Moving the Adverse Outcome Pathways Framework towards Practical Utility by Integrating Compound Profiling Data and Using Deep Learning

Despite a significant investment in in vitro and in vivo screening, clinical safety concerns are still a major cause of compounds being withdrawn from the development process. A lack of mechanistic understanding of safety findings, combined with sparse data sets, restricts the development of approaches capable of predicting earlier potential safety concerns for new chemical entities. It is often unclear whether these are truly surprise events or, with hindsight, could actually have been predicted based on the available data at the time. This failure to detect early signals of safety issues may arise due to problems at the interface between three main components of predictive analysis: observer, data and technology.

In this work we propose to employ novel mathematical approaches to data representation and modelling, such as deep learning, of the processes that lead to an adverse outcome in order to reduce or eliminate events of this type. Data will be compiled from within AZ and GSK and shared via Lhasa Limited Ltd, with the student having access to both data sources for model generation. Depending on the adverse event considered, we will either compile AOP frameworks for adverse events, and/or use deep learning approaches for cases where only observational data is available, and hence extend current AOP frameworks where indicated. In particular, nodes in deep networks can be interpreted with respect to their contribution to output toxicity, and hence information for AOPs can be derived from them.

The current project will involve GSK and AstraZeneca as industrial partners, with the charity Lhasa Ltd. as an 'honest broker', to exchange and pool compound profiling data for model generation. This framework is currently already being established in the context of the Cambridge Alliance on Medicines Safety (CAMS), and it hence improves significantly upon previous analyses which were based either on public data, or data from only a single source. A pilot study on Structural Cardiotoxicity has already been completed and it is currently being written up for publication, proving the viability and benefit of exchanging data for safety prediction in this framework. We now would like to broaden the approach to other adverse events where data is available.

This proposal directly puts into practice recommendations published recently, and it represents an advancement over previous approaches (such as eTox) in being able to have broader access to internal safety data from two major pharmaceutical companies in the UK, AstraZeneca and GSK, shared under confidentiality agreements via a Trusted Broker, Lhasa. This means that all biological data will be made available in the context of the project, including experimental protocols, quantitative data on exposure of compounds, and additional target-based and biological profiling information of compounds can be used in this project. Hence, we expect that the combination of access to data, and the utilization of state-of-the-art machine learning methods, will enhance our ability to predict drug safety and toxicity events considerably.

Increasing our understanding of the molecular mechanisms by which compounds can cause adverse events will lead to the implementation of much improved in silico and in vitro screens to detect safety risks will avoid unnecessary investments in preclinical and clinical studies.