Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.787786
Title: Structure-based predictions for molecular initiating events
Author: Wedlake, Andrew John
ISNI:       0000 0004 7972 8968
Awarding Body: University of Cambridge
Current Institution: University of Cambridge
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Toxicity testing of chemicals is currently undergoing its largest ever paradigm shift, moving towards faster, cheaper and more human-relevant methods which focus on mechanistic understanding. An AOP provides a framework for organising biological knowledge and data. The gateway to an AOP is the MIE, and chemistry is key to predicting which chemicals can undergo a MIE. In silico predictions of MIEs are a vital tool in a modern, mechanism-focused approach to risk assessment of chemicals. In this project, new structural alert-based models for receptor binding MIEs have been constructed that create accurate, transparent and interpretable predictions. The alerts have been constructed with an automated workflow that uses Bayesian statistics to iteratively select substructures associated with activity. The models were constructed from balanced data sets taken from human in vitro¬ assays in the ChEMBL and ToxCast databases. The new models significantly improve on previous models, with performance metrics comparable to random forest models. Methods for further improving structural alert models are presented, including a method for generalising aromatic atoms in structural alerts to reduce the number of alerts in a model, and construction of a consensus model combining structural alerts with a random forest model. Structural alert models have been constructed for a wide range of biological targets of toxicological interest and the variation in performance across all targets has been explained by considering the proportion of activity cliffs in data sets. Having significantly improved structural alert models in terms of performance, new methods for assessing confidence in both active and inactive predictions have been developed. These involve considering similarity to relevant chemicals in the training set. The measure of confidence in active predictions allows for applicability of predictions to be evaluated, whilst the measure of confidence in inactive predictions is vital in risk assessment of chemicals. Moving beyond structural alerts, attempts to describe chemicals in terms of the key interactions made with the biological target have been made. This a step towards describing how the receptor binding MIEs work and then using this knowledge to make better activity predictions. The generalised aromatic structural alerts have been used to predict key receptor binding interactions, which are consistent with interactions derived from crystal structures. Using structural alerts to group chemicals, pharmacophore models have been developed, allowing for activity predictions in terms of general features in three-dimensional space instead of the specific combination of atoms and bonds described by structural alerts.
Supervisor: Goodman, Jonathan Sponsor: Unilever
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.787786  DOI:
Keywords: Chemistry ; Computational Chemistry ; Toxicology ; Risk Assessment ; Machine Learning ; Structure Activity Relationship ; Structural Alerts ; Pharmacophore ; Molecular Initiating Event ; Adverse Outcome Pathway ; In silico Predictions
Share: