EUTECNIQAtlas

Every page of EUTECNIQ on one mapArrow keys move through results · Enter opens · Esc closes

Reference library Agency guidance and databases Peer-reviewed methods One educational simulation

Toxicity and exposure models, and the one question each answers

Liver-injury scales, pharmacokinetic and population models, mechanistic liver simulators, structure-based hazard tools, dose–response software, an emissions model and fluid dynamics. For each: what it answers, what goes in and comes out, who maintains it, what regulators say about it and where it stops.

1,336drugs ranked in FDA's DILIrank 2.0 reference list; 217 of them in the most-DILI-concern group
≥ 1/10FDA's estimate of the rate of severe drug-induced liver injury relative to the rate of Hy's law cases
5OECD principles a (Q)SAR model should meet for regulatory use, agreed in November 2004
553chemicals with human in vitro toxicokinetic data in US EPA's httk R package at publication (2017)

FDA, DILIrank dataset page (modified April 2, 2026); FDA, Drug-Induced Liver Injury: Premarketing Clinical Evaluation (July 2009); OECD (Q)SAR validation principles (2004); Pearce et al., J Stat Softw 2017;79(4).

Status. Education, not medical advice. EUTECNIQ runs none of these models. The population simulation on this page is an educational calculation from published parameters, not a measurement or a prediction for any product.

Related: ADME, PK and PBPK · Academy: toxicology and dose · Toxicology and exposure assessment · Toxicology and risk assessment

Overview

Each model answers one kind of question, and none of them alone decides that a product is safe

A causality scale judges one patient's injury. A pharmacokinetic model predicts concentrations. A hazard tool flags a structure. Safety conclusions combine several of these with exposure and the actual studies; the course on toxicology and dose shows how.

Seventeen models and tools against eight questionsQualitative map
Matrix of 17 toxicity and exposure models against eight questions; filled marks show the question each model was built to answer, open marks show where it contributes an input or a secondary outputLiver injury:cause orsignalExposureover timeVariabilitybetweenpeopleMechanismof toxicityHazardwithout newanimal dataDose for aset responseImpact perkg emittedFlow andmixing inorgansLiverTox likelihood scaleLiverTox likelihood scale · Liver injury: cause or signal: highRUCAM causality scoreRUCAM causality score · Liver injury: cause or signal: highFDA Hy's lawFDA Hy's law · Liver injury: cause or signal: highFDA DILIrankFDA DILIrank · Liver injury: cause or signal: highCompartmental PK, PK/PDCompartmental PK, PK/PD · Exposure over time: highPopulation PK (NLME)Population PK (NLME) · Exposure over time: highPopulation PK (NLME) · Variability between people: highPBPK platformsPBPK platforms · Exposure over time: highPBPK platforms · Variability between people: lowUS EPA httkUS EPA httk · Exposure over time: highUS EPA httk · Variability between people: lowUS EPA httk · Hazard without new animal data: lowDILIsym (QST)DILIsym (QST) · Liver injury: cause or signal: lowDILIsym (QST) · Variability between people: lowDILIsym (QST) · Mechanism of toxicity: high(Q)SAR models(Q)SAR models · Hazard without new animal data: highRead-acrossRead-across · Hazard without new animal data: highTTC with Cramer classesTTC with Cramer classes · Hazard without new animal data: highToxCast and Tox21ToxCast and Tox21 · Mechanism of toxicity: lowToxCast and Tox21 · Hazard without new animal data: highAdverse outcome pathwaysAdverse outcome pathways · Mechanism of toxicity: highAdverse outcome pathways · Hazard without new animal data: lowBenchmark dose (BMDS)Benchmark dose (BMDS) · Dose for a set response: highUSEtoxUSEtox · Impact per kg emitted: highCFD of gut and airwaysCFD of gut and airways · Flow and mixing in organs: highbuilt to answer this question (per its documentation or regulator)contributes an input or secondary outputList of 17 toxicity and exposure models with the question each was built to answer and the questions it contributes toLiverTox likelihood scaleAnswers: liver-injury cause or signalRUCAM causality scoreAnswers: liver-injury cause or signalFDA Hy's lawAnswers: liver-injury cause or signalFDA DILIrankAnswers: liver-injury cause or signalCompartmental PK, PK/PDAnswers: exposure over timePopulation PK (NLME)Answers: exposure over time; variability betweenpeoplePBPK platformsAnswers: exposure over timeContributes: variability between peopleUS EPA httkAnswers: exposure over timeContributes: variability between people; hazard withoutnew animal dataDILIsym (QST)Answers: mechanism of toxicityContributes: liver-injury cause or signal; variabilitybetween people(Q)SAR modelsAnswers: hazard without new animal dataRead-acrossAnswers: hazard without new animal dataTTC with Cramer classesAnswers: hazard without new animal dataToxCast and Tox21Answers: hazard without new animal dataContributes: mechanism of toxicityAdverse outcome pathwaysAnswers: mechanism of toxicityContributes: hazard without new animal dataBenchmark dose (BMDS)Answers: dose for a set responseUSEtoxAnswers: impact per kg emittedCFD of gut and airwaysAnswers: flow and mixing in organs
Filled marks: the purpose each tool's maintainer, its documentation or the cited regulator states (sources in the sections below). Open marks: the tool supplies an input or a secondary output for that question. A judgment by this page, not a ranking of quality.
Text version: the eight questions
  • Liver-injury cause or signal: LiverTox likelihood scale, RUCAM, FDA Hy's law, FDA DILIrank; DILIsym contributes simulated enzyme and bilirubin rises.
  • Exposure over time: compartmental PK and PK/PD, population PK, PBPK platforms, US EPA httk.
  • Variability between people: population PK; PBPK platforms, httk and DILIsym contribute through simulated populations.
  • Mechanism of toxicity: DILIsym, adverse outcome pathways; ToxCast and Tox21 contribute target-level activity.
  • Hazard without new animal data: (Q)SAR, read-across, threshold of toxicological concern, ToxCast and Tox21; httk and adverse outcome pathways contribute.
  • Dose for a set response: benchmark-dose modeling (US EPA BMDS, EFSA guidance).
  • Impact per kg emitted: USEtox, in life-cycle assessment.
  • Flow and mixing in organs: computational fluid dynamics of the stomach and airways.
Knowledge bases and causality

