How robust are transcriptomics-based points of departure when translated to in vivo doses by kinetic-modeling? A case study with the herbicide pendimethalin using HepG2 and HepaRG cells

Technical Summary: Robustness of Transcriptomics-Based Points of Departure for Pendimethalin

Problem Statement
Traditional risk assessment for agrochemicals relies heavily on in vivo animal studies to derive Points of Departure (PODs) for regulatory guidance values. While New Approach Methodologies (NAMs) offer a pathway to estimate protective exposure levels without new animal data, their predictive performance and robustness require validation against established in vivo outcomes. Specifically, there is a need to evaluate whether transcriptomics-based workflows, when combined with Physiologically-Based Kinetic (PBK) modeling for in vitro-to-in vivo extrapolation (IVIVE), can generate biologically plausible systemic toxicity thresholds. Furthermore, the relative contribution of uncertainty from toxicokinetic modeling versus transcriptomic data analysis remains unclear.

Methodology
This case study utilized pendimethalin (PDM), an agrochemical with extensive existing toxicological and kinetic reference data, to evaluate a transcriptomics-based NAM workflow. The study employed two distinct in vitro liver models and transcriptomic platforms:

1. Cell Models: HepG2 (immortalized hepatocellular carcinoma) and HepaRG (differentiated hepatic stem cells).
2. Transcriptomics Platforms: Targeted TempO-Seq in HepG2 cells and untargeted RNA sequencing in HepaRG cells.
3. Exposure: Cells were exposed to PDM for 48 hours across concentration ranges (0.07–202 µM for HepaRG; 0.8–56 µM for HepG2), selected based on cytotoxicity limits (IC25).
4. tPOD Derivation: Gene expression changes were analyzed using BMDExpress 3.0. Four statistical approaches were applied to derive Transcriptomic Points of Departure (tPODs):
   • 25th gene Benchmark Concentration (BMC).
   • 5th percentile of the BMC distribution.
   • 1st mode of the BMC density distribution.
   • No Observed Transcriptomic Effect Level (NOTEL-5), defined as the highest concentration with fewer than 5 differentially expressed genes.
5. IVIVE and PBK Modeling: Derived tPODs were translated to external doses using reverse dosimetry. Three distinct rat PBK models were utilized: the EPA's httk R package, the EFSA's TK-Plate online tool, and a BASF-internal 9-compartment TK-Estimator.
6. Human Population Modeling: A human population-based PBK model (using the httk package with a virtual population of 100,000 individuals) was used to derive a 5th-percentile Oral Equivalent Dose (OED) from the most sensitive tPOD.
7. Sensitivity Analysis: Sensitivity coefficients were calculated for key input parameters (fraction unbound, logKow, intrinsic clearance, and apparent permeability) to assess their impact on plasma $C_{max}$ predictions.

Key Results

• Transcriptomic Sensitivity: Both cell lines and platforms detected concentration-dependent gene expression changes. The most sensitive tPOD identified was 0.65 µM (25th gene BMC in HepG2 via TempO-Seq), while the HepaRG RNA-seq analysis yielded a most sensitive tPOD of 0.68 µM (NOTEL-5).
• IVIVE-Derived PODs: When translated to external doses, the IVIVE-derived PODs generally fell within the range of regulatory in vivo PODs reported for PDM.
  • The most sensitive prediction for the rat yielded a POD of 6.0 mg/kg bw/day.
  • The human population-based modeling resulted in a 5th-percentile OED of 0.068 mg/kg bw/day.
• Comparison to Regulatory Values: The predicted human OED (0.068 mg/kg bw/day) is close to the established regulatory Acceptable Daily Intake (ADI) of 0.125 mg/kg bw/day and the Acceptable Operator Exposure Level (AOEL) of 0.17 mg/kg bw/day. The rat POD (6.0 mg/kg bw/day) aligns with the NOAEL of 12.5 mg/kg bw/day derived from a chronic dog study (used for the ADI).
• Sources of Variability: The variability introduced by the choice of toxicokinetic model (PBK) was found to be of a similar magnitude to the variability associated with different transcriptomic analysis approaches (cell line, platform, and statistical method). Sensitivity analysis indicated that absorption and partitioning parameters (specifically $P_{app}$ and logKow) were the primary drivers of uncertainty in the kinetic predictions.
• Model Performance: Predicted plasma $C_{max}$ values from the PBK models were within 0.5 to 1.0-fold of measured in vivo rat data following a single oral dose, falling within the acceptable accuracy range for qualified PBPK models.

Significance and Claims
The authors claim that this study supports the feasibility of combining in vitro transcriptomics with IVIVE to estimate systemic toxicity thresholds for agrochemicals. The results suggest that transcriptomics-based NAM workflows can provide estimates of protective exposure levels that are comparable to those derived from traditional in vivo studies.

Crucially, the paper highlights that uncertainty in IVIVE contributes at least as much to the final prediction as uncertainty in transcriptomic data analysis. This finding underscores that improvements in Next Generation Risk Assessment (NGRA) workflows depend as much on advances in toxicokinetic modeling, dosimetry, and kinetic parameterization as on the refinement of transcriptomic methodologies.

The study concludes that while the approach shows promise for identifying protective systemic toxicity thresholds, broader evaluation across a wider range of agrochemicals and further methodological refinement (including the use of experimentally determined kinetic parameters and in vitro dosimetry) are required before routine regulatory application can be established. The authors maintain a modest stance, presenting the work as a proof-of-concept case study rather than a definitive replacement for current regulatory frameworks.