Machine learning has become increasingly important in drug discovery where viable molecular structures are searched or designed for therapeutic efficacy. In particular, the costly pre-clinical in vitro and in vivo testing of drug candidates can be focused to the most promising molecules, if accurate in silico models are available [7]. During the last decade kernel methods [3, 7, 2, 1, 10] have emerged as an effective way for modelling the activity of candidate drug molecules. However, classification methods focusing on a single target variable at a time are not optimally suited to drug screening applications where a large number of target cell lines are to be handled. In this paper we propose, to our knowledge, the first multilabel learning approach for molecular classification. Our method belongs to the structured output prediction family [6, 8, 4, 5], where graphical models and kernels have been successfully married in recent years. In our approach, the drug targets (cancer cell lines) are organized in a network, drug molecules are represented by kernels and discriminative max-margin training is used to learn the parameters. We demonstrate the benefits of the multilabel classification approach on a dataset of 60 cancer cell lines and 4554 candidate molecules.

Juho Rousu @VideoLectures.net