We present a new approach to large-scale graph mining based on so-called backbone refinement classes. The method efficiently mines tree-shaped subgraph descriptors under minimum frequency and significance constraints, using classes of fragments to reduce feature set size and running times. The classes are defined in terms of fragments sharing a common backbone. The method is able to optimize structural inter-feature entropy as opposed to occurrences, which is characteristic for open or closed fragment mining. In the experiments, the proposed method reduces feature set sizes by >90 % and >30 % compared to complete tree mining and open tree mining, respectively. Evaluation using crossvalidation runs shows that their classification accuracy is similar to the complete set of trees but significantly better than that of open trees. Compared to open or closed fragment mining, a large part of the search space can be pruned due to an improved statistical constraint (dynamic upper bound adjustment), which is also confirmed in the experiments in lower running times compared to ordinary (static) upper bound pruning. Further analysis using large-scale datasets yields insight into important properties of the proposed descriptors, such as the dataset coverage and the class size represented by each descriptor. A final cross-validation run confirms that the novel descriptors render large training sets feasible which previously might have been intractable.

Abstract Motivation: The Predictive Toxicology Challenge (PTC) was initiated to stimulate the development of advanced techniques for predictive toxicology models. The goal of this challenge was to compare different approaches for the prediction of rodent carcinogenicity, based on the experimental results of the US National Toxicology Program (NTP). Results: 111 sets of predictions for 185 compounds have been evaluated on quantitative and qualitative scales to select the most predictive models and those with the highest toxicological relevance. The accuracy of the submitted predictions was between 25 and 79 %. An evaluation of the most accurate models by toxicological experts showed, that it is still hard for domain experts to interpret the submitted models and to put them into relation with toxicological knowledge. Availability: PTC details and data can be found at: http://www.predictive-toxicology.org/ptc/ Contact: [email protected] * To whom correspondence should be addressed. † Present address: Institute for Computer Science, Technical University Munich, D-85748, Garching, Germany.

MOTIVATION: The development of in silico models to predict chemical carcinogenesis from molecular structure would help greatly to prevent environmentally caused cancers. The Predictive Toxicology Challenge (PTC) competition was organized to test the state-of-the-art in applying machine learning to form such predictive models. RESULTS: Fourteen machine learning groups generated 111 models. The use of Receiver Operating Characteristic (ROC) space allowed the models to be uniformly compared regardless of the error cost function. We developed a statistical method to test if a model performs significantly better than random in ROC space. Using this test as criteria five models performed better than random guessing at a significance level p of 0.05 (not corrected for multiple testing). Statistically the best predictor was the Viniti model for female mice, with p value below 0.002. The toxicologically most interesting models were Leuven2 for male mice, and Kwansei for female rats. These models performed well in the statistical analysis and they are in the middle of ROC space, i.e. distant from extreme cost assumptions. These predictive models were also independently judged by domain experts to be among the three most interesting, and are believed to include a small but significant amount of empirically learned toxicological knowledge. AVAILABILITY: PTC details and data can be found at: http://www.predictive-toxicology.org/ptc/.

We present the application of Feature Mining techniques to the Developmental Therapeutics Program's AIDS antiviral screen database. The database consists of 43576 compounds, which were measured for their capability to protect human cells from HIV-1 infection. According to these measurements, the compounds were classified as either active, moderately active or inactive. The distribution of classes is extremely skewed: Only 1.3 % of the molecules is known to be active, and 2.7 % is known to be moderately active. Given this database, we were interested in molecular substructures (i.e., features) that are frequent in the active molecules, and infrequent in the inactives. In data mining terms, we focused on features with a minimum support in active compounds and a maximum support in inactive compounds. We analyzed the database using the levelwise version space algorithm that forms the basis of the inductive query and database system MOLFEA (Molecular Feature Miner). Within this framework, it is possible to declaratively specify the features of interest, such as the frequency of features on (possibly different) datasets as well as on the generality and syntax of them. Assuming that the detected substructures are causally related to biochemical mechanisms, it should be possible to facilitate the development of new pharmaceuticals with improved activities.

Abstract Summary: We initiated the Predictive Toxicology Challenge (PTC) to stimulate the development of advanced SAR techniques for predictive toxicology models. The goal of this challenge is to predict the rodent carcinogenicity of new compounds based on the experimental results of the US National Toxicology Program (NTP). Submissions will be evaluated on quantitative and qualitative scales to select the most predictive models and those with the highest toxicological relevance. Availability: http://www.informatik.uni-freiburg.de/~ml/ptc/ Contact: [email protected] * To whom correspondence should be addressed.