Search

The use of comparative genomics for functional, evolutionary, and epidemiological studies requires methods to classify gene families in terms of occurrence in a given species. These methods usually lack multivariate statistical models to infer the partitions and the optimal number of classes and don’t account for genome organization. We introduce a graph structure to model pangenomes in which nodes represent gene families and edges represent genomic neighborhood. Our method, named PPanGGOLiN, partitions nodes using an Expectation-Maximization algorithm based on multivariate Bernoulli Mixture Model coupled with a Markov Random Field. This approach takes into account the topology of the graph and the presence/absence of genes in pangenomes to classify gene families into persistent, cloud, and one or several shell partitions. By analyzing the partitioned pangenome graphs of isolate genomes from 439 species and metagenome-assembled genomes from 78 species, we demonstrate that our method is effective in estimating the persistent genome. Interestingly, it shows that the shell genome is a key element to understand genome dynamics, presumably because it reflects how genes present at intermediate frequencies drive adaptation of species, and its proportion in genomes is independent of genome size. The graph-based approach proposed by PPanGGOLiN is useful to depict the overall genomic diversity of thousands of strains in a compact structure and provides an effective basis for very large scale comparative genomics. The software is freely available at https://github.com/labgem/PPanGGOLiN., Author summary Microorganisms have the greatest biodiversity and evolutionary history on earth. At the genomic level, it is reflected by a highly variable gene content even among organisms from the same species which explains the ability of microbes to be pathogenic or to grow in specific environments. We developed a new method called PPanGGOLiN which accurately represents the genomic diversity of a species (i.e. its pangenome) using a compact graph structure. Based on this pangenome graph, we classify genes by a statistical method according to their occurrence in the genomes. This method allowed us to build pangenomes even for uncultivated species at an unprecedented scale. We applied our method on all available genomes in databanks in order to depict the overall diversity of hundreds of species. Overall, our work enables microbiologists to explore and visualize pangenomes alike a subway map.

International audience; Graphical network inference is used in many fields such as genomics or ecology to infer the conditional independence structure between variables, from measurements of gene expression or species abundances for instance. In many practical cases, not all variables involved in the network have been observed, and the samples are actually drawn from a distribution where some variables have been marginalized out. This challenges the sparsity assumption commonly made in graphical model inference, since marginalization yields locally dense structures, even when the original network is sparse. We present a procedure for inferring Gaussian graphical models when some variables are unobserved, that accounts both for the influence of missing variables and the low density of the original network. Our model is based on the aggregation of spanning trees, and the estimation procedure on the Expectation-Maximization algorithm. We treat the graph structure and the unobserved nodes as missing variables and compute posterior probabilities of edge appearance. To provide a complete methodology, we also propose several model selection criteria to estimate the number of missing nodes. A simulation study and an illustration flow cytometry data reveal that our method has favorable edge detection properties compared to existing graph inference techniques. The methods are implemented in an R package.

We consider the problem of incorporating evolutionary information (e.g. taxonomic or phylogenic trees) in the context of metagenomics differential analysis. Recent results published in the literature propose different ways to leverage the tree structure to increase the detection rate of differentially abundant taxa. Here, we propose instead to use a different hierachical structure, in the form of a correlation-based tree, as it may capture the structure of the data better than the phylogeny. We first show that the correlation tree and the phylogeny are significantly different before turning to the impact of tree choice on detection rates. Using synthetic data, we show that the tree does have an impact: smoothing p-values according to the phylogeny leads to equal or inferior rates as smoothing according to the correlation tree. However, both trees are outperformed by the classical, non hierachical, Benjamini-Hochberg (BH) procedure in terms of detection rates. Other procedures may use the hierachical structure with profit but do not control the False Discovery Rate (FDR) a priori and remain inferior to a classical Benjamini-Hochberg procedure with the same nominal FDR. On real datasets, no hierarchical procedure had significantly higher detection rate that BH. Although intuition advocates the use of a hierachical structure, be it the phylogeny or the correlation tree, to increase the detection rate in microbiome studies, current hierachical procedures are still inferior to non hierachical ones and effective procedures remain to be invented.

