Search

Motivation: Over the last decades, vast numbers of sequences were deposited in public databases. Bioinformatics tools allow homology and consequently functional inference for these sequences. New profile-based homology search tools have been introduced, allowing reliable detection of remote homologs, but have not been systematically benchmarked. To provide such a comparison, which can guide bioinformatics workflows, we extend and apply our previously developed benchmark approach to evaluate the 'next generation' of profile-based approaches, including CS-BLAST, HHSEARCH and PHMMER, in comparison with the non-profile based search tools NCBI-BLAST, USEARCH, UBLAST and FASTA. Method: We generated challenging benchmark datasets based on protein domain architectures within either the PFAM+Clan, SCOP/Superfamily or CATH/Gene3D domain definition schemes. From each dataset, homologous and non-homologous protein pairs were aligned using each tool, and standard performance metrics calculated. We further measured congruence of domain architecture assignments in the three domain databases. Results: CSBLAST and PHMMER had overall highest accuracy. FASTA, UBLAST and USEARCH showed large trade-offs of accuracy for speed optimization. Conclusion: Profile methods are superior at inferring remote homologs but the difference in accuracy between methods is relatively small. PHMMER and CSBLAST stand out with the highest accuracy, yet still at a reasonable computational cost. Additionally, we show that less than 0.1% of Swiss-Prot protein pairs considered homologous by one database are considered non-homologous by another, implying that these classifications represent equivalent underlying biological phenomena, differing mostly in coverage and granularity.

Statistical regularisation methods such as LASSO and related L-1 regularised regression methods are commonly used to construct models of gene regulatory networks. Although they can theoretically infer the correct network structure, they have been shown in practice to make errors, i.e. leave out existing links and include non-existing links. We show that L-1 regularisation methods typically produce a poor network model when the analysed data are ill-conditioned, i.e. the gene expression data matrix has a high condition number, even if it contains enough information for correct network inference. However, the correct structure of network models can be obtained for informative data, data with such a signal to noise ratio that existing links can be proven to exist, when these methods fail, by using least-squares regression and setting small parameters to zero, or by using robust network inference, a recent method taking the intersection of all non-rejectable models. Since available experimental data sets are generally ill-conditioned, we recommend to check the condition number of the data matrix to avoid this pitfall of L-1 regularised inference, and to also consider alternative methods.

The InParanoid database (http://InParanoid.sbc.su.se) provides a user interface to orthologs inferred by the InParanoid algorithm. As there are now international efforts to curate and standardize complete proteomes, we have switched to using these resources rather than gathering and curating the proteomes ourselves. InParanoid release 8 is based on the 66 reference proteomes that the 'Quest for Orthologs' community has agreed on using, plus 207 additional proteomes from the UniProt complete proteomes-in total 273 species. These represent 246 eukaryotes, 20 bacteria and seven archaea. Compared to the previous release, this increases the number of species by 173% and the number of pairwise species comparisons by 650%. In turn, the number of ortholog groups has increased by 423%. We present the contents and usages of InParanoid 8, and a detailed analysis of how the proteome content has changed since the previous release., AuthorCount:2

Differential gene expression analysis between healthy and diseased groups is a widely used approach to understand the molecular underpinnings of disease. A wide variety of experimental and bioinformatics techniques are available for this type of analysis, yet their impact on the reliability of the results has not been systematically studied. We performed a large scale comparative analysis of clinical expression data, using several background corrections and differential expression metrics. The agreement between studies was analyzed for study pairs of same cancer type, of different cancer types, and between cancer and non-cancer studies. We also replicated the analysis using differential coexpression. We found that agreement of differential expression is primarily dictated by the microarray platform, while differential coexpression requires large sample sizes. Two studies using different differential expression metrics may show no agreement, even if they agree strongly using the same metric. Our analysis provides practical recommendations for gene (co)expression analysis., AuthorCount:2

Gene regulatory network inference (that is, determination of the regulatory interactions between a set of genes) provides mechanistic insights of central importance to research in systems biology. Most contemporary network inference methods rely on a sparsity/regularization coefficient, which we call zeta (zeta), to determine the degree of sparsity of the network estimates, that is, the total number of links between the nodes. However, they offer little or no advice on how to select this sparsity coefficient, in particular, for biological data with few samples. We show that an empty network is more accurate than estimates obtained for a poor choice of zeta. In order to avoid such poor choices, we propose a method for optimization of zeta, which maximizes the accuracy of the inferred network for any sparsity-dependent inference method and data set. Our procedure is based on leave-one-out cross-optimization and selection of the zeta value that minimizes the prediction error. We also illustrate the adverse effects of noise, few samples, and uninformative experiments on network inference as well as our method for optimization of zeta. We demonstrate that our zeta optimization method for two widely used inference algorithms-Glmnet and NIR-gives accurate and informative estimates of the network structure, given that the data is informative enough., QC 20130627

