Prediction of Drug-to-Drug Interactions through Zero-Shot Learning

Abstract

Drug-to-drug interactions (DDIs) are a crucial aspect of medication management. While estimates vary, some studies suggest that DDIs may be responsible for up to 20% of the adverse drug reactions requiring hospitalization. Conventional methods for predicting those interactions rely on analyzing the pharmaceutical properties of drugs, clinical findings, and literature references. In recent years, approaches based on machine learning have emerged as a promising alternative, taking advantage of the vast biomedical data currently available, to identify relations between drugs and side effects, leading to highly accurate predictions. In this thesis, we differentiate by adopting the Zero-shot learning (ZSL) paradigm to tackle the challenge of DDI prediction. ZSL is a modern ML technique, that enables models to generalize beyond the classes encountered during training and make predictions for unseen classes. To achieve this, we leveraged a ZSL framework that relies on feature vectors extracted from both instances and classes. The framework effectively tries to capture and simplify the complex underlying relationships between different drug pairs and side effects. We should mention that a single drug combination can result in multiple side effects, necessitating appropriate modifications to account for this possibility. Our goal is to develop a DDI prediction pipeline that, with the necessary adjustments, can serve as a valuable resource for identifying and mitigating potential drug-drug interactions.