Abstract: We present a new method for drug bioactivity classification based on learning an ensemble of multi-task classifiers. As the base classifiers of the ensemble we use Maximum Margin Conditional Random Field (MMCRF) odels, which have previously obtained the state-of-the-art accuracy in this problem. MMCRF relies on a network structure coupling the set of tasks together, and thus turns the multi-task learning problem into a graph labeling problem. We study different ways of obtaining the graph connecting the tasks and how it affects the predictive accuracy of the model. We compare different ways to extract that graph from the correlation matrix given by auxiliary data, such as maximum weight spanning tree, correlation thresholding and sparse inverse covariance learning. In addition, we study methods based on random networks, namely random pairing and random spanning tree of the tasks. We experiment with NCI-Cancer data containing the cancer inhibition potential of drug-like moleculesagainst 60 cancer cell lines. In our experiments we find that graphlabeling based on auxiliary data perform better than random graphs when the predictor is a single MMCRF. However, MMCRF ensembles based on random graphs behave as well or better than methods based on auxiliary data.

Computational Systems Biology and Bioinformatics Research Group