Network analysis is an important tool for functional annotation of genes and proteins. A common approach to discern structure in a global network is to infer network clusters, or modules, and assume a functional coherence within each module, which may represent a complex or a pathway. It is however not trivial to define optimal modules. Although many methods have been proposed, it is unclear which methods perform best in general. It seems that most methods produce far from optimal results but in different ways. MGclus is a new algorithm designed to detect modules with a strongly interconnected neighborhood in large scale biological interaction networks. In our benchmarks we found MGclus to outperform other methods when applied to random graphs with varying degree of noise, and to perform equally or better when applied to biological protein interaction networks. MGclus is implemented in Java and utilizes the JGraphT graph library. It has an easy to use command-line interface and is available for download from http://sonnhammer.sbc.su.se/download/software/MGclus/.

In this study, we describe a novel gene expression signature of platelet-derived growth factor (PDGF) activated fibroblasts, which is able to identify breast cancers with a PDGF-stimulated fibroblast stroma and displays an independent and strong prognostic significance. Global gene expression was compared between PDGF-stimulated human fibroblasts and cultured resting fibroblasts. The most differentially expressed genes were reduced to a gene expression signature of 113 genes. The biological significance and prognostic capacity of this signature were investigated using four independent clinical breast cancer data sets. Concomitant high expression of PDGF beta receptor and its cognate Ligands is associated with a high PDGF signature score. This supports the notion that the signature detects tumors with PDGF-activated stroma. Subsequent analyses indicated significant associations between high PDGF signature score and clinical characteristics, including human epidermal growth factor receptor 2 positivity, estrogen receptor negativity, high tumor grade, and large tumor size. A high PDGF signature score is associated with shorter survival in univariate analysis. Furthermore, the high PDGF signature score acts as a significant marker of poor prognosis in multivariate survival analyses, including classic prognostic markers, Ki-67 status, a proliferation gene signature, or other recently described stroma-derived gene expression signatures., AuthorCount:11

Motivation: Analyzing groups of functionally coupled genes or proteins in the context of global interaction networks has become an important aspect of bioinformatic investigations. Assessing the statistical significance of crosstalk enrichment between or within groups of genes can be a valuable tool for functional annotation of experimental gene sets. Results: Here we present CrossTalkZ, a statistical method and software to assess the significance of crosstalk enrichment between pairs of gene or protein groups in large biological networks. We demonstrate that the standard z-score is generally an appropriate and unbiased statistic. We further evaluate the ability of four different methods to reliably recover crosstalk within known biological pathways. We conclude that the methods preserving the second-order topological network properties perform best. Finally, we show how CrossTalkZ can be used to annotate experimental gene sets using known pathway annotations and that its performance at this task is superior to gene enrichment analysis (GEA). Availability and Implementation: CrossTalkZ (available at http://sonnhammer.sbc.su.se/download/software/CrossTalkZ/) is implemented in C++, easy to use, fast, accepts various input file formats, and produces a number of statistics. These include z-score, p-value, false discovery rate, and a test of normality for the null distributions., AuthorCount:4

An accurate inference of orthologs is essential in many research fields such as comparative genomics, molecular evolution, and genome annotation. Existing methods for genome-scale orthology inference are mostly based on all-versus-all similarity searches that scale quadratically with the number of species. This limits their application to the increasing number of available large-scale datasets. Here, we present Hieranoid, a new orthology inference method using a hierarchical approach. Hieranoid performs pairwise orthology analysis using In Paranoid at each node in a guide tree as it progresses from its leaves to the root. This concept reduces the total runtime complexity from a quadratic to a linear function of the number of species. The tree hierarchy provides a natural structure in multi-species ortholog groups, and the aggregation of multiple sequences allows for multiple alignment similarity searching techniques, which can yield more accurate ortholog groups. Using the recently published orthobench benchmark, Hieranoid showed the overall best performance. Our progressive approach presents a new way to infer orthologs that combines efficient graph-based methodology with aspects of compute-intensive tree-based methods. The linear scaling with the number of species is a major advantage for large-scale applications and makes Hieranoid well suited to cope with vast amounts of sequenced genomes in the future. Hieranoid is an open source and can be downloaded at Hieranoid.sbc.su.se., AuthorCount:2

