We apply our methods to different benchmarks. In each case we evaluate the success of clustering the data in the selected feature spaces, by measuring Jaccard scores with respect to known classifications. We demonstrate that feature filtering according to CE outperforms the variance method and gene-shaving. There are cases where the analysis, based on a small set of selected features, outperforms the best score reported when all information was used. Our method calls for an optimal size of the relevant feature set. This turns out to be just a few percents of the number of genes in the two Leukemia datasets that we have analyzed. Moreover, the most favored selected genes turn out to have significant GO enrichment in relevant cellular processes.

We find that the UFF quality degrades smoothly with information loss. It remains successful even under substantial damage. Our method allows for selection of a best imputation method on a dataset treated by UFF. More importantly, scoring features according to their stability under information loss is shown to be correlated with biological importance in cancer studies. This scoring may lead to novel biological insights.

BackgroundA hierarchy, characterized by tree-like relationships, is a natural method of organizing data in various domains. When considering an unsupervised machine learning routine, such as clustering, a bottom-up hierarchical (BU, agglomerative) algorithm is used as a default and is often the only method applied.Methodology/Principal FindingsWe show that hierarchical clustering that involve global considerations, such as top-down (TD, divisive), or glocal (global-local) algorithms are better suited to reveal meaningful patterns in the data. This is demonstrated, by testing the correspondence between the results of several algorithms (TD, glocal and BU) and the correct annotations provided by experts. The correspondence was tested in multiple domains including gene expression experiments, stock trade records and functional protein families. The performance of each of the algorithms is evaluated by statistical criteria that are assigned to clusters (nodes of the hierarchy tree) based on expert-labeled data. Whereas TD algorithms perform better on global patterns, BU algorithms perform well and are advantageous when finer granularity of the data is sought. In addition, a novel TD algorithm that is based on genuine density of the data points is presented and is shown to outperform other divisive and agglomerative methods. Application of the algorithm to more than 500 protein sequences belonging to ion-channels illustrates the potential of the method for inferring overlooked functional annotations. ClustTree, a graphical Matlab toolbox for applying various hierarchical clustering algorithms and testing their quality is made available.ConclusionsAlthough currently rarely used, global approaches, in particular, TD or glocal algorithms, should be considered in the exploratory process of clustering. In general, applying unsupervised clustering methods can leverage the quality of manually-created mapping of proteins families. As demonstrated, it can also provide insights in erroneous and missed annotations.

No abstract

Abstract. Clustering algorithms are employed in many bioinformatics tasks, including categorization of protein sequences and analysis of gene-expression data. Although these algorithms are routinely applied, many of them suffer from the following limitations: (i) relying on predetermined parameters tuning, such as a-priori knowledge regarding the number of clusters; (ii) involving nondeterministic procedures that yield inconsistent outcomes. Thus, a framework that addresses these shortcomings is desirable. We provide a datadriven framework that includes two interrelated steps. The first one is SVDbased dimension reduction and the second is an automated tuning of the algorithm's parameter(s). The dimension reduction step is efficiently adjusted for very large datasets. The optimal parameter setting is identified according to the internal evaluation criterion known as Bayesian Information Criterion (BIC). This framework can incorporate most clustering algorithms and improve their performance. In this study we illustrate the effectiveness of this platform by incorporating the standard K-Means and the Quantum Clustering algorithms. The implementations are applied to several gene-expression benchmarks with significant success.