A Neural Collaborative Filtering Model for Adverse Drug Reaction Prediction

As adverse drug reactions (ADRs) can cause serious consequences to medication treatment, identifying and predicting adverse drug reactions (ADRs) play a crucial role in drug effects and drug use safety. In this paper, we reviewed existing machine learning methods in ADR prediction and proposed a neural collaborative filtering model (NCF) for the prediction of monopharmacy ADRs. NCF is based on dimension reduction by matrix factorization (MF) and combined with an MLP model to enhance prediction performance by handling complex non-linear relationships using MLN. Our NCF model optimizes the MF model by replacing dot product in the MF model with MLP, integrating drug and ADR embeddings from monopharmacy ADR benchmark data and drug-related chemical, physical and biological descriptors in a two-layer deep neural network. The model was tested by using 10-fold cross-validation. On 10-fold cross-validation, the resulting AUC, AUPR, and running time indicated a better performance and higher efficiency on drug-ADR prediction achieved by applying our NCF model.