Whether a product caused a liver injury is judged with expert scales, not with a laboratory test

Drug-induced liver injury is a diagnosis of exclusion. LiverTox, published by the U.S. National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) on the NLM Bookshelf, covers prescription and nonprescription drugs and selected herbal and dietary supplements, and grades each agent on a likelihood scale.

LiverTox likelihood categories for drug-induced liver injury
CategoryMeaningEvidence behind it
AWell known cause of liver injuryCharacteristic signature; more than 50 reported cases, including case series
BKnown or highly likely cause of idiosyncratic injuryCharacteristic signature; 12 to 50 cases, including small case series
CProbable causeUncommon reports; no characteristic signature; fewer than 12 cases without significant case series
DPossible, rare causeSingle case reports; fewer than 3 cases in the literature
ENot believed or unlikely to cause injuryNo evidence despite extensive use; any single case reports largely unconvincing
E*Suspected, unprovenNo convincing cases in the literature
XUnknownInadequate information on risk, for an agent recently introduced or rarely used

LiverTox, Categorization of the Likelihood of Drug Induced Liver Injury, updated May 4, 2019 (NBK548392). Agents that injure the liver only at high doses, typically overdoses, carry the suffix [HD] after category A, B, C or D; LiverTox's examples are aspirin, acetaminophen, niacin and vitamin A. LiverTox grades green tea “A (well established cause of clinically apparent liver injury)” (entry updated November 20, 2020); see Green tea and the course section on botanical liver injury.

Causality in one patient: DILIN expert opinion and RUCAM

In the U.S. Drug-Induced Liver Injury Network (DILIN), three experts grade each case independently on five levels tied to a likelihood: definite (above 95%), very likely (75–95%), probable (50–75%), possible (25–50%) and unlikely (below 25%). Disagreements are reconciled in a meeting.

RUCAM, the Roussel Uclaf Causality Assessment Method, was introduced in 1993 by Danan and Benichou. It scores each suspect agent on seven items: time to onset, course after stopping, risk factors, concomitant drugs, non-drug causes, prior information on the agent's hepatotoxicity and response to re-exposure. Danan and Teschke published an update in 2016 for drug- and herb-induced liver injury.

RUCAM injury pattern from the R ratio and causality categories from the total score
StepRule
Injury patternR = (ALT ÷ ALT upper limit of normal) ÷ (alkaline phosphatase ÷ its upper limit of normal), from the first abnormal test. R above 5: hepatocellular; 2 to 5: mixed (scored as cholestatic); below 2: cholestatic
Score range−9 to +14 in theory; −7 to +11 without re-exposure
Causality≤ 0 excluded · 1–2 unlikely · 3–5 possible · 6–8 probable · above 8 highly probable (LiverTox's RUCAM chapter; its Causality chapter gives 3 or below as unlikely and 4–5 as possible)

LiverTox, Causality, updated November 20, 2019 (NBK548049); LiverTox, Roussel Uclaf Causality Assessment Method (RUCAM) in Drug Induced Liver Injury, updated May 4, 2019 (NBK548272); Danan G, Teschke R. Int J Mol Sci 2016;17(1):14.

Limits stated by LiverTox: some RUCAM descriptions are ambiguous and are interpreted differently by different users; alcohol use is not defined, and what counts as a compatible time to onset for a concomitant drug is unclear; ALT upper limits vary by laboratory, age and sex; for cholestatic and mixed injury the score range is narrower, so often only “probable” can be assigned; and RUCAM gives no instructions for the R ratio in pre-existing liver disease.

Signals across a trial or a drug list: Hy's law and DILIrank

FDA's 2009 guidance defines a Hy's law case by three findings: ALT or AST at least 3 times the upper limit of normal, more often than with placebo or a non-hepatotoxic control; total bilirubin above 2 times the upper limit of normal in some of those subjects, without initial signs of cholestasis (raised alkaline phosphatase); and no other explanation, such as viral hepatitis, pre-existing liver disease or another hepatotoxic drug.

FDA states that such cases identify a drug likely to cause severe injury, fatal or requiring transplant, at a rate of at least one tenth the rate of Hy's law cases. One case in a clinical database is worrisome; two are considered highly predictive. Finding none does not show that a drug with aminotransferase rises is free of risk.

