Improved Detection of Drug-drug Interaction

Lead Research Organisation: University of Liverpool
Department Name: Electrical Engineering and Electronics

Abstract

Adverse drug-drug interactions (DDIs) harm large numbers of patients every year. Since clinical trials for a drug are necessarily limited in their scale relative to the downstream use of the drug, not all DDIs are known when new medicines are made available to the general public. As a result, reporting databases have been developed to collate information about individuals who experience adverse events that could be side-effects caused by a DDI. Databases of known DDIs exist, but it is possible that reporting databases contain information about other DDIs that are currently unknown to the world and which could be assessed for clinical plausibility (some potential DDIs might look possible from the data but not be plausible from a clinical perspective). Mature algorithms exist to analyse reporting databases. These mature algorithms detect when the number of reports of an adverse event is statistically different when people take a specific single drug relative to when they do not. However, the algorithms for detecting DDIs in the same way are less mature.
Recent work in a project involving signal processing (i.e. Electronic Engineering) and pharmacology (i.e. Translational Medicine) experts at the University of Liverpool and the Patient Safety team at AstraZeneca has highlighted that existing algorithms for detecting drugs' side-effects can be reformulated in terms of hypothesis testing. This has enabled existing algorithms to be rearticulated in an explicit Bayesian context, providing a strong foundation for developing new algorithms that consider the detection of DDIs in terms of such Bayesian inference. This is important because it transpires that the hypothesis space will be large when considering potential interactions between large numbers of drugs: to test if one specific set of N drugs (of which there will be very many) are interacting, it appears that there are 2^(2^N-1) hypotheses that need to be considered. To make it feasible to navigate the space of all the hypotheses relevant to potential DDIs hidden in a database of reports for many drugs, there is therefore a need to capitalise on efficient numerical Bayesian inference techniques. The focus of this PhD is to understand the existing state-of-the-art in terms of analysing databases of adverse events before developing, applying, assessing and extending a Bayesian approach to detecting DDIs. The aim is to ensure that the resulting algorithms are readily used by AstraZeneca, as an exemplar pharmaceutical company keen to use algorithms that outperform the existing state-of-the-art in terms of detecting potential DDIs.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R51231X/1 01/10/2017 30/09/2021
1959810 Studentship EP/R51231X/1 25/09/2017 15/01/2022 Elpida Kontsioti
 
Description Three well-established online clinical resources for drug-drug interactions (DDIs) were compared in terms of listing of DDIs, as well as the consistency of information related to severity, evidence and clinical management options. Based on the information provided by those three DDI resources, their intersection was used in order to establish a reference set of positive test cases (i.e. DDIs that are known to lead to clinically identifiable adverse events) as well negative test cases (i.e. drug pairs for which no evidence in terms of a potential DDI could be found in the resources under consideration). This reference set has been used to evaluate the performance of three existing signal detection algorithms for DDIs when scanning the US Food and Drug Administration Adverse Event Reporting System (FAERS), a database that contains information on adverse event and medication error reports. Additionally, a set of design criteria for reference sets in pharmacovigilance (i.e. inclusion of test cases that meet specific requirements, such as evidence availability, minimum report count, etc) has been applied to the established reference set to assess the performance fluctuation and relative impact on the algorithms under consideration.
Exploitation Route One of the project aims is to release an open-access version of the established reference set for DDIs. Although a definitive reference standard including the complete set of DDIs cannot exist, the automatic extraction and aggregation of information from multiple clinical resources on DDIs enabled us to construct, share and advocate a reference set that can be used to facilitate research in signal detection and allow a common ground for comparing methodologies. This scalable approach requires less manual effort for future updates, considering the dynamic nature of data and evidence availability.

Also, the industry partner, AstraZeneca (AZ), has been actively involved in the project, by providing continuous support and feedback. It is expected that the outcomes of this project (i.e. reference sets, algorithms, other methodologies and frameworks that will be developed related to biological plausibility of generated signals) can be readily used by AZ, as an exemplar pharmaceutical company keen to use methodologies and algorithms that outperform the existing state-of-the-art in terms of detecting potential DDIs.
Sectors Pharmaceuticals and Medical Biotechnology

 
Description Collaboration with AstraZeneca Patient Safety Centre of Excellence Team 
Organisation AstraZeneca
Country United Kingdom 
Sector Private 
PI Contribution This collaboration aims to provide AstraZeneca (AZ) with novel algorithms for detection of signals related to drug-drug interactions using data sources traditionally used in pharmacovigilance (i.e. spontaneous reporting databases), but potentially other alternative data streams that have been recently identified (i.e. social media, literature) as well. Procedures and methodologies have been intended to be automated to the extent possible, in order to enable handling of large amounts of data, reproducibility and future modifications. This approach can be identified as novel in the field of pharmacovigilance, where a considerable amount of manual effort still characterises most of the activities.
Collaborator Contribution Regular discussions with researchers from AZ regarding progress and any challenges related to the project can be identified as significant contribution, given their level of expertise. Also, access to AZ data has been facilitated by a company's laptop that has been made available to EK (PhD Student).
Impact To date, curation of the FDA Adverse Event Reporting System database, one of the largest spontaneous reporting databases worldwide, and the creation of a reference set suitable for algorithm evaluation have been completed.
Start Year 2017