Open Source Toxicokinetics Package Adds Brain and Adipose Tissue Compartments

Computational toxicologists have extended the open-source High-throughput Toxicokinetics R package, known as httk, by building explicit physiological compartments for both brain and adipose tissues. Traditionally, chemical risk assessment has depended heavily on in vivo animal experiments to evaluate internal organ exposures, but conventional testing remains constrained by low throughput and difficulties translating animal findings to human outcomes. To help evaluate the tens of thousands of industrial and commercial compounds that currently lack comprehensive toxicity profiles, the United States Environmental Protection Agency originally established generic physiologically based toxicokinetic models within the httk environment. In earlier versions of this software suite, however, adipose tissue and brain structures were not represented as distinct physiological spaces but were instead aggregated into an undifferentiated generic compartment labeled rest of body, limiting the utility of the framework for neurotoxicity screening and lipophilic accumulation.

The newly engineered compartments address distinct biological mechanisms of chemical distribution across mammalian organ systems. In the updated architecture, adipose tissue is treated as a perfusion-limited organ, which assumes that the uptake and release of chemicals are primarily controlled by blood perfusion rather than physical tissue membranes. This design directly reflects the role of body fat as a highly vascularized storage depot that sequester lipophilic environmental contaminants, including persistent organic pollutants such as dioxins and polychlorinated biphenyls. Because plasma concentrations often fail to correlate closely with tissue concentrations in body fat, explicitly tracking adipose disposition offers a direct computational approach to evaluating long-term retention and bioaccumulation. In contrast, the brain compartment incorporates a blood–brain barrier component using a diffusion-limited assumption, mathematically restricting chemical transfer between brain capillary blood and the surrounding brain tissue matrix based on membrane permeability.

To translate the conceptual biology into numerical simulations, the authors constructed systems of ordinary differential equations using the GNU MCSim simulation environment alongside the R statistical computing framework. The differential equations account for specific vascular parameters, chemical-specific unbound fractions in plasma, and blood-to-plasma concentration ratios while tracking mass balance across physiological spaces. Adipose tissue volume and blood flow were derived by allocating specific fractions of total body weight and cardiac output, simultaneously subtracting these fractions from the remaining rest of body compartment to preserve whole-body physiological consistency. For the brain compartment, the blood–brain barrier permeability parameter was formulated using an effective permeability coefficient multiplied by vascular surface area, normalized against regional blood flow.

Partitioning of chemicals into tissues was estimated using a modified Schmitt method implemented through dedicated R routines within httk. This computational technique estimates tissue-specific partition coefficients using fundamental physicochemical inputs, such as the octanol-water partition coefficient, compound ionization, plasma protein binding, and tissue composition, calibrated against historical datasets of experimental partition coefficients. To parameterize the diffusion-limited blood–brain barrier, the team relied primarily on effective permeability values obtained from the parallel artificial membrane permeability assay for brain membranes, known as PAMPA-BBB. Where laboratory measurements were not accessible, predictions from a quantitative structure–activity relationship model trained on over one hundred chemicals were used. The resulting ordinary differential equations were compiled into optimized C code via MCSim to allow rapid numerical solving within the httk package source code.

To evaluate the predictive validity of the expanded software, researchers conducted a retrospective validation comparing model-simulated concentration-time profiles against curated in vivo experimental data extracted from the literature, the EPA CvT database, and the National Institute of Environmental Health Sciences Chemical Effects in Biological Systems database. The evaluation dataset was restricted to oral and intravenous dosing regimens in rats and humans. Chemicals requiring active transporter mechanisms at the blood–brain barrier, such as substrates of P-glycoprotein or breast cancer resistance protein, were explicitly excluded because the passive diffusion model cannot account for active efflux transport. After applying rigorous quality curation and applicability domain criteria, the final validation set comprised eight chemicals for the adipose compartment across various dosing regimens and 22 chemicals encompassing 62 dosing regimens for the brain compartment.

Evaluation of the adipose tissue compartment revealed robust linear alignment with observed experimental peak concentrations across the evaluated chemicals. Model predictions achieved a coefficient of determination of 0.96 with a ninety-five percent confidence interval spanning from 0.94 to 0.98, alongside a regression slope of 0.94. The log10 root mean square error for adipose predictions was 0.70, while the mean absolute log error reached 0.57 and the geometric mean fold error was 3.7. Across all individual experimental dosing scenarios evaluated in fat tissue, 31.7 percent of model predictions fell within a two-fold error window, 70.7 percent fell within five-fold, and 78.0 percent remained within a ten-fold margin of experimental maximum concentrations. On a compound-by-compound basis, five of the eight tested chemicals demonstrated predicted maximum concentrations that stayed within a five-fold range of observed experimental values.