Gene expression analysis is often used to investigate the molecular and functional underpinnings of a phenotype. However, differential expression of individual genes is limited in that it does not consider how the genes interact with each other in networks. To address this shortcoming we propose a number of network-based analyses that give additional functional insights into the studied process. These were applied to a dataset of sex-specific gene expression in the chicken gonad and brain at different developmental stages. We first constructed a global chicken interaction network. Combining the network with the expression data showed that most sex-biased genes tend to have lower network connectivity, that is, act within local network environments, although some interesting exceptions were found. Genes of the same sex bias were generally more strongly connected with each other than expected. We further studied the fates of duplicated sex-biased genes and found that there is a significant trend to keep the same pattern of sex bias after duplication. We also identified sex-biased modules in the network, which reveal pathways or complexes involved in sex-specific processes. Altogether, this work integrates evolutionary genomics with systems biology in a novel way, offering new insights into the modular nature of sex-biased genes., QC 20130201

mRNA expression is widely used as a proxy for protein expression. However, their true relation is not known and two genes with the same mRNA levels might have different abundances of respective proteins. A related question is whether the coexpression of mRNA for gene pairs is reflected by the corresponding protein pairs. We examined the mRNA-protein correlation for both expression and coexpression. This analysis yielded insights into the relationship between mRNA and protein abundance, and allowed us to identify subsets of greater mRNA-protein coherence. The correlation between mRNA and protein was low for both expression and coexpression, 0.12 and 0.06 respectively. However, applying the best-performing quality measure, high-quality subsets reached a Spearman correlation of 0.31 for expression, 034 for coexpression and 0.49 for coexpression when restricted to functionally coupled genes. Our methodology can thus identify subsets for which the mRNA levels are expected to be the strongest correlated with protein levels., 2

This chapter reviews the current research on how protein domain architectures evolve. We begin by summarizing work on the phylogenetic distribution of proteins, as this directly impacts which domain architectures can be formed in different species. Studies relating domain family size to occurrence have shown that they generally follow power law distributions, both within genomes and larger evolutionary groups. These findings were subsequently extended to multidomain architectures. Genome evolution models that have been suggested to explain the shape of these distributions arc reviewed, as well as evidence for selective pressure to expand certain domain families more than others. Each domain has an intrinsic combinatorial propensity, and the effects of this have been studied using measures of domain versatility or promiscuity. Next, we study the principles of protein domain architecture evolution and how these have been inferred from distributions of extant domain arrangements. Following this, we review inferences of ancestral domain architecture and the conclusions concerning domain architecture evolution mechanisms that can be drawn from these. Finally, we examine whether all known cases of a given domain architecture can be assumed to have a single common origin (monophyly) or have evolved convergently (polyphyly).

Background. As orthologous proteins are expected to retain function more often than other homologs, they are often used for functional annotation transfer between species. However, ortholog identification methods do not take into account changes in domain architecture, which are likely to modify a protein's function. By domain architecture we refer to the sequential arrangement of domains along a protein sequence. To assess the level of domain architecture conservation among orthologs, we carried out a large-scale study of such events between human and 40 other species spanning the entire evolutionary range. We designed a score to measure domain architecture similarity and used it to analyze differences in domain architecture conservation between orthologs and paralogs relative to the conservation of primary sequence. We also statistically characterized the extents of different types of domain swapping events across pairs of orthologs and paralogs. Results. The analysis shows that orthologs exhibit greater domain architecture conservation than paralogous homologs, even when differences in average sequence divergence are compensated for, for homologs that have diverged beyond a certain threshold. We interpret this as an indication of a stronger selective pressure on orthologs than paralogs to retain the domain architecture required for the proteins to perform a specific function. In general, orthologs as well as the closest paralogous homologs have very similar domain architectures, even at large evolutionary separation. The most common domain architecture changes observed in both ortholog and paralog pairs involved insertion/deletion of new domains, while domain shuffling and segment duplication/deletion were very infrequent. Conclusions. On the whole, our results support the hypothesis that function conservation between orthologs demands higher domain architecture conservation than other types of homologs, relative to primary sequence conservation. This supports th

Orthology is one of the most important tools available to modern biology, as it allows making inferences from easily studied model systems to much less tractable systems of interest, such as ourselves. This becomes important not least in the study of genetic diseases. We here review work on the orthology of disease-associated genes and also present an updated version of the InParanoid-based disease orthology database and web site OrthoDisease, with 14-fold increased species coverage since the previous version. Using this resource, we survey the taxonomic distribution of orthologs of human genes involved in different disease categories. The hypothesis that paralogs can mask the effect of deleterious mutations predicts that known heritable disease genes should have fewer close paralogs. We found large-scale support for this hypothesis as significantly fewer duplications were observed for disease genes in the OrthoDisease ortholog groups.

