of twenty thousand. On 19 April 2023 the European Food Safety Authority published a re-evaluation of the tolerable daily intake for bisphenol A in the EU. The new TDI is 0.2 nanograms per kilogram of body weight per day, calculated from a reference point of 8.2 ng/kg/day with an overall uncertainty factor of 50. The previous t-TDI, set in 2015, was 4 micrograms per kilogram per day. The new figure is twenty thousand times lower.

The triggering data was Th17 cells. EFSA identified an increase in the proportion of T-helper cells producing interleukin-17 — the immune-system signature of inflammatory and autoimmune disease — in mice exposed to BPA at doses well below the previous TDI. The agency's verbatim language: 'An effect on Th17 cells in mice was identified as the critical effect; these cells are pivotal in cellular immune mechanisms and involved in the development of inflammatory conditions, including autoimmunity and lung inflammation' EFSA 2023. The reference point was derived from the pivotal Th17 study; the agency added an uncertainty factor of 2 to the standard inter-species and intra-species 100-fold safety factor to cover the probability that other endpoints sit at lower doses. Total uncertainty factor 50, multiplied through, lands on 0.2 ng/kg/day. Twenty thousand times. The regulator's previous TDI had been the official safe dose for fifteen years.

The evidence base supporting the revision included the CLARITY-BPAConsortium Linking Academic and Regulatory Insights on BPA Toxicity — a joint NTP/FDA programme that ran a single rat study with both regulator-facing and academic-investigator analytical arms, designed to bridge the methodological dispute between FDA and the academic EDC field Core Study, a joint NTP/FDA project that dosed rats with five concentrations of BPA across two arms — a continuous-dose arm and a stop-dose arm where exposure ended at postnatal day 21. The study showed a statistically significant increase in mammary gland adenocarcinoma at the lowest tested dose, 2.5 µg/kg/day, in the stop-dose arm only: 22% incidence vs 6% in controls (p=0.016) NTP CLARITY-BPA Core Study 2018. The CLARITY findings were contested at the time of publication. The NTP writeup itself considered the mammary signal 'unlikely to be a plausible BPA-related lesion' given that the cancer appeared in only one dose group and one arm. The academic CLARITY investigators, in a synthesis led by Prins, Patisaul, Belcher, and Vandenberg, drew the opposite conclusion: that the developmental-window pattern — effect persists into adulthood after exposure stopped at PND21 — is exactly the signature predicted by EDC mechanisms, and that the lowest-dose-only pattern is consistent with the inverted-U dose-response that other BPA studies had reported Prins et al. 2019. Their summary: 'Collectively, the findings highlighted herein corroborate a significant body of evidence that documents adverse effects of BPA at doses relevant to human exposures and emphasizes the need for updated risk assessment analysis.' EFSA's 2023 revision sided with the academic interpretation.

Germany's Federal Institute for Risk Assessment — the BfRBundesinstitut für Risikobewertung — the German federal-state risk assessment body whose work the EU regularly defers to on chemical risk, the member-state body the EU usually defers to on chemical risk — published a competing opinion the same day. BfR's TDI was 0.2 µg/kg/day: a thousand times higher than EFSA's. The agencies were obliged under EU law to file a joint Article 30 report to the Commission identifying the contentious science. BfR and the European Medicines Agency, both disagreeing with EFSA, jointly stated 'insufficient evidence to conclude that the intermediate endpoint of increased Th17 cell percentage leads to an adverse immune outcome' BfR-EFSA Article 30 joint report 2023. The dispute is unresolved as of 2026. The 0.2 ng/kg figure is the operative EU TDI for food-contact materials, but member states have implemented enforcement at varying speeds and the underlying science remains in dispute. The interesting thing for this article isn't who's right. It's that two regulators looking at the same data set, the same week, derived TDIs that differed by three orders of magnitude. The framework's confidence intervals turn out to be narrower than its inter-agency disagreement.

As of April 2026, no major regulator has formally integrated non-monotonic dose-response analysis into the standard chemical-testing framework. EFSA's 2021 Scientific Committee Opinion concluded that the existing approach remains valid but acknowledged the need for international harmonised guidance. The US EPA's Endocrine Disruptor Screening Program does not include NMDR as a Tier 1 or Tier 2 decision criterion. OECD Guidance Document 150 — the international reference for endocrine-disruptor testing — organises tests by biological complexity but does not require NMDR analysis at any level. The position of every major regulator on NMDR, in summary: acknowledged in the discussion, not embedded in the standard test.

The current EFSA position lives in the Scientific Committee's October 2021 opinion on NMDR, which built on the agency's 2016 external report by Beausoleil and colleagues Beausoleil et al. 2016. The opinion's verbatim conclusion: 'the current risk assessment approach based on evaluating adverse outcomes seen in standard animal tests… remains valid' — and, significantly, the committee 'recommends a concerted international effort on developing internationally agreed guidance and harmonised frameworks for identifying and addressing NMDRs in the risk assessment process' EFSA 2021. The status quo holds. The acknowledgement that the status quo is missing something also holds. The two are simultaneously true.

The 2017 Solecki consensus statement, written by 21 international scientists at a BfR-led expert workshop, took a more conservative position: that the threshold-based safety-factor framework should be retained as the default and NMDR claims should require unusually strong evidence to justify regulatory departure Solecki et al. 2017. The academic side of the field, represented by Vandenberg, Hill, Myers, and the broader Endocrine Society, argues the opposite: that the threshold-based framework is the source of the systematic blindness in the first place, and that revising it is overdue. Six years on, the consensus statement is still cited by regulators and the academic side is still being cited by the chemicals being regulated.

The framework gap is real, slow to close, and unlikely to be resolved across the chemicals you encounter daily inside the next decade. The practical lever for consumers is the same as the editorial position underneath every chemical profile in this library: where a chemical is hormone-active and a non-hormone-active alternative exists at similar cost, switch to the alternative — regardless of where the regulator's TDI currently sits. A TDI in active inter-agency dispute, or that has been revised by 20,000× in twenty years, is not the kind of number a switching decision should depend on.

The corollary — that the TDI is wrong, when it's wrong, in only one direction — is also worth holding. Twenty thousand times lower, not twenty thousand times higher. Where regulators have moved on hormone-active chemicals, the move has consistently been toward more stringent limits, not less. Where regulators are still debating, the academic literature is generally already further along — closer in some cases to BfR's higher TDI than EFSA's, but more often closer to EFSA's tighter figure. The pattern across the library is consistent: regulators catch up. They catch up late, by an order of magnitude or two, after the academic data has accumulated past the point where the framework can ignore it. The case for switching now is the case for not waiting fifteen years for a number to move in the direction the literature was already pointing.

The kitchen scale is a fair analogy for industrial poisons. For the chemicals on this article's list, the scale measures one part of the picture. The other part shows up in mice at doses too small for the scale to register, in mammary glands at the lowest dose group of regulator-led studies, in the gap between two member-state agencies looking at the same data and arriving at TDIs three orders of magnitude apart.

The combination-effects article and this one cover two different ways the safety framework underestimates harm. The combination-effect failure is in how chemicals add up across exposures — the regulator tests them one at a time. The non-monotonic-dose-response failure is in how a single chemical's curve bends below the dose the regulator measured — the regulator picks three or four data points on a curve that needs seven. Together, the two arguments cover most of the conceptual gap between what the label says and what the chemistry does. The framework was honest about its assumptions in 1954. The chemicals on the bathroom counter are not the chemicals it was designed for.