Computer-Aided Discovery and Redesign for Respiratory Sensitization: A Tiered Mechanistic Model to Deliver Robust Performance Across a Diverse Chemical Space

Abstract

Asthma is among the most common occupational diseases with considerable public health and economic costs. Chemicals that induce hypersensitivity in the airways can cause respiratory distress and comorbidities with respiratory infections such as COVID. Robust predictive models for this end point are still elusive due to the lack of an experimental benchmark and the over-reliance of existing in silico tools on structural alerts and structural (vs chemical) similarities. The Computer-Aided Discovery and REdesign (CADRE) platform is a proven strategy for providing robust computational predictions for hazard end points using a tiered hybrid system of expert rules, molecular simulations, and quantum mechanics calculations. The recently developed CADRE model for respiratory sensitization is based on a highly curated data set of structurally diverse chemicals with high-fidelity biological data. The model evaluates absorption kinetics in lung mucosa using Monte Carlo simulations, assigns reactive centers in a molecule and possible biotransformations via expert rules, and determines subsequent reactivity with cell proteins via quantum-mechanics calculations using a multi-tiered regression. The model affords an accuracy above 0.90, with a series of external validations based on literature data in the range of 0.88−0.95. The model is applicable to all lowmolecular-weight organics and can inform not only chemical substitution but also chemical redesign to advance development of safer alternatives.

A Quantum-Mechanical Approach to Predicting Carcinogenic Potency of N-nitrosamine Impurities in Pharmaceuticals

“…69 In CADRE models (and models based on fundamentals of chemical interactions in general), the concept of applicability domain is less obvious. 70 Presently, we report out-of-domain predictions (based on out-of-range descriptor values) using a confidence score, 15,29,34 but we have repeatedly found that even lowerconfidence predictions are of high quality and aligned with experiment. 14,31 To this end, it can be proposed that models anchored in the underlying chemistry (vs. chemicals in the training set) have a 'softer' edge in their applicability domain, with fewer inter-and extrapolation concerns, as long as the molecular transformations are modeled appropriately and can be assumed to initiate the toxic endpoint in question.…”

“…38 Computational modeling. The tiered structure of the N-nitrosamine model mimics that of the CADRE skin and respiratory sensitization models, 15,34 given our continued reliance on modeling of molecular interactions and the relevance of covalent binding in KIEs across all three endpoints. For all compounds here, ionization, an important factor in nitroso carcinogenicity, 39,40 was assessed at biological pH (7.4).…”

“…28,30 Paired with transparent Linear Discriminant Analysis (LDA) and Multivariate Linear Regressions (MLR) modeling, this strategy has delivered robust predictions across multiple toxic endpoints and a diverse chemical space. 15,[31][32][33][34] In particular, CADRE has outperformed other tools when considering larger, heavily functionalized and biologically active APIs and their synthetic intermediates. 14,31 Here we report an extension to the CADRE platform for N-nitrosamines, describing our approach to developing a predictive model for their carcinogenic potency…”

Quantum-Mechanical Approach to Predicting the Carcinogenic Potency of N-Nitroso Impurities in Pharmaceuticals

“…The tiered structure of the N -nitroso model mimics that of the CADRE skin and respiratory sensitization models, , given our continued reliance on the modeling of molecular interactions and the relevance of covalent binding in KEs across all three endpoints. For all compounds here, ionization, an important factor in nitroso carcinogenicity, , was assessed at biological pH (7.4).…”

“…In dealing with mechanistic complexity and underlying uncertainty, a computational toxicologist can integrate explicit modeling of “known” events with statistical methods (assuming a large and chemically diverse dataset) and/or nonspecific reactivity approaches. , In our previous work, we have shown that robust toxicological models can be developed using the modular CADRE (computer-aided discovery and REdesign) platform. , CADRE relies on a tiered approach to bioavailability, metabolic activation, and covalent haptenation of biological targets by integrating an expert system with molecular simulations and QM calculations (Figure ). Because uncertainty is of concern even for the well-characterized toxicity pathways (e.g., dermal sensitization), CADRE balances site-specific QM calculations of known transformations with descriptors derived from frontier molecular orbital (FMO) theory, which captures the reactivity broadly. , Paired with transparent LDA and multivariate linear regression (MLR) modeling, this strategy has delivered robust predictions across multiple toxic endpoints and a diverse chemical space. ,− In particular, CADRE has outperformed other tools when considering larger, heavily functionalized, and biologically active APIs and synthetic intermediates. , …”