MotivationThe availability of large metagenomic data offers great opportunities for the population genomic analysis of uncultured organisms, especially for small eukaryotes that represent an important part of the unexplored biosphere while playing a key ecological role. However, the majority of these species lacks reference genome or transcriptome which constitutes a technical barrier for classical population genomic analyses.ResultsWe introduce the metavariant species (MVS) model, a representation of the species only by intra-species nucleotide polymorphism. We designed a method combining reference-free variant calling, multiple density-based clustering and maximum weighted independent set algorithms to cluster intra-species variant into MVS directly from multisample metagenomic raw reads without reference genome or reads assembly. The frequencies of the MVS variants are then used to compute population genomic statistics such asFSTin order to estimate genomic differentiation between populations and to identify loci under natural selection. The MVSs construction was tested on simulated and real metagenomic data. MVs showed the required quality for robust population genomics and allowed an accurate estimation of genomic differentiation (ΔFST<0.0001 and < 0.03 on simulated and real data respectively). Loci predicted under natural selection on real data were all found by MVSs. MVSs represent a new paradigm that may simplify and enhance holistic approaches for population genomics and evolution of microorganisms.AvailabilityThe method was implemented in a R package,metaVaR.https://github.com/madoui/MetaVaRContactamadoui@genoscope.cns.fr

Allele-specific expression (ASE) is a widely studied molecular mechanism at cell, tissue and organism levels. Here, we extrapolated the concept of ASE to the population-scale (psASE), aggregating ASEs detected at smaller scales. We developed a novel approach to detect psASE based on metagenomic and metatranscriptomic data of environmental samples containing communities of organisms. This approach which measures the deviation between the frequency and the relative expression of biallelic loci, was applied on samples collected during the Tara Oceans expedition (2009-2013), in combination to new Oithona similis transcriptomes, a widespread marine copepod. Among a total of 25,768 single nucleotide variants (SNVs) of O. similis, 587 (2.3%) were targeted by psASE in at least one population. The distribution of SNVs targeted by psASE in different populations is significantly shaped by population genomic differentiation (p-value = 9.3×10−9), supporting a partial genetic control of psASE. To investigate the link between evolution and psASE, loci under selection were compared to loci under psASE. A significant amount of SNVs (0.6%) were targeted by both selection and psASE (p-values < 9.89×10−3), supporting the hypothesis that natural selection and ASE may lead to the same phenotype. Population-scale ASE offers new insights into the gene regulation control in populations and its link with natural selection.

High-throughput RNA-sequencing has become the gold standard method for whole-transcriptome gene expression analysis, and is widely used in numerous applications to study cell and tissue transcriptomes. It is also being increasingly used in a number of clinical applications, including expression profiling for diagnostics and alternative transcript detection. However, despite its many advantages, RNA sequencing can be challenging in some situations, for instance in cases of low input amounts or degraded RNA samples. Several protocols have been proposed to overcome these challenges, and many are available as commercial kits. In this study, we comprehensively test three recent commercial technologies for RNA-seq library preparation (TruSeq, SMARTer and SMARTer Ultra-Low) on human reference tissue preparations, using standard (1μg), low (100 and 10 ng) and ultra-low (< 1 ng) input amounts, and for mRNA and total RNA, stranded or unstranded. The results are analyzed using read quality and alignment metrics, gene detection and differential gene expression metrics. Overall, we show that the TruSeq kit performs well with an input amount of 100 ng, while the SMARTer kit shows degraded performance for inputs of 100 and 10 ng, and the SMARTer Ultra-Low kit performs relatively well for input amounts < 1 ng. All the results are discussed in detail, and we provide guidelines for biologists for the selection of a RNA-seq library preparation kit.

International audience

International audience; Forest trees are long-lived organisms subject to repeated environmental constraints throughout their long lifetimes. They have developed various mechanisms enabling them to cope with fluctuating environmental conditions during their life span, and to survive to current climate change. Epigenetics has recently emerged as a powerful set of mechanisms regulating various developmental processes, plant growth and responses to environmental variations. Such epigenetic mechanisms, which may remain stable along tree life or across generations, constitute a source of rapid phenotypic variations potentially improving adaptation of the plants in situations in which naturally occurring mutations are very rare.In this review, we summarize recent advances in forest tree epigenomics. We first draw the particularities of trees and the available (epi) genomics resources and strategies. Then, we discuss the potential contributions of epigenetics to cope with global climate change and regulate various developmental processes, such as developmental transitions during the annual cycle, phenotypic plasticity in response to environmental variations and stress memory, as well as local adaptation. Finally, we propose some challenges for forest management and highlighted the need to take epigenetics into account in forest tree breeding strategies.

