Methods and Applications: Recommended Occupational Exposure Limits for GMA Using Benchmark Dose and Bayesian Model Averaging

This investigation utilized data from a 104-week inhalation carcinogenicity study conducted in Japan. The experimental design comprised four groups: three treatment groups and one control group, with 50 female and 50 male rats per group, totaling 400 animals. Subjects were exposed to GMA via inhalation for 6 hours daily, 5 days weekly, throughout the 104-week period. The administered concentrations were 0 (control), 0.6, 2.5, and 10 ppm for both sexes. Detailed study information is accessible at https://anzeninfo.mhlw.go.jp/user/anzen/kag/pdf/gan/0795MAIN.pdf.

Given the absence of definitive human epidemiological evidence for GMA exposure, with existing case reports and occupational investigations lacking precise exposure quantification and being confounded by other potential sensitizing agents, these studies were deemed unsuitable for OEL determination. The selected study adhered to Good Laboratory Practice standards, featured appropriate duration, and employed the relevant exposure route (inhalation), as acknowledged in IARC’s assessment. Furthermore, this study serves as the foundational evidence for GMA OELs established by JSOH, ECHA, and other regulatory bodies, validating its selection as the primary toxicological evidence for establishing GMA OELs.

For OEL assessment, chronic toxicity and carcinogenicity were identified as the primary critical effects of GMA. The study results were systematically analyzed, incorporating various uncertainty factors and categorizing endpoints into non-neoplastic and neoplastic lesions, with stratification by sex. Carcinogenicity outcomes were classified by overall and terminal rates, all of which were incorporated into the BMD analysis. Only endpoints demonstrating statistical significance (P<0.05) were included in the analysis.

The BMD methodology employs statistical models to estimate toxic response probabilities at specified doses, facilitating the identification of dose-response relationships and determination of lower safe doses. Using BMDS software (version 3.3.2, EPA, the United States), this study analyzed statistically significant endpoints (P≤0.05) using nine models: Dichotomous Hill, Gamma, Log-Logistic, Multistage Degree, Weibull, Logistic, Log-Probit, Probit, and Quantal Linear. The BMR level was set at 0.1, with a confidence level of 0.95. The endpoint yielding the lowest BMDL₁₀ was selected as the BMD for GMA. Model fit assessment incorporated goodness-of-fit analysis, statistical testing, residual analysis, Akaike information criterion (AIC) value evaluation, P value examination, and nested testing.

This study utilized BMDS software for modeling key effects. The software assigns prior probabilities to each model based on model selection criteria, with equal default weights assigned to all models. Using binomial sampling for dichotomous endpoints and Normal or Lognormal distributions for continuous data, the software employs Laplace approximation to correct prior density. It then performs BMD estimation through maximum a posteriori probability estimation, computes posterior probabilities across multiple models, and assigns differential weights for model averaging calculations. Bayesian model averaging enhances estimation accuracy by combining results from multiple models. The incorporation of prior information substantially reduces BMD estimation uncertainty and prevents the selection of extreme models that might occur when relying solely on AIC values. By comprehensively evaluating all candidate models, this approach minimizes model selection bias, thereby improving result accuracy and calculation reliability (9).

A PoD represents the dose at which an adverse effect manifests following specific exposure, whether determined empirically or through dose-response modeling (10). In our analysis, we employed BMA to assign differential weights across models for averaging calculations and selected the BMDL₁₀ as the PoD, following EFSA and EPA recommendations for quantal data.

Our analysis incorporated uncertainty factors (UFs) for interspecies and intraspecies differences, along with effect severity. This study applied an interspecies factor of 2.5 and, following ECHA recommendations, an intraspecies factor of 5 for worker populations (compared to 10 for general populations) when establishing derived ineffective response levels. For GMA-induced non-neoplastic lesions, which are reversible, this study applied an uncertainty factor of 1. However, given the severity of GMA-induced neoplastic lesions, this study implemented an uncertainty factor of 10.

The final PoD serves as the basis for OEL calculation. Using equal prior probabilities for all models, this study derived GMA OEL values using the following equation:

where, OEL means occupational exposure limit (ppm); PoD means point of departure; UFs means uncertainty factors.

Analysis of non-neoplastic lesions in male mice revealed 10 endpoints with BMDL₁₀ values ranging from 0.100 to 8.157 ppm. The respiratory metaplasia of the nasal cavity olfactory epithelium yielded the lowest BMDL (0.103 ppm), with optimal fit achieved using the Log-Probit model. In female mice, nine non-neoplastic endpoints produced BMDL₁₀ values between 0.077 and 6.825 ppm, with nasopharyngeal eosinophilic change showing the lowest BMDL₁₀ (0.077 ppm) using the Dichotomous Hill model. Detailed BMD information is presented in Table 1.

Analysis of tumorigenic endpoints revealed 16 distinct endpoints in male mice, with BMDL₁₀ values ranging from 0.756 to 10.197 ppm. The terminal rate of nasal cavity hemangioma demonstrated the lowest BMDL₁₀, best fitted by the Dichotomous Hill model. Female mice exhibited 16 neoplastic endpoints with BMDL₁₀ values ranging from 0.791 to 7.434 ppm, with uterine histiocytic sarcoma showing the lowest BMDL₁₀, optimally fitted using the Log-Logistic model. Table 2 summarizes these findings, which identify the respiratory system as the primary target organ for GMA toxicity in female rats.

