Knowledge Graphs for Indication Expansion: An Explainable Target-Disease Prediction Method

Abstract

Indication expansion aims to find new indications for existing targets in order to accelerate the process of launching a new drug for a disease on the market. The rapid increase in data types and data sources for computational drug discovery has fostered the use of semantic knowledge graphs (KGs) for indication expansion through target centric approaches, or in other words, target repositioning. Previously, we developed a novel method to construct a KG for indication expansion studies, with the aim of finding and justifying alternative indications for a target gene of interest. In contrast to other KGs, ours combines human-curated full-text literature and gene expression data from biomedical databases to encode relationships between genes, diseases, and tissues. Here, we assessed the suitability of our KG for explainable target-disease link prediction using a glass-box approach. To evaluate the predictive power of our KG, we applied shortest path with tissue information- and embedding-based prediction methods to a graph constructed with information published before or during 2010. We also obtained random baselines by applying the shortest path predictive methods to KGs with randomly shuffled node labels. Then, we evaluated the accuracy of the top predictions using gene-disease links reported after 2010. In addition, we investigated the contribution of the KG’s tissue expression entity to the prediction performance. Our experiments showed that shortest path-based methods significantly outperform the random baselines and embedding-based methods outperform the shortest path predictions. Importantly, removing the tissue expression entity from the KG severely impacts the quality of the predictions, especially those produced by the embedding approaches. Finally, since the interpretability of the predictions is crucial in indication expansion, we highlight the advantages of our glass-box model through the examination of example candidate target-disease predictions.

An Open Access Strategy for the Drug Repurposing Community

“…Other researches develop hybrid architectures that make predictions based on embeddings or other non-interpretable methods and apply some technique to provide explanations [17]- [20]. In [11], KGML-xDTD is developed to predict repurposing candidates using embedding methods and random forest.…”

Explainable drug repurposing via path-based knowledge graph completion

“…Other researches develop hybrid architectures that make predictions based on embeddings or other non-interpretable methods and apply some technique to provide explanations [17]- [20]. In [11], KGML-xDTD is developed to predict repurposing candidates using embedding methods and random forest.…”

pyRDF2Vec: A Python Implementation and Extension of RDF2Vec

“…Shi et al[19,18] use pyRDF2Vec to calculate semantic similarity between concepts in several datasets. Gurbuz et al[8] evaluate many different techniques, including pyRDF2Vec, for explainable target-disease link prediction. Steenwinckel et al[21] compare their newly proposed technique, INK, to state-of-the-art techniques such as pyRDF2Vec.…”