International audience; Motivation: Genome-Wide Association Studies (GWAS) seek to identify causal genomic variants associated with rare human diseases. The classical statistical approach for detecting these variants is based on univariate hypothesis testing, with healthy individuals being tested against affected individuals at each locus. Given that an individual's genotype is characterized by up to one million SNPs, this approach lacks precision, since it may yield a large number of false positives that can lead to erroneous conclusions about genetic associations with the disease. One way to improve the detection of true genetic associations is to reduce the number of hypotheses to be tested by grouping SNPs. Results: We propose a dimension-reduction approach which can be applied in the context of GWAS by making use of the haplotype structure of the human genome. We compare our method with standard univariate and multivariate approaches on both synthetic and real GWAS data, and we show that reducing the dimension of the predictor matrix by aggregating SNPs gives a greater precision in the detection of associations between the phenotype and genomic regions.

Numerous variable selection methods rely on a two-stage procedure, where a sparsity-inducing penalty is used in the first stage to predict the support, which is then conveyed to the second stage for estimation or inference purposes. In this framework, the first stage screens variables to find a set of possibly relevant variables and the second stage operates on this set of candidate variables, to improve estimation accuracy or to assess the uncertainty associated to the selection of variables. We advocate that more information can be conveyed from the first stage to the second one: we use the magnitude of the coefficients estimated in the first stage to define an adaptive penalty that is applied at the second stage. We give two examples of procedures that can benefit from the proposed transfer of information, in estimation and inference problems respectively. Extensive simulations demonstrate that this transfer is particularly efficient when each stage operates on distinct subsamples. This separation plays a crucial role for the computation of calibrated p-values, allowing to control the False Discovery Rate. In this setup, the proposed transfer results in sensitivity gains ranging from 50% to 100% compared to state-of-the-art.

HAL Id : hal-01246419, version 1HAL Id : hal-01246419, version 1; We are interested in problems where two hierarchically structured datasets may be complementary for a learning process. This case may arise in biological applications where genomic and metagenomic analyses may be collected for studying the genomic features of an organism along with its environnement. In this work, we propose a model to assess the relevant interactions between the two datasets. We use a compressed representation that of the original data to cope with the high dimensionality of the problem. We show that the collection of models, characterized through the hierarchical structures, forms a partially ordered set and take advantage of this organization to define a greedy approach to solve the problem more efficiently. Finally, we illustrate the behavior of the resulting algorithm on numerical simulations.

Currently there is much interest in using microarray gene-expression data to form prediction rules for the diagnosis of patient outcomes. A process of gene selection is usually carried out first to find those genes that are most useful according to some criterion for distinguishing between the given classes of tissue samples. However, there is a bias (selection bias) introduced in the estimate of the final version of a prediction rule that has been formed from a smaller subset of the genes that have been selected according to some optimality criterion. In this paper, we focus on the bias that arises when a full data set is not available in the first instance and the prediction rule is formed subsequently by working with the top-ranked genes from the full set. We demonstrate how large the subset of top genes must be before this selection bias is not of practical consequence.

This paper introduces a pattern recognition method specially dedicated to the detection and isolation of after-sale failures on an automotive vehicle. This method is well suited to deal with binary data. It is based on the similarity between an input pattern and some reference patterns characteristic of each failure class. The initial space representation is divided into several subspaces in order to allow classification of patterns which belong simultaneously to several classes. Experiments with real data demonstrate the good performance and the adapted structure of this classification system.