There is a great need for standards in the orthology field. Users must contend with different ortholog data representations from each provider, and the providers themselves must independently gather and parse the input sequence data. These burdensome and redundant procedures make data comparison and integration difficult. We have designed two XML-based formats, SeqXML and OrthoXML, to solve these problems. SeqXML is a lightweight format for sequence records the input for orthology prediction. It stores the same sequence and metadata as typical FASTA format records, but overcomes common problems such as unstructured metadata in the header and erroneous sequence content. XML provides validation to prevent data integrity problems that are frequent in FASTA files. The range of applications for SeqXML is broad and not limited to ortholog prediction. We provide read/write functions for BioJava, BioPerl, and Biopython. OrthoXML was designed to represent ortholog assignments from any source in a consistent and structured way, yet cater to specific needs such as scoring schemes or meta-information. A unified format is particularly valuable for ortholog consumers that want to integrate data from numerous resources, e. g. for gene annotation projects. Reference proteomes for 61 organisms are already available in SeqXML, and 10 orthology databases have signed on to OrthoXML. Adoption by the entire field would substantially facilitate exchange and quality control of sequence and orthology information., authorCount :4

Background: With the wealth of genomic data available it has become increasingly important to assign putative protein function through functional transfer between orthologs. Therefore, correct elucidation of the evolutionary relationships among genes is a critical task, and attempts should be made to further improve the phylogenetic inference by adding relevant discriminating features. It has been shown that introns can maintain their position over long evolutionary timescales. For this reason, it could be possible to use conservation of intron positions as a discriminating factor when assigning orthology. Therefore, we wanted to investigate whether orthologs have a higher degree of intron position conservation (IPC) compared to non-orthologous sequences that are equally similar in sequence. Results: To this end, we developed a new score for IPC and applied it to ortholog groups between human and six other species. For comparison, we also gathered the closest non-orthologs, meaning sequences close in sequence space, yet falling just outside the ortholog cluster. We found that ortholog-ortholog gene pairs on average have a significantly higher degree of IPC compared to ortholog-closest non-ortholog pairs. Also pairs of inparalogs were found to have a higher IPC score than inparalog-closest non-inparalog pairs. We verified that these differences can not simply be attributed to the generally higher sequence identity of the ortholog-ortholog and the inparalog-inparalog pairs. Furthermore, we analyzed the agreement between IPC score and the ortholog score assigned by the InParanoid algorithm, and found that it was consistently high for all species comparisons. In a minority of cases, the IPC and InParanoid score ranked inparalogs differently. These represent cases where sequence and intron position divergence are discordant. We further analyzed the discordant clusters to identify any possible preference for protein functions by looking for enriched GO terms and Pfam prot, authorCount :3

Genes involved in cancer susceptibility and progression can serve as templates for searching protein networks for novel cancer genes. To this end, we introduce a general network searching method, MaxLink, and apply it to find and rank cancer gene candidates by their connectivity to known cancer genes. Using a comprehensive protein interaction network, we searched for genes connected to known cancer genes. First, we compiled a new set of 812 genes involved in cancer, more than twice the number in the Cancer Gene Census. Their network neighbors were then extracted. This candidate list was refined by selecting genes with unexpectedly high levels of connectivity to cancer genes and without previous association to cancer. This produced a list of 1891 new cancer candidates with up to 55 connections to known cancer genes. We validated our method by cross-validation, Gene Ontology term bias, and differential expression in cancer versus normal tissue. An example novel cancer gene candidate is presented with detailed analysis of the local network and neighbor annotation. Our study provides a ranked list of high priority targets for further studies in cancer research. Supplemental material is included., authorCount :3

