P-glycoprotein (P-gp, MDR1) is a promiscuous drug efflux pump of substantial pharmacological importance. Taking advantage of large-scale cytotoxicity screening data involving 60 cancer cell lines, we correlated the differential biological activities of ∼13,000 compounds against cellular P-gp levels. We created a large set of 934 high-confidence P-gp substrates or nonsubstrates by enforcing agreement with an orthogonal criterion involving P-gp overexpressing ADR-RES cells. A support vector machine (SVM) was 86.7% accurate in discriminating P-gp substrates on independent test data, exceeding previous models. Two molecular features had an overarching influence: nearly all P-gp substrates were large (>35 atoms including H) and dense (specific volume of <7.3 Å(3)/atom) molecules. Seven other descriptors and 24 molecular fragments ("effluxophores") were found enriched in the (non)substrates and incorporated into interpretable rule-based models. Biological experiments on an independent P-gp overexpressing cell line, the vincristine-resistant VK2, allowed us to reclassify six compounds previously annotated as substrates, validating our method's predictive ability. Models are freely available at http://pgp.biozyne.com .

In many real-life problems, obtaining labelled data can be a very expensive and laborious task, while unlabeled data can be abundant. The availability of labeled data can seriously limit the performance of supervised learning methods. Here, we propose a semi-supervised classification tree induction algorithm that can exploit both the labelled and unlabeled data, while preserving all of the appealing characteristics of standard supervised decision trees: being non-parametric, efficient, having good predictive performance and producing readily interpretable models. Moreover, we further improve their predictive performance by using them as base predictive models in random forests. We performed an extensive empirical evaluation on 12 binary and 12 multi-class classification datasets. The results showed that the proposed methods improve the predictive performance of their supervised counterparts. Moreover, we show that, in cases with limited availability of labeled data, the semi-supervised decision trees often yield models that are smaller and easier to interpret than supervised decision trees

We address the task of hierarchical multi-label classification (HMC). HMC is a task of structured output prediction where the classes are organized into a hierarchy and an instance may belong to multiple classes. In many problems, such as gene function prediction or prediction of ecological community structure, classes inherently follow these constraints. The potential for application of HMC was recognized by many researchers and several such methods were proposed and demonstrated to achieve good predictive performances in the past. However, there is no clear understanding when is favorable to consider such relationships (hierarchical and multi-label) among classes, and when this presents unnecessary burden for classification methods. To this end, we perform a detailed comparative study over 8 datasets that have HMC properties. We investigate two important influences in HMC: the multiple labels per example and the information about the hierarchy. More specifically, we consider four machine learning tasks: multi-label classification, hierarchical multi-label classification, single-label classification and hierarchical single-label classification. To construct the predictive models, we use predictive clustering trees (a generalized form of decision trees), which are able to tackle each of the modelling tasks listed. Moreover, we investigate whether the influence of the hierarchy and the multiple labels carries over for ensemble models. For each of the tasks, we construct a single tree and two ensembles (random forest and bagging). The results reveal that the hierarchy and the multiple labels do help to obtain a better single tree model, while this is not preserved for the ensemble models.