In the context of cancer diagnosis and treatment, we consider the problem of constructing an accurate prediction rule on the basis of a relatively small number of tumor tissue samples of known type containing the expression data on very many (possibly thousands) genes. Recently, results have been presented in the literature suggesting that it is possible to construct a prediction rule from only a few genes such that it has a negligible prediction error rate. However, in these results the test error or the leave-one-out cross-validated error is calculated without allowance for the selection bias. There is no allowance because the rule is either tested on tissue samples that were used in the first instance to select the genes being used in the rule or because the cross-validation of the rule is not external to the selection process; that is, gene selection is not performed in training the rule at each stage of the cross-validation process. We describe how in practice the selection bias can be assessed and corrected for by either performing a cross-validation or applying the bootstrap external to the selection process. We recommend using 10-fold rather than leave-one-out cross-validation, and concerning the bootstrap, we suggest using the so-called .632+ bootstrap error estimate designed to handle overfitted prediction rules. Using two published data sets, we demonstrate that when correction is made for the selection bias, the cross-validated error is no longer zero for a subset of only a few genes.

This paper applies recent developments of mixture models to the prediction of peak episodes of ozone in Lyon (a major French city). Forecasting for the next day should be available at 14:00 h GMT. One day ahead prediction of such events allow public authorities to warn the population of the danger of outdoor exercises and activities. Compared to a standard discrimination problem, the database has many unusual characteristics among which missing measurements and missing decisions. All this peculiarities are well taken into account by mixture models, which additionally allow class densities to have complex shapes and decision boundaries to be non-linear. The proposed approach is compared to the persistence method, which consists in forecasting the same decision as the day before. The results show that mixture models outperform this simple predictor and offer the advantage of properly handling the missing labels and data.

Networks are a commonly used mathematical model to describe the rich set of interactions between objects of interest. Many clustering methods have been developed in order to partition such structures, among which several rely on underlying probabilistic models, typically mixture models. The relevant hidden structure may however show overlapping groups in several applications. The Overlapping Stochastic Block Model (2011) has been developed to take this phenomenon into account. Nevertheless, the problem of the choice of the number of classes in the inference step is still open. To tackle this issue, we consider the proposed model in a Bayesian framework and develop a new criterion based on a non asymptotic approximation of the marginal log-likelihood. We describe how the criterion can be computed through a variational Bayes EM algorithm, and demonstrate its efficiency by running it on both simulated and real data., Comment: article

International audience; Inferring the structure of populations has many applications for genetic research. In addition to providing information for evolutionary studies, it can be used to account for the bias induced by population stratification in association studies. To this end, many algorithms have been proposed to cluster individuals into genetically homogeneous sub-populations. The parametric algorithms, such as Structure, are very popular but their underlying complexity and their high computational cost led to the development of faster parametric alternatives such as Admixture. Alternatives to these methods are the non-parametric approaches. Among this category, AWclust has proven efficient but fails to properly identify population structure for complex datasets. We present in this article a new clustering algorithm called Spectral Hierarchical clustering for the Inference of Population Structure (SHIPS), based on a divisive hierarchical clustering strategy, allowing a progressive investigation of population structure. This method takes genetic data as input to cluster individuals into homogeneous sub-populations and with the use of the gap statistic estimates the optimal number of such sub-populations. SHIPS was applied to a set of simulated discrete and admixed datasets and to real SNP datasets, that are data from the HapMap and Pan-Asian SNP consortium. The programs Structure, Admixture, AWclust and PCAclust were also investigated in a comparison study. SHIPS and the parametric approach Structure were the most accurate when applied to simulated datasets both in terms of individual assignments and estimation of the correct number of clusters. The analysis of the results on the real datasets highlighted that the clusterings of SHIPS were the more consistent with the population labels or those produced by the Admixture program. The performances of SHIPS when applied to SNP data, along with its relatively low computational cost and its ease of use make this method a promising solution to infer fine-scale genetic patterns.

It is now widely accepted that knowledge can be acquired from networks by clustering their vertices according to the connection profiles. Many methods have been proposed and in this paper we concentrate on the Stochastic Block Model (SBM). The clustering of vertices and the estimation of SBM model parameters have been subject to previous work, and numerous inference strategies such as variational expectation maximization (EM) and classification EM have been proposed. However, SBM still suffers from a lack of criteria to estimate the number of components in the mixture. To our knowledge, only one model-based criterion, Integrated Complete-data Likelihood (ICL), has been derived for SBM in the literature. It relies on an asymptotic approximation of the integrated complete-data likelihood and recent studies have shown that it tends to be too conservative in the case of small networks. To tackle this issue, we propose a new criterion that we call Integrated Likelihood Variational Bayes (ILvb), based on a non-asymptotic approximation of the marginal likelihood. We describe how the criterion can be computed through a variational Bayes EM algorithm.

