Catálogo de publicaciones - libros

A new gene selection method for analyzing microarray experiments pertaining to two classes of tissues and for determining relevant genes characterizing differences between the two classes is proposed. The new technique is based on Switching Neural Networks (SNN), learning machines that assign a relevance value to each input variable, and adopts Recursive Feature Addition (RFA) for performing gene selection. The performances of SNN-RFA are evaluated by considering its application on two real and two artificial gene expression datasets generated according to a proper mathematical model that possesses biological and statistical plausibility. Comparisons with other two widely used gene selection methods are also shown.

Analysis and visualization of high-dimensional clinical proteomic spectra obtained from mass spectrometric measurements is a complicated issue. We present a wavelet based preprocessing combined with an unsupervised and supervised analysis by Self-Organizing Maps and a fuzzy variant thereof. This leads to an optimal encoding and a robust classifier incorporating the possibility of fuzzy labels.

In analyzing gene expression data from micro-array, a major challenge is the definition of a feature selection criterion to judge the goodness of a subset of features with respect to a particular classification model. This paper presents a cost-sensitive approach feature selection that focuses on two fundamental requirements: (1) the quality of the features in order to promote the classifier accuracy and (2) the cost of computation due to the complexity that occurs during training and testing the classifier. The paper describes the approach in detail and includes a case study for a publicly available micro-array dataset. Results show that the proposed process yields state-of-art performance and uses only a small fraction of features that are generally used in competitive approaches on the same dataset.

Many real-world classification problems involve very sparse and high-dimensional data. The successes of LIKNON - linear programming support vector machine (LPSVM) for feature selection, motivates a more thorough analysis of the method when applied to sparse, multivariate data. Due to the sparseness, the selection of a classification model is greatly influenced by the characteristics of that particular dataset. Robust feature/model selection methods are desirable. LIKNON is claimed to have such robustness properties. Its feature selection operates by selecting the groups of features with large differences between the resultants of the two classes. The degree of desired difference is controlled by the regularization parameter. We study the practical value of LIKNON-based feature/model selection for microarray data. Our findings support the claims about the robustness of the method.

Prediction of Alternative Splicing has been traditionally based on expressed sequences’ study, helped by homology considerations and the analysis of local discriminative features. More recently, Machine Learning algorithms have been developed that try avoid the use of a priori information, with partial success. Here we approach the prediction of Alternative Splicing as a time series analysis problem and we show that it is possible to obtain results similar or better than the state of the art without any explicit modeling of homology, positions in the splice site, nor any use of other local features. As a consequence, our method has a better generality and a broader and simpler applicability with respect to previous ones. Results on pre-mRNA sequences in C.Elegans are reported.

In this paper wavelet analysis and Genetic Algorithm (GA) are performed to extract features and reduce dimensionality of mass spectrometry data. A set of wavelet features, which include detail coefficients and approximation coefficients, are extracted from mass spectrometry data. Detail coefficients are used to characterize the localized change of mass spectrometry data and approximation coefficients are used to compress mass spectrometry data, reducing the dimensionality. GA performs the further dimensionality reduction and optimizes the wavelet features. Experiments prove that this hybrid method of feature extraction is efficient way to characterize mass spectrometry data.

The emerging insights into kinase function and evolution combined with a rapidly growing number of crystal structures of protein kinases complexes have facilitated a comprehensive structural bioinformatics analysis of sequence–structure relationships in determining the binding function of protein tyrosine kinases. We have found that evolutionary signal derived solely from the tyrosine kinase sequence conservation can not be readily translated into the ligand binding phenotype. However, fingerprinting ligand–protein interactions using in silico profiling of inhibitor binding against protein tyrosine kinases crystal structures can detect a functionally relevant kinase binding signal and reconcile the existing experimental data. In silico proteomics analysis unravels mechanisms by which structural plasticity of the tyrosine kinases is linked with the conformational preferences of cancer drugs Imatinib and Dasatinib in achieving effective drug binding with a distinct spectrum of the tyrosine kinome. While Imatinib binding is highly sensitive to the activation state of the enzyme, the computed binding profile of Dasatinib is remarkably tolerant to the conformational state of ABL. A comprehensive study of evolutionary, structural, dynamic and energetic aspects of tyrosine kinases binding with clinically important class of inhibitors provides important insights into mechanisms of sequence–structure relationships in the kinome space and molecular basis of functional adaptability towards specific binding.

Reverse engineering of gene regulatory networks is a key issue for functional genomic. Indeed, unraveling complex interactions among genes is a crucial step in order to understand their role in cellular processes. High-throughput technologies such as DNA microarrays or ChIP on chip have in principle opened the door to network inference from data. However the size of available data is still limited compared to their dimension. Machine learning methods have thus to be worked out in order to respond to this challenge. In this work we focused our attention on modeling gene regulatory networks with Bayesian networks. Bayesian networks offer a probabilistic framework for the reconstruction of biological interactions networks using data, but the structure learning problem is still a bottleneck. In this paper, we use evolutionary algorithms to stochastically evolve a set of candidate Bayesian networks structures and find the model that best explains the small number of available observational data. We propose different kinds of recombination strategies and an appropriate technique of niching that ensure diversity among candidate solutions. Tests are carried out on simulated data drawn from a biorealistic network. The effect of deterministic crowding, a niching method, is compared to mutation for different kinds of recombination strategies and is shown to improve significantly the performances. Enhanced by deterministic crowding, our evolutionary approach outperforms K2, Greedy-search and MCMC, for training sets whose size is small compared to the standard in machine learning.

The immune system is a large and complex system still incompletely understood. The synergy of biological knowledge included in data bases, mathematical and computational models and bioinformatics tools can help in gain a deeper understanding. In this paper, we try to understand how one can characterize a mouse in order to find the better vaccination protocol against mammary carcinoma for that individual. This study will be expanded in the framework of ImmunoGrid EC project.

The need to determine phases is a major bottleneck in a fully automated X-ray crystallography pipeline. The problem commonly called phasing can be solved by a computational method called molecular replacement (MR). With the deposition of more and more proteins into the Protein Data Bank (PDB), it has been shown that the MR yields better initial models . In this paper, ab initio first model generation is addressed. A novel scheme using PHASER is proposed which does not require any a priori information about the structure. The input to the system is the target structure factors and the sequence. We created a unique set of supersecondary structure (fragment) dataset and used them in creation of the first model. The method was evaluated with log-likelihood gain ( LLG ) and translational Z -score ( TFZ ) as defined by PHASER . The results obtained are highly encouraging with translation Z -scores of 7 and above for the first model. The proposed scheme is tested on six proteins, two each from α , β and α  +  β classes with very good results.