Abstract 3664: SafetyNet: a knowledge graph framework for toxicity management and safety profile prediction

Abstract Background: Anti-cancer treatments can range from systemic therapy, like chemotherapy, to targeted approaches, such as antibody-drug conjugates. In any case, such therapies cause adverse events (AEs) in cancer patients. In this study, we have developed a dedicated knowledge graph (KG) for AE risk factor prediction - SafetyNet. Specifically, this study aims to deliver class-specific risk factors, inform toxicity management guidance, and elicit a strategy for risk mitigation and monitoring. Ultimately, we aim to devise a standardized framework to monitor AEs in clinical trials. Methods: SafetyNet was built from synthetic data of 21 oncology trials, which was created from and follows the same distribution as AstraZeneca trial data and is therefore a suitable interim substitute. Real ADC trial data will replace the synthetic data in the near future. In addition, SafetyNet includes biomedical entities and relationships from AstraZeneca’s internal Biological Insights Knowledge Graph, with prior knowledge from literature as well as safety intelligence data from OFF-X. We find trends in the data by querying SafetyNet via the personalized PageRank (PPR) method. Briefly, PPR works by simulating multiple random walks in the graph, from which we count the number of visits to each node. Nodes that are visited more often have higher PPR, meaning they are more important to the starting node. Results: There are different strategies that can be employed to obtain insights. We can, for example, focus on a given AE: for patients that had nausea as an AE of the primary treatment, we found hypertension to be the most frequent comorbidity, with 367 patients affected. To complement this observation, we found that there are 543 concomitant medications that have both nausea and hypertension as a known AE. We can also analyze trends between serious and non-serious cases of a given AE: serious vomiting was more tightly linked to the use of tinzaparin sodium, an antithrombotic drug. Accordingly, pulmonary embolism and deep vein thrombosis were more frequent in patients experiencing serious vomiting vs non-serious. We can equally analyze patients that had a particular concomitant medication: diarrhea, nausea and vomiting were all higher ranked in patients that were taking cisplatin concomitantly with the primary treatment. Conclusions: By utilizing SafetyNet, we can identify the risk factors associated with AEs for a given ADC as well as their relationship with concomitant medications. By doing these analyses across ADC trials, we gain generalized insights that can help our understanding of AEs in the ADC space. In turn, these insights can be used to create standardized guidelines for ADC toxicity guidance and AE monitoring. Furthermore, SafetyNet brings considerable time saving and human error reduction relative to manually analyzing all the clinical, literature and OFF-X data. Citation Format: Miguel Goncalves, Mark O’Donoghue, Abigail Morrison, Heidi Lloyd-Williams, Sanjana Singarayer, Neki Patel, Samad Jahandideh, Ken Twomey, Krishna C. Bulusu. SafetyNet: a knowledge graph framework for toxicity management and safety profile prediction [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3664.