Summary Random-graph mixture models are very popular for modelling real data networks. Parameter estimation procedures usually rely on variational approximations, either combined with the expectation–maximization (EM) algorithm or with Bayesian approaches. Despite good results on synthetic data, the validity of the variational approximation is, however, not established. Moreover, these variational approaches aim at approximating the maximum likelihood or the maximum a posteriori estimators, whose behaviour in an asymptotic framework (as the sample size increases to ∞) remains unknown for these models. In this work, we show that, in many different affiliation contexts (for binary or weighted graphs), parameter estimators based either on moment equations or on the maximization of some composite likelihood are strongly consistent and √n convergent, when the number n of nodes increases to ∞. As a consequence, our result establishes that the overall structure of an affiliation model can be (asymptotically) caught by the description of the network in terms of its number of triads (order 3 structures) and edges (order 2 structures). Moreover, these parameter estimates are either explicit (as for the moment estimators) or may be approximated by using a simple EM algorithm, whose convergence properties are known. We illustrate the efficiency of our method on simulated data and compare its performances with other existing procedures. A data set of cross-citations among economics journals is also analysed.

International audience; Genome-Wide Association Studies are powerful tools to detect genetic variants associated with diseases. Their results have, however, been questioned, in part because of the bias induced by population stratification. This is a consequence of systematic differences in allele frequencies due to the difference in sample ancestries that can lead to both false positive or false negative findings. Many strategies are available to account for stratification but their performances differ, for instance according to the type of population structure, the disease susceptibility locus minor allele frequency, the degree of sampling imbalanced, or the sample size. We focus on the type of population structure and propose a comparison of the most commonly used methods to deal with stratification that are the Genomic Control, Principal Component based methods such as implemented in Eigenstrat, adjusted Regressions and Meta-Analyses strategies. Our assessment of the methods is based on a large simulation study, involving several scenarios corresponding to many types of population structures. We focused on both false positive rate and power to determine which methods perform the best. Our analysis showed that if there is no population structure, none of the tests led to a bias nor decreased the power except for the Meta-Analyses. When the population is stratified, adjusted Logistic Regressions and Eigenstrat are the best solutions to account for stratification even though only the Logistic Regressions are able to constantly maintain correct false positive rates. This study provides more details about these methods. Their advantages and limitations in different stratification scenarios are highlighted in order to propose practical guidelines to account for population stratification in Genome-Wide Association Studies.

Complex systems in nature and in society are often represented as networks, describing the rich set of interactions between objects of interest. Many deterministic and probabilistic clustering methods have been developed to analyze such structures. Given a network, almost all of them partition the vertices into disjoint clusters, according to their connection profile. However, recent studies have shown that these techniques were too restrictive and that most of the existing networks contained overlapping clusters. To tackle this issue, we present in this paper the Overlapping Stochastic Block Model. Our approach allows the vertices to belong to multiple clusters, and, to some extent, generalizes the well-known Stochastic Block Model [Nowicki and Snijders (2001)]. We show that the model is generically identifiable within classes of equivalence and we propose an approximate inference procedure, based on global and local variational techniques. Using toy data sets as well as the French Political Blogosphere network and the transcriptional network of Saccharomyces cerevisiae, we compare our work with other approaches., Comment: Published in at http://dx.doi.org/10.1214/10-AOAS382 the Annals of Applied Statistics (http://www.imstat.org/aoas/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We present a weighted-Lasso method to infer the parameters of a first-order vector auto-regressive model that describes time course expression data generated by directed gene-to-gene regulation networks. These networks are assumed to own a prior internal structure of connectivity which drives the inference method. This prior structure can be either derived from prior biological knowledge or inferred by the method itself. We illustrate the performance of this structure-based penalization both on synthetic data and on two canonical regulatory networks, first yeast cell cycle regulation network by analyzing Spellman et al's dataset and second E. coli S.O.S. DNA repair network by analysing U. Alon's lab data.

