Abstract 6312: The adverse events atlas, towards a strategy to predict synergistic adverse events of combination therapies

The current gold-standard of estimating adverse events of a drug are clinical trials. However, these trials fail to represent real-life practice where patients have additional diseases (comorbidities) and commonly use numerous drugs when entering the clinic. Therefore, there is a rise of interest in combinational therapies, also because effective combinations are expected to prevent therapy resistance. At this moment it is not feasible to predict the adverse events of new combination therapies due to lack of available information both from a dimensionality (i.e., number of adverse events recorded per patient) as well as from a patient-number perspective. When available, this information allows to choose combinational therapies with acceptable adverse events. In this study we developed a preliminary method to predict adverse events of drug combinations in order to select combinations with a mild adverse event profile. We used the FAERS, an FDA post-marketing adverse events registry as data source containing 15 million adverse-event records. First, we developed a method to visualize the adverse events profiles of monotherapy and combination therapy using dimension reduction to accurately represent the relation between adverse events over many patients. These adverse-event profiles are then fed to a convolutional neural network (CNN) to generate an explainable prediction-model for adverse events occurring in combination therapies. The CNN trained on monotherapy is able to learn from the data and recognize adverse event patterns. The learned pattern information, as stored in the so-called latent space, can be converted back onto the original adverse event profiles. This showed a high similarity to the original data, also for unseen combination therapy effects. Furthermore, a t-SNE analysis on the latent space of the CNN is able to separate additive and synergistic adverse event patterns in combination therapy. Our CNN model can successfully learn complex adverse-event patterns for single drugs and their combinations, which are all encoded in the latent space. The developed method is therefore applicable to determine the combinatorial effects of highly complex adverse event profiles. Citation Format: Asli Kucukosmanoglu, Silvia Scoarta, Thomas Wijnands, George Kanev, Bart Westerman, Bert Kiewit, David Noske, Tom Wurdinger. The adverse events atlas, towards a strategy to predict synergistic adverse events of combination therapies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 6312.