Performance of pharmacokinetic models developed using in-vitro-to-in-vivo extrapolation (IVIVE) methods may be improved by refining assumptions regarding fraction absorbed (Fabs) through the intestine, a component of oral bioavailability (Fbio). Although in vivo measures of Fabs are often unavailable for non-pharmaceuticals, in vitro measures of apparent permeability (Papp) using the Caco-2 cell line have been highly correlated with Fabs. We measured bidirectional Papp for over 400 non-pharmaceutical chemicals using the Caco-2 assay. A random forest quantitative structure-property relationship (QSPR) model was developed using these and peer-reviewed pharmaceutical data. Both Caco-2 data (R2 = 0.37) and the QSPR model (R2 = 0.29) were better at predicting human bioavailability compared to in vivo rat data (R2 = 0.23). After incorporation into a high-throughput toxicokinetics (HTTK) framework for IVIVE, the Caco-2 data were used to estimate in vivo administered equivalent dose (AED) for bioactivity assessed in vitro. The HTTK-predicted plasma steady state concentrations (Css) for IVIVE were revised, with modest changes predicted for poorly absorbed chemicals. Experimental data were evaluated for sources of measurement uncertainty, which were then accounted for using the Monte Carlo method. Revised AEDs were subsequently compared with exposure estimates to evaluate effects on bioactivity:exposure ratios, a surrogate for risk. Only minor changes in the margin between chemical exposure and predicted bioactive doses were observed due to the preponderance of highly absorbed chemicals.
Keywords: Caco-2 assay; QSPR model; bioavailability; in vitro-in vivo extrapolation; intestinal absorption; toxicokinetics.
When assessing any chemical risk posed to the public health, there is a crucial difference between a dose ingested orally and the amount that enters the bloodstream by absorption from the gut (the rest of the chemical is eliminated from the body). The in vitro Caco-2 permeability assay is used by the pharmaceutical industry to identify chemical compounds that will be well absorbed. Here we have used the assay to screen more than 400 chemicals occurring in commerce and the environment. We have further developed a machine learning model to predict this property for other chemicals without experimental data. Due to differences between species, in vitro and machine learning approaches both appear to work better than animal studies for predicting human bioavailability. We predict that many, but not all, non-pharmaceuticals are well absorbed. We show how these new data refine health risk-based chemical prioritizations.