Volume 11, Issue 1

Hypotheses tests in bioinformatics can often be set in a tree structure in a very natural way, e.g. when tests are performed at probe, gene, and chromosome level. Exploiting this graph structure in a multiple testing procedure may result in a gain in power or increased interpretability of the results.We present the inheritance procedure, a method of familywise error control for hypotheses structured in a tree. The method starts testing at the top of the tree, following up on those branches in which it finds significant results, and following up on leaf nodes in the neighborhood of those leaves. The method is a uniform improvement over a recently proposed method by Meinshausen. The inheritance procedure has been implemented in the globaltest package which is available on www.bioconductor.org.

We discuss the definition and application of design criteria for evaluating the efficiency of 2-color microarray designs. First, we point out that design optimality criteria are defined differently for the regression and block design settings. This has caused some confusion in the literature and warrants clarification. Linear models for microarray data analysis have equivalent formulations as ANOVA or regression models. However, this equivalence does not extend to design criteria. We discuss optimality criterion, and argue against applying regression-style D-optimality to the microarray design problem. We further disfavor E- and D-optimality (as defined in block design) because they are not attuned to scientific questions of interest.

To extract full information from samples of DNA sequence data, it is necessary to use sophisticated model-based techniques such as importance sampling under the coalescent. However, these are limited in the size of datasets they can handle efficiently. Chen and Liu (2000) introduced the idea of stopping-time resampling and showed that it can dramatically improve the efficiency of importance sampling methods under a finite-alleles coalescent model. In this paper, a new framework is developed for designing stopping-time resampling schemes under more general models. It is implemented on data both from infinite sites and stepwise models of mutation, and extended to incorporate crossover recombination. A simulation study shows that this new framework offers a substantial improvement in the accuracy of likelihood estimation over a range of parameters, while a direct application of the scheme of Chen and Liu (2000) can actually diminish the estimate. The method imposes no additional computational burden and is robust to the choice of parameters.

Testing for unequal variances is usually performed in order to check the validity of the assumptions that underlie standard tests for differences between means (the t-test and anova). However, existing methods for testing for unequal variances (Levene's test and Bartlett's test) are notoriously non-robust to normality assumptions, especially for small sample sizes. Moreover, although these methods were designed to deal with one hypothesis at a time, modern applications (such as to microarrays and fMRI experiments) often involve parallel testing over a large number of levels (genes or voxels). Moreover, in these settings a shift in variance may be biologically relevant, perhaps even more so than a change in the mean. This paper proposes a parsimonious model for parallel testing of the equal variance hypothesis. It is designed to work well when the number of tests is large; typically much larger than the sample sizes. The tests are implemented using an empirical Bayes estimation procedure which `borrows information' across levels. The method is shown to be quite robust to deviations from normality, and to substantially increase the power to detect differences in variance over the more traditional approaches even when the normality assumption is valid.

Where causal SNPs (single nucleotide polymorphisms) tend to accumulate within biological pathways, the incorporation of prior pathways information into a statistical model is expected to increase the power to detect true associations in a genetic association study. Most existing pathways-based methods rely on marginal SNP statistics and do not fully exploit the dependence patterns among SNPs within pathways.We use a sparse regression model, with SNPs grouped into pathways, to identify causal pathways associated with a quantitative trait. Notable features of our “pathways group lasso with adaptive weights” (P-GLAW) algorithm include the incorporation of all pathways in a single regression model, an adaptive pathway weighting procedure that accounts for factors biasing pathway selection, and the use of a bootstrap sampling procedure for the ranking of important pathways. P-GLAW takes account of the presence of overlapping pathways and uses a novel combination of techniques to optimise model estimation, making it fast to run, even on whole genome datasets.In a comparison study with an alternative pathways method based on univariate SNP statistics, our method demonstrates high sensitivity and specificity for the detection of important pathways, showing the greatest relative gains in performance where marginal SNP effect sizes are small.

MicroRNAs (miRNAs) are non-coding, short (21-23nt) regulators of protein-coding genes that are generally transcribed first into primary miRNA (pri-miR), followed by the generation of precursor miRNA (pre-miR). This finally leads to the production of the mature miRNA. A large amount of information is available on the pre- and mature miRNAs. However, very little is known about the pri-miRs, due to a lack of knowledge about their transcription start sites (TSSs). Based on the genomic loci, miRNAs can be categorized into two types —intragenic (intra-miR) and intergenic (inter-miR). While it is already an established fact that intra-miRs are commonly transcribed in conjunction with their host genes, the transcription machinery of inter-miRs is poorly understood. Although it is assumed that miRNA promoters are similar in structure to gene promoters, since both are transcribed by RNA polymerase II (Pol II), computational validations exhibit poor performance of gene promoter prediction methods on miRNAs. In this paper, we concentrate on the problem of TSS prediction for miRNAs. The present study begins with the identification of positive and negative promoter samples from recently published data stemming from RNA-sequencing studies. From these samples of experimentally validated miRNA TSSs, a number of standard sequence features are extracted. Furthermore, to account for potential footprints related to promoter regulation by CpG dinucleotide targeted DNA methylation, a number of novel features are defined. We develop a support vector machine (SVM) with RBF kernel for the prediction of miRNA TSSs trained on human miRNA promoters. A novel feature reduction technique based on archived multi-objective simulated annealing (AMOSA) identifies the final set of features. The resulting model trained on miRNA promoters shows improved performance over the one trained on protein-coding gene promoters in terms of classification accuracy, sensitivity and specificity. Results are also reported for a completely independent biologically validated test set. In a part of the investigation, the proposed approach is used to predict protein-coding gene TSSs. It shows a significantly improved performance when compared to previously published gene TSS prediction methods.