Background: In order to generate hypotheses regarding the mechanisms by which 2,3,7,8-tetrachlorodibenzo-p-dioxin (dioxin) causes toxicity, we analyzed global gene expression changes in developing zebrafish embryos exposed to this potent toxicant in the context of a dynamic gene network. For this purpose, we also computationally inferred a zebrafish (Danio rerio) interactome based on orthologs and interaction data from other eukaryotes. Methodology/Principal Findings: Using novel computational tools to analyze this interactome, we distinguished between dioxin-dependent and dioxin-independent interactions between proteins, and tracked the temporal propagation of dioxin-dependent transcriptional changes from a few genes that were altered initially, to large groups of biologically coherent genes at later times. The most notable processes altered at later developmental stages were calcium and iron metabolism, embryonic morphogenesis including neuronal and retinal development, a variety of mitochondria-related functions, and generalized stress response (not including induction of antioxidant genes). Within the interactome, many of these responses were connected to cytochrome P4501A (cyp1a) as well as other genes that were dioxin-regulated one day after exposure. This suggests that cyp1a may play a key role initiating the toxic dysregulation of those processes, rather than serving simply as a passive marker of dioxin exposure, as suggested by earlier research. Conclusions/Significance: Thus, a powerful microarray experiment coupled with a flexible interactome and multi-pronged interactome tools (which are now made publicly available for microarray analysis and related work) suggest the hypothesis that dioxin, best known in fish as a potent cardioteratogen, has many other targets. Many of these types of toxicity have been observed in mammalian species and are potentially caused by alterations to cyp1a., authorCount :7

Transmembrane (TM) proteins are proteins that span a biological membrane one or more times. As their 3-D structures are hard to determine, experiments focus on identifying their topology (i. e. which parts of the amino acid sequence are buried in the membrane and which are located on either side of the membrane), but only a few topologies are known. Consequently, various computational TM topology predictors have been developed, but their accuracies are far from perfect. The prediction quality can be improved by applying a consensus approach, which combines results of several predictors to yield a more reliable result. RESULTS: A novel TM consensus method, named MetaTM, is proposed in this work. MetaTM is based on support vector machine models and combines the results of six TM topology predictors and two signal peptide predictors. On a large data set comprising 1460 sequences of TM proteins with known topologies and 2362 globular protein sequences it correctly predicts 86.7% of all topologies. CONCLUSION: Combining several TM predictors in a consensus prediction framework improves overall accuracy compared to any of the individual methods. Our proposed SVM-based system also has higher accuracy than a previous consensus predictor. MetaTM is made available both as downloadable source code and as DAS server at http://MetaTM.sbc.su.se.

No single experimental method can discover all connections in the interactome. A computational approach can help by integrating data from multiple, often unrelated, proteomics and genomics pipelines. Reconstructing global networks of functional coupling (FC) faces the challenges of scale and heterogeneity--how to efficiently integrate huge amounts of diverse data from multiple organisms, yet ensuring high accuracy. We developed FunCoup, an optimized Bayesian framework, to resolve these issues. Because interactomes comprise functional coupling of many types, FunCoup annotates network edges with confidence scores in support of different kinds of interactions: physical interaction, protein complex member, metabolic, or signaling link. This capability boosted overall accuracy. On the whole, the constructed framework was comprehensively tested to optimize the overall confidence and ensure seamless, automated incorporation of new data sets of heterogeneous types. Using over 50 data sets in seven organisms and extensively transferring information between orthologs, FunCoup predicted global networks in eight eukaryotes. For the Ciona intestinalis network, only orthologous information was used, and it recovered a significant number of experimental facts. FunCoup predictions were validated on independent cancer mutation data. We show how FunCoup can be used for discovering candidate members of the Parkinson and Alzheimer pathways. Cross-species pathway conservation analysis provided further support to these observations.

RISC (RNA-induced silencing complex) is a central protein complex in RNAi, into which a siRNA strand is assembled to become effective in gene silencing. By using an in vitro RNAi reaction based on Drosophila embryo extract, an asymmetric model was recently proposed for RISC assembly of siRNA strands, suggesting that the strand that is more loosely paired at its 5' end is selectively assembled into RISC and results in target gene silencing. However, in the present study, we were unable to establish such a correlation in cell-based RNAi assays, as well as in large-scale RNAi data analyses. This suggests that the thermodynamic stability of siRNA is not a major determinant of gene silencing in mammalian cells. Further studies on fork siRNAs showed that mismatch at the 5' end of the siRNA sense strand decreased RISC assembly of the antisense strand, but surprisingly did not increase RISC assembly of the sense strand. More interestingly, measurements of melting temperature showed that the terminal stability of fork siRNAs correlated with the positions of the mismatches, but not gene silencing efficacy. In summary, our data demonstrate that there is no definite correlation between siRNA stability and gene silencing in mammalian cells, which suggests that instead of thermodynamic stability, other features of the siRNA duplex contribute to RISC assembly in RNAi.