FDA's DILIrank 2.0 places 217 of 1,336 drugs in its most-concern groupFDA dataset
Bar chart of FDA DILIrank 2.0: 217 drugs most-DILI-concern, 351 less-DILI-concern, 414 no-DILI-concern and 354 ambiguous-DILI-concern, 1,336 in totalvMost-DILI-concernvMost-DILI-concern: 217 drugs217vLess-DILI-concernvLess-DILI-concern: 351 drugs351vNo-DILI-concernvNo-DILI-concern: 414 drugs414Ambiguous-DILI-concernAmbiguous-DILI-concern: 354 drugs354Number of drugs, total 1,336 (DILIrank 2.0)Bar chart of FDA DILIrank 2.0: 217 drugs most-DILI-concern, 351 less-DILI-concern, 414 no-DILI-concern and 354 ambiguous-DILI-concern, 1,336 in totalvMost-DILI-concernvMost-DILI-concern: 217 drugs217vLess-DILI-concernvLess-DILI-concern: 351 drugs351vNo-DILI-concernvNo-DILI-concern: 414 drugs414Ambiguous-DILI-concernAmbiguous-DILI-concern: 354 drugs354Number of drugs, total 1,336 (DILIrank 2.0)
FDA National Center for Toxicological Research (NCTR), Liver Toxicity Knowledge Base, Drug-Induced Liver Injury Rank (DILIrank 2.0) dataset page, modified April 2, 2026. Version 1.0: Chen M, et al. Drug Discov Today 2016;21:648–653.
Show data: DILIrank 2.0 categories
Number of drugs in each DILIrank 2.0 category
CategoryDrugsCausal relationship with liver injury
vMost-DILI-concern217Established
vLess-DILI-concern351Established
vNo-DILI-concern414Established
Ambiguous-DILI-concern354Undetermined
Total1,3361,036 in version 1.0 plus 300 drugs approved 2010–2021

FDA DILIrank dataset page. Classes come from the hepatotoxicity descriptions in FDA-approved labeling combined with causality evidence from the literature. The list covers FDA-approved drugs only.

Pharmacokinetic and physiological models

Compartmental, PBPK and toxicokinetic models predict concentrations; they do not predict harm by themselves

The one-compartment equations, the Hill model and the organ-and-blood-flow structure of a PBPK model are on ADME, PK and PBPK. This section covers how regulators treat PBPK, the main platforms, and a toxicokinetic package developed at US EPA for screening chemicals.

FDA and EMA positions on PBPK analyses
Regulator and documentStatusWhat it says about acceptance
FDA, Physiologically Based Pharmacokinetic Analyses — Format and ContentFinal guidance, CDER; the PDF cover is dated August 2018 (page footer 09/03/18), and FDA's guidance page lists the issue date as September 2018PBPK predictions can inform “whether, when, and how to conduct certain clinical pharmacology studies” and support labeling dose recommendations. Accepting them in place of clinical data is decided “on a case-by-case basis.” The guidance “does not address methodological considerations and best practices.”
EMA, Guideline on the reporting of PBPK modelling and simulation (EMA/CHMP/458101/2016)Final, adopted December 13, 2018; in effect July 1, 2019The platform must be qualified for its intended use, and “qualification is purpose and platform version specific.” Requirements rise with regulatory impact (high, moderate, low): using a model to waive an interaction study, or to set a child's dose with little or no clinical exposure data, is high impact. Applies to commercial and in-house platforms.

FDA guidance PDF and FDA guidance page (docket FDA-2016-D-3969); EMA guideline PDF, sections on qualification and regulatory impact. Status as of October 2026.

Three PBPK platforms: maintainer, description and license
PlatformMaintainerWhat the peer-reviewed description saysLicense
Simcyp SimulatorCertara (Sheffield, UK)A population-based “bottom-up” simulator combining in vitro enzyme and transporter data and physicochemistry with demographic, physiological and genetic data on healthy and disease populations; built for multi-drug interactions, metabolite profiles and time-dependent effects such as auto-inductionCommercial
GastroPlusSimulations Plus (Lancaster, California)Absorption by the advanced compartmental absorption and transit (ACAT) model, nine compartments for gut segments, linked to a PBPK model of fourteen tissue compartments. Agoram and colleagues describe ACAT as differential equations for release, dissolution, passive and carrier-mediated absorption, saturable metabolism and efflux, predicting fraction absorbed, bioavailability and plasma profilesCommercial
PK-Sim and MoBi (Open Systems Pharmacology suite)Open Systems Pharmacology community, developed on GitHubFormerly commercial; released as open source from version 7.0 in early 2017. Covers PBPK across preclinical species and humans, drug interactions and special populations, with published qualification repositoriesOpen source, GPLv2

Jamei M, et al. Expert Opin Drug Metab Toxicol 2009;5:211–223 (Simcyp); Agoram B, Woltosz WS, Bolger MB. Adv Drug Deliv Rev 2001;50 Suppl 1:S41–S67 (ACAT); Lippert J, et al. CPT Pharmacometrics Syst Pharmacol 2019;8:878–882 (Open Systems Pharmacology). Maintainers as named in the methods of Choules MP, et al. J Clin Pharmacol 2025;65:1096–1105 (Simcyp, Certara) and Wang L, et al. Clin Transl Sci 2026;19:e70715 (GastroPlus version 9.9, Simulations Plus; ACAT and tissue compartments).

