PharMeBINet: a general-Purpose approach for the integration of heterogeneous biomedical data Sources and its application to drug-drug interaction prediction

Biological and medical data has become available in large quantities, especially from high-throughput technologies. This increase in data causes problems for storage and analysis. As a solution, data is combined into databases, resulting in an increase in the number of biomedical databases in the last years. For some analyses, a combination of multiple databases is needed to gather all information relevant to uncover unknown associations. A common way to represent such combined heterogeneous data is a knowledge graph (KG). However, their robustness and quality need to be considered, as low-quality KGs can lead to poor Analysis results and decision-making. KGs with an approved quality are used for multiple tasks, like drug repurposing, or the prediction of drug-target interactions, drug-drug interactions, or side effects for polypharmacy.<br /><br /> This thesis analyzes and compares existing KGs regarding their adherence to defined essential features for KGs. A workflow is implemented for the construction of the pharmacological medical biochemical network (PharMeBINet) KG. PharMeBINet is compared to existing KGs and compared to the same features. An accompanying web portal is implemented allowing the inspection and analysis of the KG. An enrichment analysis module provides a broad spectrum of options for users to choose from, such as multiple tests, correction methods, input types, and enrichment types. The enrichment analysis is compared to existing enrichment tools using a published dataset. Additional analysis modules are implemented and described. Overall, PharMeBINet fulfills all defined KG features, in comparison to other KGs, and provides a robust and well-connected view on biomedical information. Existing drug-drug interaction (DDI) prediction methods are analyzed for their reproducibility. Furthermore, several DDI prediction methods are implemented. All methods are compared regarding their performance on the same dataset extracted from the PharMeBINet KG. The results demonstrate the applicability of PharMeBINet for machine learning-based DDI prediction.<br /><br /> Additional use cases are described, demonstrating the overall reusability and applicability of the PharMeBINet KG for various analyses. PharMeBINet is open-source and freely available at https://pharmebi.net.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2025

Author dc:creator

• Königs, Cassandra

Repository record source_url

https://pub.uni-bielefeld.de/record/3000996

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:3000996