Task-Driven Knowledge Graph Filtering Improves Prioritizing Drugs for Repurposing

Abstract

Background Drug repurposing aims at finding new targets for already developed drugs. It becomes more relevant as the cost of discovering new drugs steadily increases. To find new potential targets for a drug, an abundance of methods and existing biomedical knowledge from different domains can be leveraged. Recently, knowledge graphs have emerged in the biomedical domain that integrate information about genes, drugs, diseases and other biological domains. Knowledge graphs can be used to predict new connections between compounds and diseases, leveraging the interconnected biomedical data around them. While real world use cases such as drug repurposing are only interested in one specific relation type, widely used knowledge graph embedding models simultaneously optimize over all relation types in the graph. This can lead the models to underfit the data that is most relevant for the desired relation type. We propose a method that leverages domain knowledge in the form of metapaths and use them to filter two biomedical knowledge graphs (Hetionet and DRKG) for the purpose of improving performance on the prediction task of drug repurposing while simultaneously increasing computational efficiency. Results We find that our method reduces the number of entities by 60% on Hetionet and 26% on DRKG, while leading to an improvement in prediction performance of up to 40.8% on Hetionet and 12.4% on DRKG, with an average improvement of 17.5% on Hetionet and 5.1% on DRKG. Additionally, prioritization of antiviral compounds for SARS CoV-2 improves after task-driven filtering is applied. Conclusion Knowledge graphs contain facts that are counter productive for specific tasks, in our case drug repurposing. We also demonstrate that these facts can be removed, resulting in an improved performance in that task and a more efficient learning process.

Ensembles of knowledge graph embedding models improve predictions for drug discovery

“…We attempted to mitigate this factor by including two KGs in our benchmark which share equivalent triples but vary in size. Furthermore, we followed a similar strategy as Ratajczak and colleagues [ 10 ] by training the KGEMs using validation and test splits that exclusively contained drug–disease triples. Outstanding questions remain regarding the performance of other non-benchmarked models and the inclusion of novel KGs.…”

“…We employed 10 different KGEMs: RESCAL [ 16 ], TransE [ 17 ], DistMult [ 23 ], ERMLP [ 18 ], TransH [ 24 ], ComplEx [ 25 ], HolE [ 26 ], ConvE [ 27 ], RotatE [ 28 ] and MuRE [ 29 ]. These models have been selected based on: (i) their variability in terms of modeling paradigms [ 11 ], (ii) their performance on benchmarks [ 11 ] and (iii) their prior use for applications in drug discovery [ 10 , 30 , 31 ]. Supplementary Table 1 summarizes the key properties of the models.…”

“…Furthermore, some of the node types removed (i.e., protein functions and side effects for BioKG, and phenotypes, pathways, anatomy, and biological function for OpenBioLink) can also be used as node properties (e.g., protein function) and are present in the KG for other biomedical applications (e.g., side effect prediction). For instance, in a recent drug repurposing study using KGEMs, the authors showed how filtering out entities that do not appear in metapaths connecting chemicals and diseases significantly improved the performance of the models [ 10 ]. This study showed that after this filtering, although the relative frequency of chemicals and diseases nodes increased, the frequency of other node types, such as pathway or cellular component nodes (the ones we removed) was significantly reduced.…”

“…Furthermore, similar to Ratajczak et al . [ 10 ], we distributed the drug–disease triples along the three splits with ~80% of the drug–disease triples in the train, ~10% in the validation and ~10% in the test split. Accordingly, the overall split ratio for all triples is ~96, 2 and 2% for the train, validation and test datasets, respectively.…”

“…Although for most non-biomedical KGs one may want to optimize the model to learn all relations present in the KG in equal proportions (e.g., predicting heterogeneous triples in social media networks), biomedical KGs typically entail that the model is focused on a specific relation type which corresponds to a biomedical application (e.g., drug–disease for drug discovery, protein–disease for target prioritization and drug-side effect for side effect prediction) [ 9 ]. This distinct condition typically required by applications of biomedical KGs can adversely affect the performance of KGEMs as the majority of these models have been benchmarked for non-biomedical KGs which require the simultaneous optimization of the model to predict all or multiple relation types [ 10 ].…”