Figure 1 illustrates the dose-response relationships derived from Bayesian model averaging across different endpoints. Table 3 presents the posterior probabilities and BMD values post-model averaging. The model averaging approach, which incorporated all viable alternative models while excluding extreme cases, yielded more robust results than classical single-model analysis. For male rats’ olfactory epithelial nasal cavity respiratory metaplasia, the Multistage, Quantal Linear, and Weibull models demonstrated superior fit and significantly influenced BMDL₁₀ calculations, receiving greater computational weight and yielding a model-averaged BMDL₁₀ of 0.118 ppm. Female rats’ nasopharyngeal eosinophilia change was best characterized by the Multistage, Quantal Linear, and Log-Logistic models, producing a BMDL₁₀ of 0.157 ppm. For mice carcinogenicity endpoints, the Probit, Multistage, and Quantal Linear models provided optimal fit, yielding BMDL₁₀ values of 1.733 and 1.081 ppm for males and females, respectively. The lower PoD for non-carcinogenic effects compared to carcinogenic effects indicates that intranasal lesions represent the most sensitive endpoint for GMA inhalation exposure. Application of UFs to the model-averaged results produced OEL values of 0.0094, 0.0126, 0.0139, and 0.0086 ppm, aligning with established limits in the EU (0.016 ppm), Japan (0.012 ppm), and the US (0.01 ppm). Current evidence supports 0.01 ppm as a protective PC-TWA for occupational GMA exposure.

Figure 1. 

Dose-response relationships of bayesian modeling average. (A) Male mice nasal cavity respiratory metaplasia: olfactory epithelium. (B) Female mice Nasopharynx eosinophilic change. (C) Male mice nasal cavity hemangioma. (D) Female mice uterus histiocytic sarcoma.

Abbreviation: Ma=model average; BMD=benchmark dose; BMDL=benchmark dose lower confidence limit.

This study employed BMD analysis and BMA for GMA risk assessment, utilizing animal studies to identify primary sites of toxic effects. The results demonstrated that GMA’s principal adverse effects manifest at initial exposure sites, specifically the foregut following oral exposure and respiratory tract after inhalation exposure. Chronic GMA exposure in mice induced carcinogenic effects, evidenced by increased tumor incidence in multiple sites including the nasal cavity, lungs, stomach, and uterus. Animal studies (3) have also established GMA’s reproductive toxicity, while several case reports (11-12) have documented allergic reactions in humans exposed to GMA.

The BMD approach demonstrates superior sensitivity compared to the NOAELs/LOAELs methodology, ensuring comprehensive identification of potentially sensitive endpoints for risk assessment. Our analysis yielded BMDL₁₀ values of 0.103 and 0.077 ppm as general toxicity PoDs, substantially lower than the NOAEL/LOAEL-derived PoD (0.6 ppm). These results produced OEL values (0.01 ppm) slightly below ECHA’s 8h-TWA (0.016 ppm) and JSOH’s OEL-M (0.012 ppm). The BMD approach offers distinct advantages: it transcends experimental dose limitations, shows reduced sensitivity to dose spacing, and incorporates both dose-response curve characteristics and statistical uncertainties from data quality. When statistical power is constrained by limited data points or high variability, the BMD approach provides more robust conclusions by considering the complete dose-response curve and addressing statistical limitations more effectively than NOAEL. Consequently, our derived PoD incorporates more comprehensive information and better reflects GMA’s actual toxic effect profile. The alignment of final OEL values, despite different methodological approaches (ECHA and JSOH applying a 10-fold uncertainty factor for LOAEL to NOAEL extrapolation), validates 0.01 ppm as a reasonable PC-TWA for GMA under current evidence.

The BMA methodology employed here utilizes a comprehensive dose-response model with weighted mathematical components to calculate model means, generating reliable estimates and confidence intervals. This approach leverages prior information to enhance parameter estimation precision while accounting for model uncertainty. Traditional single-model statistical approaches risk introducing “model selection error (13),” which can be mitigated through model averaging techniques (14). The method’s inherent capacity to address model uncertainty provides enhanced flexibility and reliability in both model selection and parameter estimation (9), ultimately establishing a more robust foundation for risk quantification (7).

This study was subject to some limitations. The selection of appropriate prior distributions for BMA in BMDS software presents challenges, particularly without comprehensive background knowledge. Furthermore, our reliance on animal test data for OEL recommendations may not fully reflect actual plant operational conditions, necessitating additional field studies for developing more practical OELs.

Our recommended PC-TWA of 0.01 ppm for GMA represents a conservative approach to worker protection. According to ECHA’s (4) T25 methodology dose-response relationship, this concentration corresponds to approximately 40 additional cancer cases per 100,000 exposed workers. Given GMA’s demonstrated high sensitization potential in animal tests and case reports, dermal exposure remains an important area for future research. While modern closed-system manufacturing processes under controlled conditions (2) facilitate maintaining low workplace GMA concentrations, current national OELs require revision. Specifically, establishing PC-TWA and revising maximum allowable concentration are crucial steps toward enhanced worker health protection.