Mapping Chemical Respiratory Sensitization: How Useful Are Our Current Computational Tools?

Chemical respiratory sensitization is an immunological process that manifests clinically mostly as occupational asthma and is responsible for 1 in 6 cases of adult asthma, although this may be an underestimate of the prevalence, as it is under-diagnosed. Occupational asthma results in unemployment for roughly one-third of those affected due to severe health issues. Despite its high prevalence, chemical respiratory sensitization is difficult to predict, as there are currently no validated models and the mechanisms are not entirely understood, creating a significant challenge for regulatory bodies and industry alike. The Adverse Outcome Pathway (AOP) for respiratory sensitization is currently incomplete. However, some key events have been identified, and there is overlap with the comparatively well-characterized AOP for dermal sensitization. Because of this, and the fact that dermal sensitization is often assessed by in vivo, in chemico, or in silico methods, regulatory bodies are defaulting to the dermal sensitization status of chemicals as a proxy for respiratory sensitization status when evaluating chemical safety. We identified a data set of known human respiratory sensitizers, which we used to investigate the accuracy of a structural alert model, Toxtree, designed for skin sensitization and the Centre for Occupational and Environmental Health (COEH)’s model, a model developed specifically for occupational asthma. While both models had a reasonable level of accuracy, the COEH model achieved the highest balanced accuracy at 76%; when the models agreed, the overall accuracy was 87%. There were important differences between the models: Toxtree had superior performance for some structural alerts and some categories of well-characterized skin sensitizers, while the COEH model had high accuracy in identifying sensitizers that lacked identified skin sensitization reactivity domains. Overall, both models achieved respectable accuracy. However, neither model addresses potency, which, along with data quality, remains a hurdle, and the field must prioritize these issues to move forward.