This article proposes a novel approach to statistical alignment of nucleotide sequences by introducing a context dependent structure on the substitution process in the underlying evolutionary model. We propose to estimate alignments and context dependent mutation rates relying on the observation of two homologous sequences. The procedure is based on a generalized pair-hidden Markov structure, where conditional on the alignment path, the nucleotide sequences follow a Markov distribution. We use a stochastic approximation expectation maximization (saem) algorithm to give accurate estimators of parameters and alignments. We provide results both on simulated data and vertebrate genomes, which are known to have a high mutation rate from CG dinucleotide. In particular, we establish that the method improves the accuracy of the alignment of a human pseudogene and its functional gene.

For a better understanding of the biology of an organism, a complete description is needed of all regions of the genome that are actively transcribed. Tiling arrays are used for this purpose. They allow for the discovery of novel transcripts and the assessment of differential expression between two or more experimental conditions such as genotype, treatment, tissue, etc. In tiling array literature, many efforts are devoted to transcript discovery, whereas more recent developments also focus on differential expression. To our knowledge, however, no methods for tiling arrays have been described that can simultaneously assess transcript discovery and identify differentially expressed transcripts. In this paper, we adopt wavelet based functional models to the context of tiling arrays. The high dimensionality of the data triggered us to avoid inference based on Bayesian MCMC methods. Instead, we introduce a fast empirical Bayes method that provides adaptive regularization of the functional effects. A simulation study and a case study illustrate that our approach is well suited for the simultaneous assessment of transcript discovery and differential expression in tiling array studies, and that it outperforms methods that accomplish only one of these tasks.

The D2 statistic, defined as the number of matches of words of some pre-specified length k, is a computationally fast alignment-free measure of biological sequence similarity. However there is some debate about its suitability for this purpose as the variability in D2 may be dominated by the terms that reflect the noise in each of the single sequences only. We examine the extent of the problem and the effectiveness of overcoming it by using two mean-centred variants of this statistic, D2* and D2c. We conclude that all three statistics are potentially useful measures of sequence similarity, for which reasonably accurate p-values can be estimated under a null hypothesis of sequences composed of identically and independently distributed letters. We show that D2 and D2c, and to a somewhat lesser extent D2*, perform well in tests to classify moderate length query sequences as putative cis-regulatory modules.

An important challenge in system biology is the inference of biological networks from postgenomic data. Among these biological networks, a gene transcriptional regulatory network focuses on interactions existing between transcription factors (TFs) and and their corresponding target genes. A large number of reverse engineering algorithms were proposed to infer such networks from gene expression profiles, but most current methods have relatively low predictive performances. In this paper, we introduce the novel TNIFSED method (Transcriptional Network Inference from Functional Similarity and Expression Data), that infers a transcriptional network from the integration of correlations and partial correlations of gene expression profiles and gene functional similarities through a supervised classifier. In the current work, TNIFSED was applied to predict the transcriptional network in Escherichia coli and in Saccharomyces cerevisiae, using datasets of 445 and 170 affymetrix arrays, respectively. Using the area under the curve of the receiver operating characteristics and the F-measure as indicators, we showed the predictive performance of TNIFSED to be better than unsupervised state-of-the-art methods. TNIFSED performed slightly worse than the supervised SIRENE algorithm for the target genes identification of the TF having a wide range of yet identified target genes but better for TF having only few identified target genes. Our results indicate that TNIFSED is complementary to the SIRENE algorithm, and particularly suitable to discover target genes of "orphan" TFs.

Profile Hidden Markov Models (pHMMs) are widely used to model nucleotide or protein sequence families. In many applications, a sequence family classified into several subfamilies is given and each subfamily is modeled separately by one pHMM. A major drawback of this approach is the difficulty of coping with subfamilies composed of very few sequences.Correct subtyping of human immunodeficiency virus-1 (HIV-1) sequences is one of the most crucial bioinformatic tasks affected by this problem of small subfamilies, i.e., HIV-1 subtypes with a small number of known sequences. To deal with small samples for particular subfamilies of HIV-1, we employ a machine learning approach. More precisely, we make use of an existing HMM architecture and its associated inference engine, while replacing the unsupervised estimation of emission probabilities by a supervised method. For that purpose, we use regularized linear discriminant learning together with a balancing scheme to account for the widely varying sample size. After training the multiclass linear discriminants, the corresponding weights are transformed to valid probabilities using a softmax function.We apply this modified algorithm to classify HIV-1 sequence data (in the form of partial-length HIV-1 sequences and semi-artificial recombinants) and show that the performance of pHMMs can be significantly improved by the proposed technique.