Catálogo de publicaciones - libros

When applied to supervised classification problems, dataset complexity determines how difficult a given dataset to classify. Since complexity is a nontrivial issue, it is typically defined by a number of measures. In this paper, we explore complexity of three gene expression datasets used for two-class cancer classification. We demonstrate that estimating the dataset complexity before performing actual classification may provide a hint whether to apply a single best nearest neighbour classifier or an ensemble of nearest neighbour classifiers.

The problem of secondary structure prediction can be formulated as a pattern classification problem and methods from statistics and machine learning are suitable. This paper proposes a new combination approach between Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) by typical sample extraction based on a UBM/GMM system for SVM in protein secondary structure prediction. Our hybrid model achieved a good performance of three-state overall per residue accuracy Q _3 = 77.6% which is comparable to the best techniques available.

Protein Structure Prediction (PSP) aims to reconstruct the 3D structure of a given protein starting from its primary structure (chain of amino acids). It is a well known fact that the 3D structure of a protein only depends on its primary structure. PSP is one of the most important and still unsolved problems in computational biology. Protein Structure Selection (PSS), instead of reconstructing a 3D model for the given chain, aims to select among a given, possibly large, number of 3D structures (called decoys) those that are closer (according to a given notion of distance) to the original (unknown) one. Each decoy is represented by a set of points in ℝ^3. Existing methods for solving PSS make use of suitably defined energy functions which heavily rely on the primary structure of the protein and on protein chemistry. In this paper we present a completely different approach to PSS which does not take advantage at all of the knowledge of the primary structure of the protein but only relies on the graph theoretic properties of the decoys graphs (vertices represent amino acids and edges represent pairs of amino acids whose euclidean distance is less than or equal to a fixed threshold). Even if our methods only rely on approximate geometric information, experimental results show that some of the graph properties we adopt score similarly to energy-based filtering functions in selecting the best decoys. Our results show the principal role of geometric information in PSS, setting a new starting point and filtering method, for existing energy function-based techniques.

For a quantitative analysis of differential protein expression, one has to overcome the problem of aligning time series of measurements from liquid chromatography coupled to mass spectrometry. When repeating experiments one typically observes that the time axis is deformed in a non-linear way. In this paper we propose a technique to align the time series based on generalized canonical correlation analysis (GCCA) for multiple datasets. The monotonicity constraint in time series alignment is incorporated in the GCCA algorithm. The alignment function is learned both in a supervised and a semi-supervised fashion. We compare our approach with previously published methods for aligning mass spectrometry data on a large proteomics dataset.

In this paper we extend kernel functions defined on generative models to embed phylogenetic information into a discriminative learning approach. We describe three generative tree kernels, a Fisher kernel, a sufficient statistics kernel and a probability product kernel, whose key features are the adaptivity to the input domain and the ability to deal with structured data. In particular, kernel adaptivity is obtained through the estimation of a tree structured model of evolution starting from the phylogenetic profiles encoding the presence or absence of specific proteins in a set of fully sequenced genomes. We report preliminary results obtained by these kernels in the prediction of the functional class of the proteins of S. Cervisae , together with comparisons to a standard vector based kernel and to a non-adaptive tree kernel function.

In this paper we describe the state of the art of the semi-automatic and automatic techniques for liver volume extraction from abdominal CT. In the recent years this research focus has gained a lot of importance in the field of medical image processing since it is the first and fundamental step of any automated technique for the automatic liver disease diagnosis, liver volume measurement, and 3D liver volume rendering from CT images.

We introduce a new clustering method for DNA microarray data that is based on space filling curves and wavelet denoising. The proposed method is much faster than the established fuzzy c-means clustering because clustering occurs in one dimension and it clusters cells that contain data, instead of data themselves. Moreover, preliminary evaluation results on data sets from Small Round Blue-Cell tumors, Leukemia and Lung cancer microarray experiments show that it can be equally or more accurate than fuzzy c-means clustering or a gaussian mixture model.

Two major problems related the unsupervised analysis of gene expression data are represented by the accuracy and reliability of the discovered clusters, and by the biological fact that classes of examples or classes of functionally related genes are sometimes not clearly defined. To face these items, we propose a fuzzy ensemble clustering approach to both improve the accuracy of clustering results and to take into account the inherent fuzziness of biological and bio-medical gene expression data. Preliminary results with DNA microarray data of lymphoma and adenocarcinoma patients show the effectiveness of the proposed approach.

Comparative genomics techniques are a powerful tool for the identification of conserved functional genomic regions, but have to be coupled with methods able to assign a functional role to the regions identified. Several methods for the characterization of conserved regions have been proposed but, to our knowledge, signal processing approaches have not been applied yet in this context, despite the proven usefulness of this technique in experiments performed at single genome level. In this article we introduce the use of signal processing in comparative genomics, presenting a method for rapid classification of genomic conserved sequences as protein coding or non coding.

In this work we describe a clustering and feature selection technique applied to the analysis of international dietary profiles. An asymmetric entropy-based measure for assessing the similarity between two clusterizations, also taking into account subclustering relationships, is at the core of the technique, together with PCA. Then, a feature analysis of the dataset with respect to its hierarchical clusterization is performed. This way, most significant features of the dataset are found and a deep understanding of the data distribution is made possible.