Machine learning-based drug-drug interaction prediction: a critical review of models, limitations, and data challenges

Background/Objectives New computational methods, based on statistical, machine learning, and deep learning techniques using drug-related entities (e.g., genes, protein bindings, etc.), help reduce the costs of in-vitro experiments through drug-drug interaction prediction (DDIp). This review examines recent advances in DDIp. It presents an in-depth review of the state-of-the-art studies relating to semi-supervised, supervised, self-supervised learning, and other techniques such as graph-based learning and matrix factorization methods for predicting DDIs. All possible interactions between drugs are not known, and accurately predicting interactions is even more difficult due to the complex nature of drug-drug interactions (DDI). Methods Of the 49 papers published in Web of Science in the last 6 years, 24 papers were considered relevant based on information presented in their titles and abstracts. The included articles focus specifically on predicting DDIs using a type of machine learning algorithm. Excluded articles focused on drug discovery, drug repurposing, molecular representation, or the extraction of biomedical interactions. The methodology, results limitations, and future research directions were studied for each paper. Common challenges, limitations, and future research directions were analyzed. Results and conclusion The main limitations are class imbalance, poor performance on new drugs, limited explainability, and the need for additional data sources.