Physiologically based toxicokinetics for thousands of chemicals: US EPA httk

The httk authors note that “pharmacokinetic” and “toxicokinetic” are used largely interchangeably, and chose “toxicokinetic” because the package gives dose context to toxicity screening. The httk R package, written by scientists at US EPA and the National Toxicology Program (NIEHS), pairs generic models with two in vitro measurements per chemical: intrinsic hepatic clearance in primary hepatocytes (Clint) and the fraction unbound in plasma (fub). Run in reverse (“reverse dosimetry”), it estimates the exposure dose that would produce a concentration found active in high-throughput screens.

US EPA httk package: models, inputs, outputs, scope and limits
ItemDetail
Models (2017 paper)pbtk (physiologically based), 3compartment, 3compartmentss (steady state) and 1compartment
Steady-state equation (3compartmentss)Css = kdose / [fub·Qgfr + (Qliver + Qgut)·fub·Clmetabolism / ((Qliver + Qgut) + fub·Clmetabolism/Rblood2plasma)]: renal filtration plus well-stirred liver clearance, for a constant dose rate kdose (mg/kg body weight per day)
Coverage at publication553 chemicals with human in vitro data, including 94 pharmaceuticals and 415 ToxCast chemicals; rat in vitro data for 67
AssumptionsPerfusion-limited tissues; fraction absorbed set to 1 when no measured value exists; renal elimination by passive glomerular filtration
Limits stated by the authorsFull kinetics may need data not available from high-throughput in vitro work: bioavailability, transporters, protein-binding kinetics, and extra-hepatic or strongly saturable metabolism
Current releaseVersion 2.7.4 on CRAN (December 2025), MIT license; includes a Monte Carlo sampler for human variability and empirically calibrated methods for tissue partition coefficients

Pearce RG, Setzer RW, Strope CL, Sipes NS, Wambaugh JF. httk: R Package for High-Throughput Toxicokinetics. J Stat Softw 2017;79(4):1–26 (sections 2.1–2.3); CRAN package description, httk 2.7.4, published December 9, 2025. Author affiliations: U.S. Environmental Protection Agency and the National Toxicology Program, NIEHS.

What a one-compartment model predicts: raising ka moves the peak earlier and higher without changing AUC; F scales the whole curve; CL lowers it and shortens the half-lifeComputed from the stated equations · illustrative inputs
Left: plasma concentration after one oral dose, annotated with Cmax, tmax, the half-life and the area under the curve. Right: three rows of four curves each, raising the absorption rate constant ka, the bioavailability F and the clearance CL in turn, with the computed change in Cmax, tmax and AUC. Inputs are illustrative and describe no compound.
Equations from ADME, PK and PBPK on this site (Grogan and Preuss, StatPearls 2023; Mould and Upton, CPT Pharmacometrics Syst Pharmacol 2013). Inputs are invented for illustration and describe no botanical compound. Every number on the figure is computed from those inputs. Drawn by EUTECNIQ. Full-size SVG
Population PK and mixed-effects models

Population models split variability into a typical value, differences between people and measurement error

A nonlinear mixed-effects model fits everyone's data at once. Fixed effects (θ) give the typical parameter values; random effects (η) give each person's deviation; a residual term (ε) absorbs assay noise and model misfit. FDA's 2022 guidance notes that population PK analysis is frequently used to guide drug development and to inform individualized dosing.

Equations of a population pharmacokinetic model
LevelEquationMeaning
Individual parameterCLi = θ1 · exp(η1i)θ1 is the population clearance; η1i is subject i's deviation, normally distributed with mean 0 and variance ω², so CLi is log-normal
Between-subject CVCV (%) = √(exp(ω²) − 1) · 100%Converts the log-scale variance ω² to a coefficient of variation
Additive residual errorY = f(θ, time) + εConstant noise; often enough for pharmacodynamic data
ProportionalY = f(θ, time) · (1 + ε)Noise scales with the prediction
ExponentialY = f(θ, time) · exp(ε)Tends to overweight low concentrations, like the proportional form
CombinedY = f(θ, time) · (1 + ε1) + ε2Often used for dense PK data because it reflects assay variability

Mould DR, Upton RN. CPT Pharmacometrics Syst Pharmacol 2012;1:e6 (fixed and random effects, two levels of unexplained variability); 2013;2:e38, equations 5–6 and Table 1; U.S. FDA, Population Pharmacokinetics, Guidance for Industry, February 2022 (final; CDER and CBER).