When assessing maximum concentration predictions in the brain compartment, the specialized diffusion-limited model substantially outperformed an exploratory perfusion-limited brain model across all diagnostic criteria. Across the 22 curated test compounds, the diffusion-limited model achieved a coefficient of determination of 0.71, a regression slope of 0.75, and a log10 root mean square error of 0.83, compared to a coefficient of determination of 0.69, a slope of 0.69, and a root mean square error of 1.00 in the simpler perfusion-limited version. The geometric mean fold error in brain predictions was reduced from 7.1 in the perfusion model down to 4.7 in the diffusion-limited framework. Furthermore, 29.0 percent of predictions fell within two-fold, 56.5 percent within five-fold, and 72.6 percent within ten-fold of observed values, marking distinct improvements over the corresponding percentages of 22.6 percent, 43.5 percent, and 64.5 percent observed when the blood–brain barrier permeability term was omitted.

Because public releases of httk feature specific flags for per- and polyfluoroalkyl substances, the authors also examined brain model performance after excluding the two PFAS compounds remaining in the test set, specifically perfluorooctanoic acid and perfluorooctanesulfonic acid. This secondary analysis demonstrated minimal variation from the complete evaluation cohort, yielding a regression slope of 0.75, a coefficient of determination of 0.72, a log10 root mean square error of 0.84, a mean absolute log error of 0.68, and a geometric mean fold error of 4.8. Under this configuration, 25.0 percent, 55.4 percent, and 71.4 percent of brain predictions fell within two-, five-, and ten-fold bands of experimental observations, respectively. The investigators noted that large prediction errors for compounds like perfluorooctanoic acid, which exhibited a geometric mean fold error of 13.31, stem from complex physiological mechanisms such as saturable plasma protein binding and active renal resorption that generic high-throughput models do not capture.

Detailed scrutiny of individual chemical deviations highlighted key boundaries within the model applicability domain. For instance, ambroxol showed a high geometric mean fold error of 14.00, which the authors attributed to its origin as a botanical extract presenting analytical purity and stability hurdles. Highly lipophilic compounds with logP values above 5, including dieldrin in brain tissue and deltamethrin in adipose tissue, displayed elevated errors of 10.90-fold and 19.22-fold, respectively, likely resulting from limitations in the Schmitt partitioning algorithm for extreme lipophilicity as well as uncertainties in measured in vitro parameters. Analytical data quality also influenced outliers; raw chromatograms for cinnamic acid and L-methylephedrine exhibited low signal-to-noise ratios in the underlying experimental records. Overall, model discrepancies tended toward overprediction rather than underprediction of tissue concentrations, providing conservative exposure estimates that align with screening-level safety thresholds in risk assessment.

The study authors highlighted several foundational limitations that define the current scope of the open-source software and guide future development. The model relies entirely on passive permeability and tissue partitioning, rendering it unsuitable for chemicals whose disposition is governed by active efflux or influx transporters at the blood–brain barrier. Furthermore, human brain evaluation data relied exclusively on lumbar cerebrospinal fluid measurements rather than parenchymal tissue biopsies, introducing an anatomical surrogate into the validation workflow. In addition, the perfusion-limited adipose structure does not account for dynamic physiological processes such as lipolysis, weight fluctuation, or physical exercise, which could alter terminal elimination rates for persistent bioaccumulative chemicals. The authors also noted that the simulations operated using fixed point parameters without formal sensitivity or uncertainty analyses, establishing a baseline platform that future releases can expand upon as experimental toxicokinetic datasets continue to grow.

Subject of Research: Physiologically based toxicokinetic modeling of brain and adipose tissue distribution

Article Title: Extending the high-throughput toxicokinetics R package to include brain and adipose compartments

Article References: Unnikrishnan, A., Kreutz, A., Chang, X., Sluka, J. P., Li, D., Linakis, M. W., Wambaugh, J. F., Kleinstreuer, N., & Hogberg, H. T. (2026). Extending the high-throughput toxicokinetics R package to include brain and adipose compartments. Frontiers in Pharmacology, 17, Article 1895108. https://doi.org/10.3389/fphar.2026.1895108

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1895108

Keywords: physiologically based toxicokinetics, httk package, toxicology, chemical risk assessment, blood-brain barrier, adipose tissue, pharmacokinetics, in vitro to in vivo extrapolation, computational toxicology, in silico modeling, perfusion-limited modeling, diffusion-limited modeling, environmental chemicals, neurotoxicity prediction, bioaccumulation, MCSim