Improving prediction accuracy of drug activities by utilising unlabelled instances with feature selection

Molecular activities can be predicted by Quantitative Structure Activity Relationship (QSAR). Because of the high cost of experiments, the number of drug molecules with known activity is much less than that of unknown, to predict molecular activities utilising unlabelled instances will be an interesting issue. Here, Semi-Supervised Learning (SSL) is introduced and a SSL method, Co-Training is investigated on predicting drug activities utilising unlabelled instances. At the same time, a novel algorithm called FESCOT is proposed, which applies feature selection to remove redundant and irrelevant features for Co-Training. Numerical experimental results show that Co-Training and feature selection helps to improve the prediction ability of Co-Training.