In this paper we adapt online estimation strategies to perform model-based clustering on large networks. Our work focuses on two algorithms, the first based on the SAEM algorithm, and the second on variational methods. These two strategies are compared with existing approaches on simulated and real data. We use the method to decipher the connexion structure of the political websphere during the US political campaign in 2008. We show that our online EM-based algorithms offer a good trade-off between precision and speed, when estimating parameters for mixture distributions in the context of random graphs., Published in at http://dx.doi.org/10.1214/10-AOAS359 the Annals of Applied Statistics (http://www.imstat.org/aoas/) by the Institute of Mathematical Statistics (http://www.imstat.org)

Summary: The R package SIMoNe (Statistical Inference for MOdular NEtworks) enables inference of gene-regulatory networks based on partial correlation coefficients from microarray experiments. Modelling gene expression data with a Gaussian graphical model (hereafter GGM), the algorithm estimates non-zero entries of the concentration matrix, in a sparse and possibly high-dimensional setting. Its originality lies in the fact that it searches for a latent modular structure to drive the inference procedure through adaptive penalization of the concentration matrix. Availability: Under the GNU General Public Licence at http://cran.r-project.org/web/packages/simone/ Contact: julien.chiquet@genopole.cnrs.fr

International audience; Large datasets with interactions between objects are common to numerous scientific fields including the social sciences and biology, as well as being a feature of specific phenomena such as the internet. The interactions naturally define a graph, and a common way of exploring and summarizing such datasets is graph clustering. Most techniques for clustering graph vertices use only the topology of connections, while ignoring information about the vertices’ features. In this paper we provide a clustering algorithm that harnesses both types of data, based on a statistical model with a latent structure characterizing each vertex both by a vector of features and by its connectivity. We perform simulations to compare our algorithm with existing approaches, and also evaluate our method using real datasets based on hypertext documents. We find that our algorithm successfully exploits whatever information is found both in the connectivity pattern and in the features.

Our concern is selecting the concentration matrix's nonzero coefficients for a sparse Gaussian graphical model in a high-dimensional setting. This corresponds to estimating the graph of conditional dependencies between the variables. We describe a novel framework taking into account a latent structure on the concentration matrix. This latent structure is used to drive a penalty matrix and thus to recover a graphical model with a constrained topology. Our method uses an $\ell_1$ penalized likelihood criterion. Inference of the graph of conditional dependencies between the variates and of the hidden variables is performed simultaneously in an iterative \textsc{em}-like algorithm. The performances of our method is illustrated on synthetic as well as real data, the latter concerning breast cancer., Comment: 35 pages, 15 figures

Networks are used in many scientific fields such as biology, social science, and information technology. They aim at modelling, with edges, the way objects of interest, represented by vertices, are related to each other. Looking for clusters of vertices, also called communities or modules, has appeared to be a powerful approach for capturing the underlying structure of a network. In this context, the Block-Clustering model has been applied on random graphs. The principle of this method is to assume that given the latent structure of a graph, the edges are independent and generated from a parametric distribution. Many EM-like strategies have been proposed, in a frequentist setting, to optimize the parameters of the model. Moreover, a criterion, based on an asymptotic approximation of the Integrated Classification Likelihood (ICL), has recently been derived to estimate the number of classes in the latent structure. In this paper, we show how the Block-Clustering model can be described in a full Bayesian framework and how the posterior distribution, of all the parameters and latent variables, can be approximated efficiently applying Variational Bayes (VB). We also propose a new non-asymptotic Bayesian model selection criterion. Using simulated data sets, we compare our approach to other strategies. We show that our criterion can outperform ICL.

International audience; BACKGROUND: Identifying gene functional modules is an important step towards elucidating gene functions at a global scale. Clustering algorithms mostly rely on co-expression of genes, that is group together genes having similar expression profiles. RESULTS: We propose to cluster genes by co-regulation rather than by co-expression. We therefore present an inference algorithm for detecting co-regulated groups from gene expression data and introduce a method to cluster genes given that inferred regulatory structure. Finally, we propose to validate the clustering through a score based on the GO enrichment of the obtained groups of genes. CONCLUSION: We evaluate the methods on the stress response of S. Cerevisiae data and obtain better scores than clustering obtained directly from gene expression.