Estimation methods and software for nonlinear mixed-effects models
SoftwareMaintainer and licenseEstimation methods
NONMEMICON (Dublin); commercial. The first population PK software, developed in San Francisco in 1978; requires a Fortran compilerFirst-order conditional estimation (FOCE; with interaction, FOCEI, in the caffeine study below), Laplace, iterative two-stage, importance sampling, SAEM and Bayesian Markov-chain Monte Carlo, as listed by ICON
MonolixLixoft (Paris) in 2019; current documentation published by Simulations Plus; commercialStochastic approximation expectation-maximization (SAEM)
nlmixr2Open-source R package (GPL ≥ 3); version 7.0.1 on CRAN, August 2026SAEM, FOCE and FOCEI, nlme; differential equations solved by the rxode2 package

Mould and Upton 2013: FOCE and Laplace approximate the likelihood with a simpler function, SAEM refines estimates stochastically, and only the original FO method is “of concern” because it can bias random-effect estimates. Fidler M, et al. CPT Pharmacometrics Syst Pharmacol 2019;8:621–633 (software, licenses and nlmixr estimation methods); Mould DR, Upton RN 2012 (NONMEM was the first population PK software and requires Fortran); ICON, NONMEM product page (estimation methods; written in ANSI Fortran 95), retrieved October 2026; Schoemaker R, et al. CPT Pharmacometrics Syst Pharmacol 2019;8:923–930 (nlmixr SAEM and FOCEI matched Monolix/SAEM and NONMEM/FOCEI closely); CRAN, nlmixr2 and nlmixr2est package descriptions, retrieved October 2026.

A worked simulation: caffeine in 1,000 virtual adults

Csajka and colleagues at the University of California, San Francisco, fitted caffeine data from 24 healthy adults (8 given an herbal product containing caffeine and ephedrine, 16 given pharmaceutical caffeine, ephedrine or both) with a one-compartment model in NONMEM (FOCE with interaction). The simulation below draws 1,000 virtual adults from their published estimates.

With the same 175 mg dose, the simulated peak ranges from about 2.9 to 5.4 mg/L across the middle 90% of 1,000 virtual adultsEducational calculation
Simulated caffeine plasma concentration over 24 hours after a single 175 mg oral dose in 1,000 virtual adults: 25 individual curves, the median and the 5th to 95th percentile band, from the published population model of Csajka et al. 200501234567Caffeine, mg/L04812162024Time after dose, h5th to 95th percentile of 1,000 simulated adultsmedian of 1,000 simulated adults5th–95th percentile,1,000 virtual adults25 of the simulatedindividualsmedianDose 175 mg, single oralCL/F 4.98 L/h (CV 38%)V/F 38.6 L (CV 20%)ka​ 3.84 /h (CV 50%)Lag 22.2 minSimulated caffeine plasma concentration over 24 hours after a single 175 mg oral dose in 1,000 virtual adults: 25 individual curves, the median and the 5th to 95th percentile band, from the published population model of Csajka et al. 200501234567Caffeine, mg/L04812162024Time after dose, h5th to 95th percentile of 1,000 simulated adultsmedian of 1,000 simulated adults5th–95th percentile, 1,000 virtual adults25 of the simulated individualsmedian175 mg; CL/F 4.98 L/h, V/F 38.6 L, ka​ 3.84 /h
Calculated from a one-compartment oral model with lag time, C(t) = (D/V)·ka/(ka − k)·(e−k(t−tlag) − e−ka(t−tlag)), k = CL/V, with parameters from Csajka C, Haller CA, Benowitz NL, Verotta D. Br J Clin Pharmacol 2005;59:335–345, Table 3 (non-users of oral contraceptives), and that paper's log-normal between-subject variability (CV converted to ω with Mould and Upton 2013, equation 6); an educational calculation, not a measurement or a prediction for any product. Python, fixed random seed 20261030.

The median curve is the typical person; the band is the spread that clearance, volume and absorption differences alone produce, before any measurement error.

Show data, method and limits
Simulated caffeine concentrations in mg/L at selected times: 5th percentile, median, 95th percentile and typical individual
Time after dose, h5th percentileMedian95th percentileTypical individual (η = 0)
0.50.861.773.391.77
12.543.795.263.91
22.783.714.883.79
41.972.883.862.94
80.701.732.721.75
120.251.041.991.05
240.010.230.910.22

Simulated peak concentration: 5th percentile 2.93, median 3.98, 95th percentile 5.35 mg/L; time of peak 0.8, 1.3 and 2.0 h. Exposure AUC0–∞ = D/(CL/F): 18.6, 35.2 and 65.6 mg·h/L. Typical half-life ln 2·V/CL = 5.37 h, against 5.3 h in the authors' Results (5.4 h in their Discussion).

Published parameters used
CL/F 0.083 L/min (4.98 L/h), between-subject CV 38%; V/F 38.6 L, CV 20%; ka 0.064 /min (3.84 /h), CV 50%; lag time 22.2 min, estimated for the herbal product only (no lag with pharmaceutical caffeine), with no between-subject variability reported. The authors report CL/F falling to 0.038 L/min in oral-contraceptive users; that subgroup is not simulated.
Dose
175 mg, the measured caffeine content of the herbal-product dose in the study (labeled 200 mg). The model is linear, so concentrations scale with dose.
Variability model
Csajka and colleagues model each caffeine parameter exponentially, Pj = θ·eηj, with independent, normally distributed η of mean zero (their “Variance models” section), so the log-normal form and the independent draws follow the paper. The paper reports the variability only as CV%; converting CV to ω with ω² = ln(1 + CV²), from equation 6 of Mould and Upton (2013), is this page's assumption: ω = 0.367 (CL), 0.198 (V), 0.472 (ka).
Not included
Residual error (the authors used a proportional error model; Table 3 does not report its value for caffeine), pre-dose caffeine, oral-contraceptive use, smoking and any other covariate. Because CL and V are apparent values (divided by bioavailability F), F is not needed.
What it does not establish
It describes neither any product nor any safe or effective amount of caffeine. It reproduces the spread implied by one published model fitted to 24 people.

