Drug Induced Liver Injury Prediction with Injective Molecular Transformer

Abstract

Drug-Induced Liver Injury (DILI), liver damage caused by drugs, represents a significant factor contributing to the failure of clinical trials. Remarkably, the drug development process, which entails an extensive timeline spanning several years and incurring costs of billions of dollars to achieve Food and Drug Administration (FDA) approval, could greatly benefit from early DILI prediction. Furthermore, through the utilization of DILI prediction, clinicians can obtain valuable insights into the potential risks associated with medication, empowering them to make more informed decisions when prescribing drugs to patients. We employ Graph Neural Networks (GNNs) to predict DILI based on drug structures. GNNs consist of node aggregation, which gathers node representations, and graph pooling, which compiles node representations to portray the graph as a single vector. While the graph pooling method built on Set Transformer outperforms existing techniques, we identify a limitation: Set Transformer uses a random seed vector as the query vector that cannot differentiate between graphs of varied structures. Moreover, it potentially lacks expressiveness, as it is randomly defined without prior knowledge and relies on a limited number of seed vectors. To overcome the issues, we introduce Molecular Transformer that employs unique molecular representations as the query vectors. We find that using drug toxicity information extracted from relevant knowledge-bases as the query vector yields the best performance.