In this paper, we develop a new algorithm to construct Multiple and Global Alignments (MGA) of primary structures, i.e., strings coding biological macromolecules. The construction of such alignments is based on the one of the (longest) Approximate Common Subsequences (ACS), made up by longer approximate substrings appearing, approximately, in the same positions in all the strings. This ACS represents a MGA. Constructing such alignments is a way to find homologies between biological macromolecules. Our algorithm is of complexity O(N 2*L 2*(log(L))2) in computing time, where N is the number of the strings and L is the length of the longest string.
Identification of genes that share common biological functions is one of the main objectives in molecular biology. The present work focuses on developing an appropriate mechanism for discovery of such genesets from microarray data. We introduce a conceptual property "cohesion' among genes as representative of common biological function, under influence of which a geneset behave coherently. Such genesets are marked as 'cohesive'. Depending on "100% cohesion" equivalence relation, the entire set of associated genes is decomposed into a number of disjoint equivalence classes, each with unique behavior. The equivalence classes form several disjoint affinity groups, members within a group having pair-wise direct interaction. Each group may be called a cohesive gene cluster. A data mining technique for cohesive geneset discovery is developed and applied on expression data to discover intracluster gene relationships for extracting natural coherent genesets. Experiments with some cancer datasets discover thousands of long confident patterns within reasonable time, showing its superiority over classical association discovery techniques. Results can provide important insight into molecular biology and biomedical research.
In microarray data analysis, visualizations based on agglomerative clustering results are widely applied to help biomedical researchers in generating a mental model of their data. In order to support a selection of the to-be-applied algorithm and parameterizations, we propose a novel cluster index, the tree index (TI), to evaluate hierarchical cluster results regarding their visual appearance and their accordance to available background information. Visually appealing cluster trees are characterized by splits that separate those homogeneous clusters from the rest of the data, which have low inner cluster variance and share a medical class label. To evaluate clustering trees regarding this property, the TI computes the likeliness of every single split in the cluster tree. Computing TIs for different algorithms and parameterizations allows to identify the most appealing cluster tree among many possible tree visualizations obtained. Application is shown on simulated data as well as on two public available cancer data sets.
Our group has previously used machine learning techniques to develop computational systems to automatically analyse fluorescence microscope images and classify the location of the depicted protein. Based on this work, we developed a system, the Subcellular Location Image Finder (SLIF), which mines images from scientific journals for analysis.For some of the images in journals, the system is able to automatically compute the pixel resolution (the physical space represented by each pixel), by identifying a scale bar and processing the caption text. However, scale bars are not always included. For those images, the pixel resolution is unknown. Blindly feeding these images into the classification pipeline results in unacceptably low accuracy.We first describe methods that minimise the impact of this problem by training resolution-insensitive classifiers. We show that these techniques are of limited use as classifiers can only be made insensitive to resolutions which are similar to each other. We then approach the problem in a different way by trying to estimate the resolution automatically and processing the image based on this prediction. Testing on digitally down-sampled images shows that the combination of these two approaches gives classification results which are essentially as good as if the resolution had been known.
