Side Information in Drug–Target Interaction Prediction

Abstract

Artificial intelligence constantly offers new ways to enhance drug discovery. The computing power of advanced models helps in predicting potential interactions of drugs and targets non-experimentally and on a large scale. This thesis presents a drug–target interaction prediction model that utilises binary identifiers of drugs and targets. Side information is also added to the model to determine whether it improves the model's prediction performance compared to using only binary identifiers. The model is based on a supervised learning method, factorisation machine, that captures and predicts the pairwise interactions of drugs and targets. Side information refers to additional data that describes the chemical properties of drug and target molecules. The side information is obtained from external data, whereas binary identifiers and the interaction labels are derived from interaction matrix describing known interactions of drugs and targets. The proposed model tries to find patterns that map drug–target data to labels. The drug–target data has three feature options: using only binary identifiers, using only side information, and using both binary identifiers and side information. The results indicate that the proposed model's performance, measured with C-index, is relatively high, and the addition of the side information does not systematically enhance the predictions. Enhancement caused by side information can be seen only in some datasets. Potential model improvements and future work possibilities are also discussed.