Sivan Sabato

DNA methylation of distal regulatory sites characterizes dysregulation of cancer genes

BackgroundAbnormal epigenetic marking is well documented in gene promoters of cancer cells, but the study of distal regulatory siteshas lagged behind.We performed a systematic analysis of DNA methylation sites connected with gene expression profilesacross normal and cancerous human genomes.ResultsUtilizing methylation and expression data in 58 cell types, we developed a model for methylation-expression relationships in gene promoters and extrapolated it to the genome. We mapped numerous sites at which DNA methylation was associated with expression of distal genes. These sites bind transcription factors in a methylation-dependent manner, and carry the chromatin marks of a particular class of transcriptional enhancers. In contrast to the traditional model of one enhancer site per cell type, we found that single enhancer sites may define gradients of expression levels across many different cell types. Strikingly, the identified sites were drastically altered in cancers: hypomethylated enhancer sites associated with upregulation of cancer-related genes and hypermethylated sites with downregulation. Moreover, the association between enhancer methylation and gene deregulation in cancerwas significantly stronger than the association of promoter methylationwith gene deregulation.ConclusionsMethylation of distal regulatory sites is closely related to gene expression levels across the genome. Single enhancers may modulate ranges of cell-specific transcription levels, from constantlyopen promoters. In contrast to the remote relationships between promoter methylation and gene dysregulation in cancer, altered methylation of enhancer sites is closely related to gene expression profiles of transformed cells.

Prediction by categorical features: generalization properties and application to feature ranking

We describe and analyze a new approach for feature ranking in the presence of categorical features with a large number of possible values. It shown that popular ranking criteria, such as the Gini index and the misclassification error, can be interpreted as the training error of a predictor that is deduced from the training set. It is then argued that using the generalization error is a more adequate ranking criterion.We propose a modification of the Gini index criterion, based on a robust estimation of the generalization error of a predictor associated with the Gini index. The properties of this new estimator are analyzed, showing that for most training sets, it produces an accurate estimation of the true generalization error. We then address the question of finding the optimal predictor that is based on a single categorical feature. It is shown that the predictor associated with the misclassification error criterion has the minimal expected generalization error. We bound the bias of this predictor with respect to the generalization error of the Bayes optimal predictor, and analyze its concentration properties.

Preprocessing Expression-Based Constraint Satisfaction Problems for Stochastic Local Search

This work presents methods for processing a constraint satisfaction problem (CSP) formulated by an expression-based language, before the CSP is presented to a stochastic local search solver. The architecture we use to implement the methods allows the extension of the expression language by user-defined operators, while still benefiting from the processing methods. Results from various domains, including industrial processor verification problems, show the strength of the methods. As one of our test cases, we introduce the concept of random-expression CSPs as a new form of random CSPs. We believe this form emulates many real-world CSPs more closely than other forms of random CSPs. We also observe a satisfiability phase transition in this type of problem ensemble.

Submodular Learning and Covering with Response-Dependent Costs

We consider interactive learning and covering problems, in a setting where actions may incur different costs, depending on the response to the action. We propose a natural greedy algorithm for response-dependent costs. We bound the approximation factor of this greedy algorithm in active learning settings as well as in the general setting. We show that a different property of the cost function controls the approximation factor in each of these scenarios. We further show that in both settings, the approximation factor of this greedy algorithm is near-optimal among all greedy algorithms. Experiments demonstrate the advantages of the proposed algorithm in the response-dependent cost setting.