Csajka et al. 2005 (study 1: herbal product; study 2: pharmaceutical caffeine, alone or with ephedrine); Mould and Upton 2013. See Caffeine and Ephedra.

Quantitative systems toxicology

DILIsym links liver exposure to three injury mechanisms and simulates enzyme rises across virtual populations

DILIsym is a mathematical model of drug-induced liver injury, developed over about a decade through a public–private partnership. It combines a compound's measured potency for oxidative stress, mitochondrial dysfunction and disruption of bile-acid handling with an estimate of liver exposure, for example from a PBPK model.

DILIsym inputs, outputs, maintainer, regulatory use and limits
ItemAs described in peer-reviewed papers
InputsIn vitro assays in human systems: reactive oxygen species production, cellular oxygen consumption (mitochondrial respiration), and inhibition constants for bile-acid transporters such as BSEP, NTCP and MRP3/4; plus PBPK estimates of liver exposure to parent compound and metabolites
OutputsSimulated serum ALT and bilirubin, the fraction of a simulated population (SimPops) above 3 times the upper limit of normal, simulated Hy's law cases, and which mechanism dominates
Published examplesTolcapone at 200 mg three times daily for one week: ALT above 3 times the upper limit in 2.2% of a simulated population, against none for entacapone, attributed mainly to mitochondrial uncoupling potency and liver exposure. Fezolinetant (version 7A): no ALT above 3 times the upper limit in simulated healthy volunteers at therapeutic doses; one simulated Hy's law case at 45 and 60 mg once daily in a population with fatty liver disease and weaker mitochondria
MaintainerCommercial. Authors' affiliation: DILIsym Services in 2016; Simulations Plus (Research Triangle Park, North Carolina) in 2026
Regulatory useIts developers report that results are increasingly included in regulatory communications and new drug application submissions. No FDA or EMA document stating acceptance criteria for it was found for this page
Limits stated in the papersOnly mechanisms with in vitro data can be represented (tolcapone: respiration and bile-acid transport only, so oxidative stress was not assessed); no virtual population matched the fezolinetant trial population, and model risk features could not be tied to age, sex or weight

Watkins PB. Methods Mol Biol 2022;2486:181–196; Longo DM, et al. CPT Pharmacometrics Syst Pharmacol 2016;5:31–39; Nielsen JC, et al. Clin Pharmacol Ther 2026;119:1016–1024; Shoda LKM, et al. Biopharm Drug Dispos 2014;35:33–49. All are co-authored by the model's developers, which readers should weigh.

Hazard without new animal data

Structure, analogues and in vitro screens can flag a hazard, each only inside its stated domain

These methods estimate a property of a chemical from its structure, from tested look-alikes, or from cell and protein assays. Regulators accept them for defined purposes and ask that the domain and the reasoning be documented.

(Q)SAR models and the OECD QSAR Toolbox

A (quantitative) structure–activity relationship predicts a property from molecular structure. In November 2004 OECD member countries agreed that, for regulatory consideration, a (Q)SAR model should have:

  1. a defined endpoint;
  2. an unambiguous algorithm;
  3. a defined domain of applicability;
  4. appropriate measures of goodness-of-fit, robustness and predictivity;
  5. a mechanistic interpretation, if possible.

ICH M7(R2) applies this to mutagenic impurities in medicines: two complementary (Q)SAR methods, one expert rule-based and one statistical, both following the OECD principles. If neither finds a structural alert, the impurity is considered of no mutagenic concern. The OECD QSAR Toolbox is free software, with OECD and ECHA named as copyright holders, that retrieves experimental data, simulates metabolism and groups chemicals for read-across.

OECD Principles for the Validation, for Regulatory Purposes, of (Q)SAR Models (37th Joint Meeting, November 2004), with notes that (Q)SARs are reductionist, that unpublished commercial algorithms hinder independent assessment, and that lack of a mechanistic interpretation does not by itself rule out regulatory use; ICH M7(R2), Step 4, April 3, 2023, section 6; OECD QSAR Toolbox website (version 4.9 available), retrieved October 2026.

Read-across, as ECHA assesses it

Read-across predicts a property of a target substance from tested source substances. ECHA's Read-Across Assessment Framework (RAAF, March 2017) separates an analogue approach (one or a few sources, no trend) from a category approach (a group with a regular pattern or a worst case). Its six scenarios rest on two hypotheses: the substances are transformed to the same compound, or different compounds cause the same type of effect.

Each assessment element, from structural similarity and exposure of the biological target to impurities and bias that influences the prediction, is graded from “acceptable with high confidence” to “not acceptable.” ECHA calls the RAAF a framework for expert judgment, “rather than a series of steps to be followed mechanically,” and states that it is not legal advice.

ECHA, Read-Across Assessment Framework (RAAF), ECHA-17-R-01-EN, March 2017.

