In a recent article in PLoS Genetics, Bock et al., (2006) undertake an extensive computational epigenetics analysis of the ability of DNA sequence-derived features, capturing attributes such as tetramer frequencies, repeats and predicted structure, to predict the methylation status of CpG islands. Their suite of analyses appears highly rigorous with regard to accompanying validation procedures, employing stringent Bonferroni corrections, stratified cross-validation, and follow-up experimental verification. Here, however, we showcase concerns with the validation steps, in part ascribable to the genome scale of the investigation, that serve as a cautionary note and indicate the heightened need for careful selection of analytic and companion validation methods. A series of new analyses of the same CpG island methylation data helps illustrate these issues, not just for this particular study, but also analogous investigations involving high-dimensional predictors with complex between-feature dependencies.

Chess and chance are seemingly strange bedfellows. Luck and/or randomness have no apparent role in move selection when the game is played at the highest levels. However, when competition is at the ultimate level, that of the World Chess Championship (WCC), chess and conspiracy are not strange bedfellows, there being a long and colorful history of accusations levied between participants. One such accusation, frequently repeated, was that all the games in the 1985 WCC (Karpov vs Kasparov) were fixed and pre-arranged move-by-move. That this claim was advanced by a former World Champion, Bobby Fischer, argues that it ought be investigated. That the only published, concrete basis for this claim consists of an observed run of particular moves, allows this investigation to be performed using probabilistic and statistical methods. In particular, we employ imbedded finite Markov chains to evaluate run statistic distributions. Further, we demonstrate how both chess computers and game databases can be brought to bear on the problem.

Diffuse large-B-cell lymphoma (DLBCL) is an aggressive malignancy of mature B lymphocytes and is the most common type of lymphoma in adults. While treatment advances have been substantial in what was formerly a fatal disease, less than 50% of patients achieve lasting remission. In an effort to predict treatment success and explain disease heterogeneity clinical features have been employed for prognostic purposes, but have yielded only modest predictive performance. This has spawned a series of high profile microarray-based gene expression studies of DLBCL, in the hope that molecular level information could be used to refine prognosis. The intent of this paper is to reevaluate these microarray-based prognostic assessments, and extend the statistical methodology that has been used in this context.

Methodological challenges arise in using patients’ gene expression profiles to predict survival endpoints on account of the large number of genes and their complex interdependence. We initially focus on the Lymphochip data and analysis of Rosenwald et al., (2002). After describing relationships between the analyses performed and gene harvesting (Hastie et al., 2001), we argue for the utility of penalized approaches, in particular LARS-Lasso (Efron et al., 2004). While these techniques have been extended to the proportional hazards / partial likelihood framework, the resultant algorithms are computationally burdensome. We develop residualbased approximations that eliminate this burden yet perform similarly. Comparisons of predictive accuracy across both methods and studies are effected using time-dependent ROC curves. These indicate that gene expression data, in turn, only delivers modest predictions of post therapy DLBCL survival. We conclude by outlining possibilities for further work.

Nucleosomes, the fundamental repeating subunits of all eukaryotic chromatin, are responsible for packaging DNA into chromosomes inside the cell nucleus and controlling gene expression. While it has been well established that nucleosomes exhibit higher affinity for select DNA sequences, until recently it was unclear whether such preferences exerted a significant, genome-wide effect on nucleosome positioning in vivo. This question was seemingly and recently resolved in the affirmative: a wide-ranging series of experimental and computational analyses provided extensive evidence that the instructions for wrapping DNA around nucleosomes are contained in the DNA itself. This subsequently labelled second genetic code was based on data-driven, structural, and biophysical considerations. It was subjected to an extensive suite of validation procedures, with one conclusion being that intrinsic, genome-encoded, nucleosome organization explains _50% of in vivo nucleosome positioning. Here, we revisit both the nature of the underlying sequence preferences, and the performance of the proposed code. A series of new analyses, employing spectral envelope (Fourier transform) methods for assessing key sequence periodicities, classification techniques for evaluating predictive performance, and discriminatory motif finding methods for devising alternate models, are applied. The findings from the respective analyses indicate that signature dinucleotide periodicities are absent from the bulk of the high affinity nucleosome-bound sequences, and that the predictive performance of the code is modest. We conclude that further exploration of the role of sequence-based preferences in genome-wide nucleosome positioning is warranted. This work offers a methodologic counterpart to a recent, high resolution determination of nucleosome positioning that also questions the accuracy of the proposed code and, further, provides illustration of techniques useful in assessing sequence periodicity and predictive performance.