A new method for function estimation and variable selection, specifically designed for additive models fitted by cubic splines is proposed. This new method involves regularizing additive models using the l"1-norm, which generalizes the lasso to the nonparametric setting. As in the linear case, it shrinks coefficients and produces some coefficients that are exactly zero. It gives parsimonious models, selects significant variables, and reveals nonlinearities in the effects of predictors. Two strategies for finding a parsimonious additive model solution are proposed. Both algorithms are based on a fixed point algorithm, combined with a singular value decomposition that considerably reduces computation. The empirical behavior of parsimonious additive models is compared to the adaptive backfitting BRUTO algorithm. The results allow to characterize the domains in which our approach is effective: it performs significantly better than BRUTO when model estimation is challenging. An implementation of this method is illustrated using real data from the Cophar 1 ANRS 102 trial. Parsimonious additive models are applied to predict the indinavir plasma concentration in HIV patients. Results suggest that this new method is a promising technique for the research and application areas.

International audience; This paper tackles the problem of model complexity in the context of additive models. Several methods have been proposed to estimate smoothing parameters, as well as to perform variable selection. However, these procedures are inefficient or computationally expensive in high dimension. To answer this problem, the lasso technique has been adapted to additive models, but its experimental performance has not been analyzed. We propose a modified lasso for additive models, performing variable selection. A benchmark is developed to examine its practical behavior, comparing it with forward selection. Our simulation studies suggest ability to carry out model selection of the proposed method. The lasso technique shows up better than forward selection in the most complex situations. The computing time of modified lasso is considerably smaller since it does not depend on the number of relevant variables.

Basing cluster analysis on mixture models has become a classical and powerful approach. It proves to be useful for understanding and suggesting significant criteria. With Gaussian clustering models, for instance, the parametrization of the variance matrix of a cluster in terms of its eigenvalue decomposition allows to propose many general clustering criteria from the simplest (k-means criterion) to the most complex. Mixture models can also deal with very different situations such as quantitative data, binary data, spatial data, missing data, partially classified samples, order constraints on clusters or outliers. This paper intends to illustrate the application of mixture models to various clustering situations from classical Gaussian models to multivariate binary observation vectors located at neighboring geographic sites.

International audience; This paper introduces MixtBoost, a variant of AdaBoost dedicated to solve problems in which both labeled and unlabeled data are available. We propose several definitions of loss for unlabeled data, from which margins are defined. The resulting boosting schemes implement mixture models as base classifiers. Preliminary experiments are analyzed and the relevance of loss choices is discussed. MixtBoost improves on both mixture models and AdaBoost provided classes are structured, and is otherwise similar to AdaBoost.

Mixture models are widely used for clustering or discrimination problems. Estimating the parameters of such models can be viewed as an incomplete data problem and has thus often been handled by the Expectation-Maximization (EM) algorithm. It has been shown that this method can integrate additional information such as the label of some observations. In this paper we propose a generalization of this approach which can take into account partial information about the observation labels. An example illustrates the relevance of the proposed method for mixture density estimation.

Basing cluster analysis on mixture models has become a classical and powerful approach. In this paper we propose an extension of this approach to fuzzy clustering. We define a fuzzy clustering criterion which generalizes both the maximum likelihood and the classification maximum likelihood. Finally using a generalization of the well-known EM and CEM algorithms we design an algorithm to optimize this criterion.

This paper presents a new method for segmenting multispectral satellite images. The proposed method is unsupervised and consists of two steps. During the first step the pixels of a learning set are summarized by a set of codebook vectors using a Probabilistic Self-Organizing Map (PSOM, Statistique et methodes neuronales, Dunod, Paris, 1997). In a second step the codebook vectors of the map are clustered using Agglomerative Hierarchical Clustering (AHC, Pattern Recognition and Neural Networks, Cambridge University Press, Cambridge, 1996). Each pixel takes the label of its nearest codebook vector. A practical application to Meteosat images illustrates the relevance of our approach.

A clustering algorithm for spatial data is presented. It seeks a fuzzy partition which is optimal according to a criterion interpretable as a penalized likelihood. We propose to penalize the energy function exhibited by Hathaway (1986) with a term taking into account spatial contiguity constraints. The structure of the EM algorithm may be used to maximize the proposed criterion. The Maximization step is then unchanged and the Expectation step becomes iterative. The efficiency of the new clustering algorithm has been tested with biological images and compared with other clustering techniques.