Threshold of toxicological concern

When a structure is known, toxicity data are limited and intake can be estimated, EFSA's 2019 guidance, built on a joint EFSA and WHO review, sets screening values: 30, 9.0 and 1.5 µg/kg body weight per day for Cramer Classes I, II and III; 0.3 for organophosphates and carbamates; 0.0025 for potential DNA-reactive mutagens or carcinogens. The full table and exclusions are on Margin of exposure and TTC and in the course section When there are no data.

EFSA Scientific Committee. Guidance on the use of the Threshold of Toxicological Concern approach in food safety assessment. EFSA J 2019;17(6):5708, abstract, section 1 and Table 2.

High-throughput screening and adverse outcome pathways

ToxCast, Tox21 and adverse outcome pathways: what they are, who runs them and their stated uses
ProgramWhat it isWho runs itStated use
ToxCastMedium- and high-throughput in vitro assays of how chemicals act on biological targets, covering almost 10,000 substances from more than 20 assay sources; curve-fitted with open-source R packages (tcpl, tcplfit2) and public through the CompTox Chemicals DashboardUS EPARanking and prioritizing chemicals for further evaluation; EPA's Endocrine Disruptor Screening Program is working to use it to identify priority chemicals
Tox21A federal collaboration to develop faster toxicity screening methods; ToxCast aggregates its dataUS EPA, NCATS, NIEHS and FDAScreening chemicals for potential to disrupt processes in the human body
Adverse outcome pathway (AOP)A documented chain from a molecular initiating event, through measurable key events, to an adverse outcome of regulatory significance. An AOP should not be specific to one stressorOECD program; AOPs are developed in the AOP-Wiki, part of the AOP Knowledge BaseOrganizing mechanistic evidence; endorsed AOPs are published in the OECD Series on Adverse Outcome Pathways

US EPA, Toxicity Forecasting (ToxCast) page, updated August 27, 2026; US EPA, Toxicology Testing in the 21st Century (Tox21) page, updated April 22, 2026; Thomas RS, et al. Toxicol Sci 2019;169:317–332 (goal of broader acceptance of these methods for higher-tier regulatory decisions); OECD AOP Developers' Handbook version 2.5, July 2022; Ankley GT, et al. Environ Toxicol Chem 2010;29:730–741 (the AOP concept).

Three ways to flag a hazard without new animal data, each valid only inside its own domainSchematic · content from this page
Hub diagram. Center: hazard estimate without new animal data. Three methods: (Q)SAR, read-across and the threshold of toxicological concern, each with the rule or values the page cites.
Content from this section: OECD (Q)SAR validation principles (2004); ICH M7(R2) section 6; ECHA RAAF (2017); EFSA TTC guidance (2019). Drawn by EUTECNIQ. Full-size SVG
Dose–response modeling

Benchmark-dose software fits the whole dose–response curve; the method is taught on two other pages

How a benchmark dose and its lower bound (BMDL) are derived, and how they compare with a NOAEL, are explained with calculators in the course section NOAEL or benchmark dose and in Dose–response and the margin. The definitions are on Dose–response vocabulary.

US EPA BMDS and EFSA benchmark dose guidance
Tool or documentMaintainer and statusWhat it states
Benchmark Dose Software (BMDS)US EPA; BMDS Online, BMDS Desktop and the pybmds Python package (version 25.2 highlighted, page updated September 21, 2026)Models for dichotomous, continuous and nested data; “used by EPA risk assessors primarily for chemical risk assessment”
EFSA, Guidance on the use of the benchmark dose approach in risk assessmentEFSA Scientific Committee; Final, EFSA J 2022;20(10):7584The BMD approach is “scientifically more advanced” than the NOAEL approach; recommends a Bayesian paradigm with model averaging as the preferred method; the BMDL is the potential reference point and the BMDU/BMDL ratio expresses uncertainty

US EPA, Benchmark Dose Software page; EFSA Scientific Committee 2022, abstract.

Life-cycle assessment

USEtox compares the toxic impact of chemical emissions in life-cycle assessment; it does not assess the safety of swallowing a supplement

USEtox is the scientific consensus model developed under the UNEP–SETAC Life Cycle Initiative to characterize the human toxicity and freshwater ecotoxicity of chemical emissions in life-cycle impact assessment. It answers how much toxic impact one kilogram released to air, water or soil causes, relative to other chemicals. Life-cycle assessment itself is covered on Sustainability and LCA.

An emission's characterization factor is the product of fate, exposure and effectSchematic, not to scale
USEtox cause-effect chain: an emission passes through a fate factor, an exposure factor and an effect factor; their product is the characterization factor. Schematic, not to scaleEmissionkg released to air,water or soilFate factor FFhow the chemicalspreads and degradesExposure factor XFhow much reachespeople or ecosystemsEffect factor EFharm per unit ofexposure; for people,from ED50​CharacterizationfactorCF = EF · XF · FFIntake fraction iF = XF · FF. Human toxicity: CF in CTUh, disease cases per kg emitted.Freshwater ecotoxicity: CF in CTUe, potentially affected fraction of species integrated over time and volume per kg emitted.USEtox cause-effect chain: an emission passes through a fate factor, an exposure factor and an effect factor; their product is the characterization factor. Schematic, not to scaleEmissionkg released to air, water or soilFate factor FFhow the chemical spreads and degradesExposure factor XFhow much reaches people or ecosystemsEffect factor EFharm per unit of exposure; for people, from ED50​Characterization factorCF = EF · XF · FFiF = XF · FF. Human toxicity: CTUh, disease cases per kgemitted. Freshwater ecotoxicity: CTUe, affected fraction ofspecies over time and volume per kg emitted.
Rosenbaum RK, et al. Int J Life Cycle Assess 2008;13:532–546; Fantke P (ed.), USEtox 2.0 Documentation (2017), section 2.2 (CF = FF × XF × EF; intake fraction iF = FF × XF; CTUh in disease cases per kg emitted; CTUe in PAF·m³·d per kg emitted).
Explanation, details and sources

