Pilar Bulacio

Sparse and stable gene selection with consensus SVM-RFE

A method is described for performing sparse and stable gene selection from a number of unstable, but low cost, SVM-RFE units referred to as SVM-RFE subunits. Using a comprehensive simulation study, we show that the introduction of a consensus constraint with respect to variations in the policy of gene removal and a stability constraint with respect to perturbations in the training data can remarkably improve gene selection precision, dimensionality reduction ratio and stability of low cost SVM-RFE subunits still guaranteeing affordable computational costs. The method, which does not require the preselection of the number of selected genes, is divided into two stages. Multiple rough gene removal policies are first applied to multiple surrogate training datasets (spreading). Multiple consensus gene sets with respect to variations in the gene removal policy are then obtained and passed through a stability filter which selects the best performing gene set (despreading). Hence, while the consensus constraint performs strong dimensionality reduction at affordable computational costs, the stability constraint ensures acceptable indexes of gene selection stability and further dimensionality reduction. The method is validated on three benchmark microarray datasets.

Full Length Article: Revised HLMS: A useful algorithm for fuzzy measure identification

An important limitation of fuzzy integrals for information fusion is the exponential growth of coefficients for an increasing number of information sources. To overcome this problem a variety of fuzzy measure identification algorithms has been proposed. HLMS is a simple gradient-based algorithm for fuzzy measure identification which suffers from some convergence problems. In this paper, two proposals for HLMS convergence improvement are presented, a modified formula for coefficients update and new policy for monotonicity check. A comprehensive experimental work shows that these proposals indeed contribute to HLMS convergence, accuracy and robustness.

k-maxitive fuzzy measures: A scalable approach to model interactions

Abstract Fuzzy measures are powerful at modeling interactions between elements. Unfortunately, they use a number of coefficients that exponentially grows with the number of elements. Beyond the computational complexity, assigning a value to any coalition of a large set of elements does not make sense. k-order measures model interactions involving at most k elements. The number of coefficients to identify is reduced and their modeling capacity is preserved in real problems where the number of interacting elements is limited. In extreme situations of full redundancy or complementariness, it is mathematically proven that the complete fuzzy measure is both k-additive and k-maxitive. A learning algorithm to identify k-maxitive measures from labeled data is designed on the basis of HLMS (Heuristic Least Mean Squares). In a classification context, the study of synthetic data with partial redundancy or complementariness supports the idea that the difference between full and partial interaction is a matter of degree, not of kind. Dealing with two real world datasets, a comparison of the complete fuzzy measure and a k-maxitive one shows the number of interacting elements is limited and the k-maxitive measures yield the same characterization of interactions and a comparable classification accuracy.

Set characterization-selection towards classification based on interaction index

In many real world datasets both the individual and coordinated action of features may be relevant for class identification. In this paper, a computational strategy for relevant feature selection based on the characterization of redundant or complementary features is proposed. The characterization is achieved using fuzzy measures and an interaction index computed from fuzzy measure coefficients. Fuzzy measure identification requires raw data to be turned into confidence degrees. This key step is carried out considering the distributions of feature values across all the classes. Fuzzy measure coefficients are then estimated with an improved version of the Heuristic Least Mean Squares algorithm that includes an efficient management of untouched coefficients. Then, a generalization of the Shapley index for an arbitrary number of features is used. Simulations experiments on synthetic datasets are performed to study the behavior of this generalized interaction index. For extreme datasets, containing either redundant or complementary features as well as noise, the index value is defined by mathematical formula. This result is used to motivate feature selection guidelines that take into account feature interactions. Experimental results on benchmark datasets show that the proposal allows for the design of compact, interpretable and competitive classification models.

