(062) Identifying Adverse Drug Reactions Through Knowledge Graph Modeling

Background: For signal detection on drug - Adverse Drug Reaction (ADR) associations, the Netherlands Pharmacovigilance Centre Lareb makes use of disproportionality analysis (Reporting Odds Ratio (ROR)). Information on drug characteristics from external databases (e.g. DrugBank) could possibly be used as supportive evidence in signal detection.

Objectives: In our work, we explore the usage of knowledge graph embeddings and link prediction to support signal detection on drug – ADR associations. By building a knowledge graph, we link our data to external databases such as DrugBank and make use of additional information, that is taking into account when the knowledge graph is embedded and aids in the link prediction process.

Methods: The association between immunosuppressive drugs and the ADR ‘photosensitivity reaction’ was used as test-case. 2358 ICSRs on immune suppressive drugs received until July 17, 2023 were extracted from the our database. Patient, drug and ADR characteristics were mapped using the Drug Repurposing Knowledge Graph (DRKG), that includes information from six existing databases: DrugBank, Hetionet, GNBR, String, IntAct and DGIdb. Drugs were mapped using their Anatomical Therapeutic Chemical (ATC) code and ADRs using Medical Dictionary for Regulatory Activities (MedDRA) ontology. Three link prediction models were tested: ComplEx, DistMult and HolE, and evaluated using the mean reciprocal rank (MRR) and Hit@n. Correlations between predicted probabilities and the ROR were assessed using Spearman rank-order correlation.

Results: We successfully linked the data to the external databases according to ATC- and MedDRA ontology. All link prediction models had similar MRR values ranging from 0.23 – 0.30. HolE outperformed the other models across all values of N (10, 3 and 1), suggesting better ranking the correct answer. Top 5 drugs with highest probability were: pirfenidone (prob=0.83), azathioprine (prob=0.74), mercaptopurine (prob=0.72), methotrexate (prob=0.70) and raltitrexed (prob=0.67). Except for raltitrexed, photosensitivity was reported in the product information of these drugs. Model predictions positively correlated with the ROR (ComplEX ρ=0.82, p=0.007, DistMult ρ=0.83, p=0.005 HolE ρ=0.68, p=0.002).

Conclusions: Knowledge graph embedding and link prediction was successfully implemented and could provide supporting evidence for drug – ADR associations. Future work on the whole dataset should be performed to find possible new signals or provide supporting evidence for existing drug-ADR suspicions.