Breiman (2001a,b) has recently developed an ensemble classification and regression approach that displayed outstanding performance with regard prediction error on a suite of benchmark datasets. As the base constituents of the ensemble are tree-structured predictors, and since each of these is constructed using an injection of randomness, the method is called ‘random forests’. That the exceptional performance is attained with seemingly only a single tuning parameter, to which sensitivity is minimal, makes the methodology all the more remarkable. The individual trees comprising the forest are all grown to maximal depth. While this helps with regard bias, there is the familiar tradeoff with variance. However, these variability concerns were potentially obscured because of an interesting feature of those benchmarking datasets extracted from the UCI machine learning repository for testing: all these datasets are hard to overfit using tree-structured methods. This raises issues about the scope of the repository.

With this as motivation, and coupled with experience from boosting methods, we revisit the formulation of random forests and investigate prediction performance on real-world and simulated datasets for which maximally sized trees do overfit. These explorations reveal that gains can be realized by additional tuning to regulate tree size via limiting the number of splits and/or the size of nodes for which splitting is allowed. Nonetheless, even in these settings, good performance for random forests can be attained by using larger (than default) primary tuning parameter values.

In a recent article in this journal, Khatun, Hamlett, and Giddings (2008) (KHG) advance a new scoring scheme for use in conjunction with tandem mass spectrometry (MS/MS) based peptide identification. As they note, such identifications are fundamental to much proteomics research but, due to MS/MS data complexity and the scale of attendant database searches, their accuracy is limited. The scoring technique they propose, which employs a hidden Markov model (HMM) over a set of states that represent key features of MS/MS data, is convincingly motivated and exhibits good performance. The purpose of this brief note is to critique the method chosen for calibrating the HMM scores, rather than the genesis of the scores themselves.

Various topologies for representing three dimensional protein structures have been advanced for purposes ranging from prediction of folding rates to ab initio structure prediction. Examples include relative contact order, Delaunay tessellations, and backbone torsion angle distributions. Here we introduce a new topology based on a novel means for operationalizing three dimensional proximities with respect to the underlying chain. The measure involves first interpreting a rank-based representation of the nearest neighbors of each residue as a permutation, then determining how perturbed this permutation is relative to an unfolded chain. We show that the resultant topology provides improved association with folding and unfolding rates determined for a set of two-state proteins under standardized conditions. Furthermore, unlike existing topologies, the proposed geometry exhibits fine scale structure with respect to sequence position along the chain, potentially providing insights into folding initiation and/or nucleation sites.

Beyond linear sequence, higher order structure of the genome influences gene regulation and has been implicated in disease. Chromatin structure is the degree of chromatin compaction at genomic loci. Chromatin organization is the spatial, three-dimensional (3D) positioning of chromatin. Here, we adapt and apply methods for next-generation sequencing analyses of chromatin structure and organization based on chromatin immunoprecipitation-sequencing (ChIP-seq) and genome-wide chromosome conformation capture (Hi-C), respectively. First, we built on a previous study that sought to classify nucleosomes containing either H2A.Z or H2A/H4 arginine 3 symmetric dimethylation (H2A/H4R3me2s) from human ChIP-seq data. We hypothesized that appropriate data preprocessing – deduplication, normalization for sequencing depth, and position-finding – in conjunction with advanced algorithms for feature selection (Discriminatory Motif Feature Selection) and classification (Random Forest) would improve performance. We achieved dramatically improved classification accuracy and identified a significant and biologically meaningful DNA motif associated with H2A/H4R3me2s: “TCCATT”, which is part of the consensus sequence of satellite II and III DNA. Second, we tested our hypothesis that there are advantages to assessing the 3D co-localization of functional annotations (e.g., centromeres) using 3D genome reconstructions from Hi-C contact data because they enable detection of multi-level interactions (assessments using contact data are inherently limited to detecting strictly pairwise interactions). We found significant 3D co-localization of sets of genes with developmentally regulated expression in Plasmodium falciparum with 3D reconstruction-based assessment but not with contact-based assessment. Further, we developed a method for 3D reconstruction-based assessment that avoids the data dichotomization of previous approaches. Third, we tested our hypothesis that analyzing ChIP-seq data in combination with 3D reconstructions could identify functional 3D hotspots. We separately overlaid a Saccharomyces cerevisiae 3D genome reconstruction with three ChIP-seq inputs and contrasted two algorithms for identifying regions in 3-space — 3D hotspots – for which mean ChIP-seq peak height is significantly elevated: k-Nearest Neighbor (k-NN) regression and the Patient Rule Induction Method (PRIM). For each ChIP-seq input, both algorithms identified significant, corresponding and biologically meaningful 3D hotspots containing distal genomic regions. Our research demonstrates that applying appropriate data preprocessing and advanced supervised learning algorithms improves the interpretability of next-generation sequencing studies of chromatin structure and organization.