DNA Barcoding through Quaternary LDPC Codes

For many parallel applications of Next-Generation Sequencing (NGS) technologies short barcodes able to accurately multiplex a large number of samples are demanded. To address these competitive requirements, the use of error-correcting codes is advised. Current barcoding systems are mostly built from short random error-correcting codes, a feature that strongly limits their multiplexing accuracy and experimental scalability. To overcome these problems on sequencing systems impaired by mismatch errors, the alternative use of binary BCH and pseudo-quaternary Hamming codes has been proposed. However, these codes either fail to provide a fine-scale with regard to size of barcodes (BCH) or have intrinsic poor error correcting abilities (Hamming). Here, the design of barcodes from shortened binary BCH codes and quaternary Low Density Parity Check (LDPC) codes is introduced. Simulation results show that although accurate barcoding systems of high multiplexing capacity can be obtained with any of these codes, using quaternary LDPC codes may be particularly advantageous due to the lower rates of read losses and undetected sample misidentification errors. Even at mismatch error rates of 10−2 per base, 24-nt LDPC barcodes can be used to multiplex roughly 2000 samples with a sample misidentification error rate in the order of 10−9 at the expense of a rate of read losses just in the order of 10−6.

Consistent prediction of GO protein localization

The GO-Cellular Component (GO-CC) ontology provides a controlled vocabulary for the consistent description of the subcellular compartments or macromolecular complexes where proteins may act. Current machine learning-based methods used for the automated GO-CC annotation of proteins suffer from the inconsistency of individual GO-CC term predictions. Here, we present FGGA-CC+, a class of hierarchical graph-based classifiers for the consistent GO-CC annotation of protein coding genes at the subcellular compartment or macromolecular complex levels. Aiming to boost the accuracy of GO-CC predictions, we make use of the protein localization knowledge in the GO-Biological Process (GO-BP) annotations to boost the accuracy of GO-CC prediction. As a result, FGGA-CC+ classifiers are built from annotation data in both the GO-CC and GO-BP ontologies. Due to their graph-based design, FGGA-CC+ classifiers are fully interpretable and their predictions amenable to expert analysis. Promising results on protein annotation data from five model organisms were obtained. Additionally, successful validation results in the annotation of a challenging subset of tandem duplicated genes in the tomato non-model organism were accomplished. Overall, these results suggest that FGGA-CC+ classifiers can indeed be useful for satisfying the huge demand of GO-CC annotation arising from ubiquitous high throughout sequencing and proteomic projects.

Formalization of Gene Ontology relationships with factor graph towards Biological Process prediction

Gene Ontology is a hierarchical controlled vocabulary for protein annotation. Its synergy with automatic classification methods, ensemble, has been widely used for the prediction of protein functions. Current classification methods use only the relation is_a and a few little part_of to generate prediction model. In this work we formalize the GO part_of, regulates; negatively_regulates and positively_regulates relationships through predicate logic. This formalization is incorporated within an ensemble method based on graph factor called Factor Graph GO Annotation. The proposed model is validated against four model organisms for GO Biological Process prediction.

Formal modeling of Gene Ontology annotation predictions based on factor graphs

Fil: Spetale, Flavio Ezequiel. Universidad Nacional de Rosario. Facultad de Ciencias Exactas, Ingenieria y Agrimensura; Argentina. Consejo Nacional de Investigaciones Cientificas y Tecnicas. Centro Cientifico Tecnologico Conicet - Rosario. Centro Internacional Franco Argentino de Ciencias de la Informacion y de Sistemas. Universidad Nacional de Rosario. Centro Internacional Franco Argentino de Ciencias de la Informacion y de Sistemas; Argentina

Application of Hierarchical Function Prediction in Solanum Lycopersicum

Predicting functional gene annotations through machine learning techniquesmay focus on the experimental val- idation reducing their cost. The hierarchical prediction method based on True Path Rule provides function results consistent and traceable with the Gene Ontology Molecular Function defi- nition. In this work, a design of a hierarchical prediction model based on True Path Rule for plants is presented. The train- ing stage is done with Arabidopsis thaliana data characterized with sequence domains and physicochemical properties feeding an ensemble of binary classifiers, one classifier for each func- tional class. The proposed model is validated against a set of well-known control sequences and with a set of sequences of S. lycopersicum without any annotation by biological experts. The discussed results are promising; the proposal can be enriched withmore organisms and with diverse sources of sequence char- acterizations.