The model came from a comparison of seven existing models (CalTOX, IMPACT 2002, USES-LCA, BETR, EDIP, WATSON and EcoSense) on 45 organic chemicals. Harmonizing their most influential elements cut the spread between models from up to 13 orders of magnitude to no more than two for any substance.

Human effect factors start from the ED50, the lifetime dose that raises disease probability by 50%, with a default linear slope; for cancer most ED50 values are derived from TD50 values in the Carcinogenic Potency Database. Characterization factors are not normalized to a reference substance. Rosenbaum and colleagues (2011) report that their comparison of carcinogenic potency by inhalation and by mouth “suggests that results by one route can reasonably be used to represent another route.”

The model works at indoor, urban, continental and global scales, with rural air, agricultural and natural soil, freshwater and coastal marine compartments. Its team asks that interim factors (called “indicative” in the 2.0 documentation) always be used together with recommended ones, because leaving them out would score those substances as having zero impact; results dominated by interim factors carry lower confidence, and a sensitivity check can rerun them with recommended factors only. The current manual adds interfaces for product applications, such as personal care products and food contact materials, and describes its outputs as relative measures for screening, alternatives assessment and life-cycle assessment.

A CTUh value says how one emission compares with another in a life-cycle inventory. It is not a reference dose, a margin of exposure or a judgment that an ingested product is safe; those come from the methods in the sections above.

Rosenbaum RK, et al. Int J Life Cycle Assess 2008;13:532–546 (abstract); Rosenbaum RK, et al. Int J Life Cycle Assess 2011;16:710–727 (abstract); usetox.org model page and USEtox manual, retrieved October 2026; Fantke P (ed.), USEtox 2.0 Documentation (2017), sections 2.2, 3.1 and 8 (characterization factors and units, interim and recommended factors, human effect factors and the CPDB).

Fluid dynamics in physiology

Fluid-dynamics models of the stomach and airways explain flow patterns, but their validation is partial

Computational fluid dynamics (CFD) solves the equations of mass and momentum conservation on a 3-D geometry. For an incompressible Newtonian liquid in laminar flow, as in the stomach model below: ∇·u = 0 and ρ(∂u/∂t + u·∇u) = −∇p + μ∇²u, where u is velocity, p pressure, ρ density and μ viscosity (the Navier–Stokes equations).

Published CFD applications in the stomach and the respiratory tract, with validation status
ApplicationWhat was modeledValidation and stated limits
Gastric mixing (Ferrua and Singh 2010)An average-sized human stomach with antral contraction waves of up to 80% occlusion, solved in Fluent for contents of 10−3 Pa·s (water-like) and 1 Pa·s (honey-like). At high viscosity the retropulsive jet and eddies weakened and pressure rose, suggesting viscous meals mix poorlyReported as in good agreement with earlier experimental data. Upper-wall contractions were “to some extent, arbitrarily defined” for lack of measurements; wall roughness and the mucus layer were not modeled
Aerosol deposition (Longest and Holbrook 2012, review)Transport and deposition of inhaled pharmaceutical aerosols in the inhaler, nose and mouth, conducting airways and alveoliThe review concludes that more refined model validations are needed, together with concurrent experiments

Ferrua MJ, Singh RP. J Food Sci 2010;75:R151–R162; Longest PW, Holbrook LT. Adv Drug Deliv Rev 2012;64:296–311. For medical devices, FDA's final guidance Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions (CDRH, November 2023) sets a risk-informed credibility framework for physics-based models.

Related pages: ADME, PK and PBPK · Toxicology and risk assessment · Toxicology and exposure assessment · Academy: toxicology and dose · Green tea · Caffeine

Health information notice

These statements have not been evaluated by the Food and Drug Administration. Nothing on this site is intended to diagnose, treat, cure or prevent any disease. The information is for general education only. It is not medical advice and does not replace advice, diagnosis or treatment from a qualified health professional, or any medication a clinician has prescribed.

Do not start, stop or change any medicine, supplement or herbal product because of something you read here. Talk to your doctor or pharmacist first, especially if you are pregnant or breastfeeding, have a medical condition, take any medicines, are preparing for surgery, or are caring for a child. Research findings, traditional uses and safety data are reported as their sources state them; they are not recommendations or claims about any product. If you think you may have a medical emergency, call your doctor or 911 at once.

EUTECNIQ Sciences sells no supplements or health products today and receives nothing from any product or ingredient mentioned on this site; any commercial interest it later has in an ingredient will be stated on that ingredient’s page.

Ready to check what you take?

Free tools built on published reference values, each with its sources and its limits. Your lists stay in your browser.