Biomarker discovery by feature ranking: Evaluation on a case study of embryonal tumors.

The task of biomarker discovery is best translated to the machine learning task of feature ranking. Namely, the goal of biomarker discovery is to identify a set of potentially viable targets for addressing a given biological status. This is aligned with the definition of feature ranking and its goal - to produce a list of features ordered by their importance for the target concept. This differs from the task of feature selection (typically used for biomarker discovery) in that it catches viable biomarkers that have redundant or overlapping information with often highly important biomarkers, while with feature selection this is not the case. We propose to use a methodology for evaluating feature rankings to assess the quality of a given feature ranking and to discover the best cut-off point. We demonstrate the effectiveness of the proposed methodology on 10 datasets containing data about embryonal tumors. We evaluate two most commonly used feature ranking algorithms (Random forests and RReliefF) and using the evaluation methodology identifies a set of viable biomarkers that have been confirmed to be related to cancer.

Error curves for evaluating the quality of feature rankings.

In this article, we propose a method for evaluating feature ranking algorithms. A feature ranking algorithm estimates the importance of descriptive features when predicting the target variable, and the proposed method evaluates the correctness of these importance values by computing the error measures of two chains of predictive models. The models in the first chain are built on nested sets of top-ranked features, while the models in the other chain are built on nested sets of bottom ranked features. We investigate which predictive models are appropriate for building these chains, showing empirically that the proposed method gives meaningful results and can detect differences in feature ranking quality. This is first demonstrated on synthetic data, and then on several real-world classification benchmark problems.

Neuroblastoma tumorigenesis is regulated through the Nm23-H1/h-Prune C-terminal interaction.

Nm23-H1 is one of the most interesting candidate genes for a relevant role in Neuroblastoma pathogenesis. H-Prune is the most characterized Nm23-H1 binding partner, and its overexpression has been shown in different human cancers. Our study focuses on the role of the Nm23-H1/h-Prune protein complex in Neuroblastoma. Using NMR spectroscopy, we performed a conformational analysis of the h-Prune C-terminal to identify the amino acids involved in the interaction with Nm23-H1. We developed a competitive permeable peptide (CPP) to impair the formation of the Nm23-H1/h-Prune complex and demonstrated that CPP causes impairment of cell motility, substantial impairment of tumor growth and metastases formation. Meta-analysis performed on three Neuroblastoma cohorts showed Nm23-H1 as the gene highly associated to Neuroblastoma aggressiveness. We also identified two other proteins (PTPRA and TRIM22) with expression levels significantly affected by CPP. These data suggest a new avenue for potential clinical application of CPP in Neuroblastoma treatment.

Finding explained groups of time-course gene expression profiles with predictive clustering trees.

In biology, analyzing time course data is usually a two-step process, beginning with clustering of similar temporal profiles. After the initial clustering, depending on the expert's knowledge, descriptions of the clusters are elucidated (e.g., Gene Ontology terms that are enriched in the clusters). In this paper, we investigate the application of so-called predictive clustering trees (PCTs) for the analysis of time series data. PCTs are a part of a more general framework of predictive clustering, which unifies clustering and prediction. Their advantage over usual clustering approaches is that they partition the time course data into homogeneous clusters while at the same time providing symbolic descriptions of the clusters. We evaluate our approach on multiple yeast microarray time series datasets. Each dataset records the change over time in the expression level of yeast genes as a response to a specific change in environmental conditions. We demonstrate that PCTs are able to cluster genes with similar temporal profiles, yield a predictive model of the temporal profiles of genes based on a cluster prototype, and provide cluster descriptions, all in a single step.