(128c) Integrating Process Safety, Risk Assessment, AI, and Systems Biology to Regulate Human Health and Eco-Risks Arising from Industrial Accidents and Occupational Exposure

In this study, a holistic approach is introduced for assessing the potential impact of industrial accidents or daily occupational exposure on worker well-being with the associated short- and long-term health implications. Our methodology leverages mechanistic models and machine learning techniques from numerous research fields including process safety, computational fluid dynamics, toxicokinetic models, multimedia models and artificial intelligence for data gap filling. To illustrate the application of our approach, a use case is developed herein, focusing on bisphenol A, a chemical of widespread use and of significant concern due to its characterization as an endocrine disruptor and its association with a plethora of metabolic syndromes and adverse outcomes. Bisphenol A finds extensive use in food packaging, thermal paper, and as an essential component in plastics manufacturing, with a global annual production estimated at approximately 7,000 kt.

Bisphenol A is primarily synthesized through the condensation reaction of phenol and acetone with the presence of an acid catalyst. Despite its apparent simplicity, this production process carries a fire hazard due to the highly flammable nature of the involved materials. The scope of accidents investigated in this industry pertains to fire incidents, encompassing all stages of the production process. Given its relatively high molecular weight, bisphenol A is non-volatile and only investigated for pool fire accidents, in contrast to the raw materials explored for flash fires, jet fires, and explosions. To assess the industrial-level consequences of potential accidents, we employed the DNV software package, Phast, which facilitated both quantitative analysis and computational fluid dynamics modeling. This approach allowed us to calculate and visualize the dispersion of hazardous plumes across the entire facility. Furthermore, utilizing spatial integration and fluid dynamics insights provided by the software, we were able to position hypothetical workers along the length of the installation, enabling the consideration of multiple exposure scenarios for quantitative risk assessment.

The functionalities of the Phast software are valuable for calculating toxicological consequences, offering insights into the probability of life-threatening outcomes using specific coefficients with the Probit function. However, for a more comprehensive examination for the impact of an accident on the health of individuals exposed, a deeper investigation is necessary. In our quest to advance this analysis, we have incorporated the INTEGRA software package into the developing methodological pipeline (Sarigiannis et al., 2014; Sarigiannis et al., 2016). INTEGRA empowers users to explore the entire trajectory from the initial exposure event to the environmental context (source-to-dose-continuum). With a detailed level three multimedia model, INTEGRA facilitates an in-depth environmental assessment of the potential consequences of an accident. This environmental assessment proves particularly vital as it evaluates the impact on micro-environments within the affected area. Leveraging the CFD model provided by Phast, we can precisely determine the concentration of bisphenol A (BPA) across the three primary exposure routes: inhalation, oral ingestion, and skin contact. This information, in turn, allows us to employ the detailed Physiologically Based Toxicokinetic (PBTK) model provided by the platform. Through the PBTK model, we convert external doses in the three exposure routes into internal doses, a crucial step that enables us to delve into molecular and organ-level interactions between xenobiotics and endogenous factors.

To achieve this, we integrate data from both in vivo and in vitro studies, along with human biomonitoring data. This comprehensive dataset is then enriched with metabolomics and transcriptomics analyses specific to bisphenol A (BPA), including numerous concentration levels corresponding to different exposure scenarios. From these data, we construct integrated dose-response curves, effectively relating concentration levels to interactions with a range of endogenous biomarkers. The next step involves the development of a comprehensive systems biology model capable of encompassing a wide array of biological mechanisms to quantify the potential impact of BPA exposure. To achieve this, multi-omics data are combined to facilitate a joint pathway analysis. These biological pathways are then integrated into the methodology presented by Karakoltzidis et al. (2023) to introduce Next Generation Systems Biology models. These models consist of ordinary differential equations accounting for all endogenous factors within a biological system, including proteins, enzymes, and endogenous metabolites. This approach offers a holistic understanding of the intricate interactions and responses within the biological system resulting from BPA exposure.

Parameterizing such a complex mathematical model presents significant challenges. To address this, we harnessed data from a multitude of publicly accessible databases, including BRENDA, SABIO RK, HMDB, BioModels, UniProt, KEGG, and more. We developed in-house tools to compile and organize this wealth of information, resulting in expansive datasets, totaling many gigabytes. These datasets served as the foundation for the development of AI models. Using Deep Learning architectures, we constructed three models capable of accurately predicting kinetic parameters, including Michaelis-Menten constants and turnover numbers. All these models achieved impressive performance, with R-squared values exceeding 0.65, signifying their robust predictive capabilities. For the calculation of initial concentrations, a Machine Learning methodology was employed. It involved the use of unified networks to establish regression correlations, estimating the potential concentrations of unknown metabolites based on their interactions with reactants and products. In cases where training temporary unsupervised models led to overfitting or underfitting, data augmentation techniques were applied to enrich the dataset, ensuring the reliability and comprehensiveness of the model's predictions.

To establish a quantitative relation between the model including biomarkers and phenotypic diseases, we applied Natural Language Processing (NLP) techniques. We developed three token classification models, one for diseases, one for endogenous metabolites, and one for biomedical text mining, which could identify a wide array of biological entities, including DNA, proteins, cell types, cell lines, genes, and RNA. These models demonstrated excellent performance, with F1 scores surpassing 0.7; they were trained based on publicly available data.

Before collecting the model outputs, the results underwent in-house evaluation. Using these NLP tools, we scoured over 28,000 articles in PubMed, ultimately identifying 143 studies that linked endogenous metabolites to disease phenotypes. To further refine our focus, we concentrated on metabolic diseases such as obesity and different types of diabetes. We did that because BPA is deemed by the toxicological community to be highly linked with these outcomes. Consequently, by conducting a series of simulations, we could not only pinpoint BPA levels in the human body but also assess whether an accident could trigger adverse outcomes for the workers. This holistic approach provides a comprehensive understanding of the potential health impacts of BPA exposure enabling an on-site assessment that encompasses both short-term and long-term consequences of accidents, providing a comprehensive view of potential risks. Furthermore, industries can use this methodology to establish regular occupational health assessments, thus fostering a healthier and more efficient workforce. This dual benefit, combining safety and worker well-being, offers a comprehensive solution for industries seeking to improve their operations and safeguard their employees.

Karakoltzidis, A., Karakitsios, S., & Sarigiannis, D. (2023). MINER: A Text Mining Approach for the Introduction of Next-Generation Systems Biology Models. Toxicology letters, 384. https://doi.org/10.1016/S0378-4274(23)00538-6

Sarigiannis, D., Karakitsios, S., Gotti, A., Loizou, G., Cherrie, J., Smolders, R., De Brouwere, K., Galea, K., Jones, K., Handakas, E., Papadaki, K., & Sleeuwenhoek, A. (2014). Integra: From global scale contamination to tissue dose. Proceedings - 7th International Congress on Environmental Modelling and Software: Bold Visions for Environmental Modeling, iEMSs 2014,

Sarigiannis, D. A., Karakitsios, S. P., Handakas, E., Simou, K., Solomou, E., & Gotti, A. (2016). Integrated exposure and risk characterization of bisphenol-A in Europe. Food and Chemical Toxicology, 98, 134-147. https://doi.org/10.1016/j.fct.2016.10.017