Identification of peptides binding to Major Histocompatibility Complex (MHC) molecules is important for accelerating vaccine development and improving immunotherapy. Accordingly, a wide variety of prediction methods have been applied in this context. In this paper, we introduce (tree-based) ensemble classifiers for such problems and contrast their predictive performance with forefront existing methods for both MHC class I and class II molecules. In addition, we investigate the impact of differing peptide representation schemes on performance. Finally, classifier predictions are used to conduct genomewide scans of a diverse collection of HIV-1 strains, enabling assessment of epitope conservation. We investigated all combinations of six classifi- cation methods (classification trees, artificial neural networks, support vector machines, as well as the more recently devised ensemble methods (bagging, random forests, boosting) with four peptide representation schemes (amino acid sequence, select biophysical properties, select quantitative structure-activity relationship (QSAR) descriptors, and the combination of the latter two) in predicting peptide binding to an MHC class I molecule (HLA-A2) and MHC class II molecule (HLA-DR4). Our results show that the ensemble methods are consistently more accurate than the other three alternatives. Furthermore, they are robust with respect to parameter tuning. Among the four representation schemes, the amino acid sequence representation gave consistently (across classifiers) best results. This finding obviates the need for feature selection strategies incurred by use of biophysical and/or QSAR properties. We obtained, and aligned, a diverse set of 32 HIV-1 genomes and pursued genomewide HLA-DR4 epitope profiling by querying with respect to classifier predictions, as obtained under each of the four peptide representation schemes. We validated those epitopes conserved across strains against known T-cell epitopes. Once again, amino acid sequence representation was at least as effective as using properties. Assessment of novel epitope predictions awaits experimental verification.

The recent availability of whole-genome scale data sets that investigate complementary and diverse aspects of transcriptional regulation has spawned an increased need for new and effective computational approaches to analyze and integrate these large scale assays. Here, we propose a novel algorithm, based on random forest methodology, to relate gene expression (as derived from expression microarrays) to sequence features residing in gene promoters (as derived from DNA motif data) and transcription factor binding to gene promoters (as derived from tiling microarrays). We extend the random forest approach to model a multivariate response as represented, for example, by time-course gene expression measures. An analysis of the multivariate random forest output reveals complex regulatory networks, which consist of cohesive, condition-dependent regulatory cliques. Each regulatory clique features homogeneous gene expression profiles and common motifs or synergistic motif groups. We apply our method to several yeast physiological processes: cell cycle, sporulation, and various stress conditions. Our technique displays excellent performance with regard to identifying known regulatory motifs, including high order interactions. In addition, we present evidence of the existence of an alternative MCB-binding pathway, which we confirm using data from two independent cell cycle studies and two other physioloigical processes. Finally, we have uncovered elaborate transcription regulation refinement mechanisms involving PAC and mRRPE motifs that govern essential rRNA processing. These include intriguing instances of differing motif dosages and differing combinatorial motif control that promote regulatory specificity in rRNA metabolism under differing physiological processes.