Artificially synthesized short interfering RNAs (siRNAs) are widely used in functional genomics to knock down specific target genes. One ongoing challenge is to guarantee that the siRNA does not elicit off-target effects. Initial reports suggested that siRNAs were highly sequence-specific; however, subsequent data indicates that this is not necessarily the case. It is still uncertain what level of similarity and other rules are required for an off-target effect to be observed, and scoring schemes have not been developed to look beyond simple measures such as the number of mismatches or the number of consecutive matching bases present. We created design rules for predicting the likelihood of a non-specific effect and present a web server that allows the user to check the specificity of a given siRNA in a flexible manner using a combination of methods. The server finds potential off-target matches in the corresponding RefSeq database and ranks them according to a scoring system based on experimental studies of specificity. AVAILABILITY: The server is available at http://informatics-eskitis.griffith.edu.au/SpecificityServer.

When using conventional transmembrane topology and signal peptide predictors, such as TMHMM and SignalP, there is a substantial overlap between these two types of predictions. Applying these methods to five complete proteomes, we found that 30-65% of all predicted signal peptides and 25-35% of all predicted transmembrane topologies overlap. This impairs predictions of 5-10% of the proteome, hence this is an important issue in protein annotation. To address this problem, we previously designed a hidden Markov model, Phobius, that combines transmembrane topology and signal peptide predictions. The method makes an optimal choice between transmembrane segments and signal peptides, and also allows constrained and homology-enriched predictions. We here present a web interface (http://phobius.cgb.ki.se and http://phobius.binf.ku.dk) to access Phobius., QC 20150624

PfamAlyzer is a Java applet that enables exploration of Pfam domain architectures using a user-friendly graphical interface. It can search the UniProt protein database for a domain pattern. Domain patterns similar to the query are presented graphically by PfamAlyzer either in a ranked list or pinned to the tree of life. Such domain-centric homology search can assist identification of distant homologs with shared domain architecture., authorCount :2

G protein-coupled receptors (GPCRs) constitute a large superfamily involved in various types of signal transduction pathways triggered by hormones, odorants, peptides, proteins, and other types of ligands. The superfamily is so diverse that many members lack sequence similarity, although they all span the cell membrane seven times with an extracellular N and a cytosolic C terminus. We analyzed a divergent set of GPCRs and found distinct loop length patterns and differences in amino acid composition between cytosolic loops, extracellular loops, and membrane regions. We configured GPCRHMM, a hidden Markov model, to fit those features and trained it on a large dataset representing the entire superfamily. GPCRHMM was benchmarked to profile HMMs and generic transmembrane detectors on sets of known GPCRs and non-GPCRs. In a cross-validation procedure, profile HMMs produced an error rate nearly twice as high as GPCRHMM. In a sensitivity-selectivity test, GPCRHMM's sensitivity was about 15% higher than that of the best transmembrane predictors, at comparable false positive rates. We used GPCRHMM to search for novel members of the GPCR superfamily in five proteomes. All in all we detected 120 sequences that lacked annotation and are potentially novel GPCRs. Out of those 102 were found in Caenorhabditis elegans, four in human, and seven in mouse. Many predictions (65) belonged to Pfam domains of unknown function. GPCRHMM strongly rejected a family of arthropod-specific odorant receptors believed to be GPCRs. A detailed analysis showed that these sequences are indeed very different from other GPCRs. GPCRHMM is available at http://gpcrhmm.cgb.ki.se., QC 20111128

An inherent problem in transmembrane protein topology prediction and signal peptide prediction is the high similarity between the hydrophobic regions of a transmembrane helix and that of a signal peptide, leading to cross-reaction between the two types of predictions. To improve predictions further, it is therefore important to make a predictor that aims to discriminate between the two classes. In addition, topology information can be gained when successfully predicting a signal Peptide leading a trans' membrane protein since it dictates that the N terminus of the mature protein must be on the non-cytoplasmic side of the membrane. Here, we present Phobius, a combined transmembrane protein topology and signal peptide predictor. The predictor is based on a hidden Markov model (HMM) that models the different sequence regions of a signal peptide and the different regions of a transmembrane protein in a series of interconnected states. Training was done on a newly assembled and curated dataset. Compared to TMHMM and SignalP, errors coming from cross-prediction between transmembrane segments and signal peptides were reduced substantially by Phobius. False classifications of signal peptides were reduced from 26.1% to 3.9% and false classifications of transmembrane helices were reduced from 19.0%, to 7.7%. Phobius was applied to the proteomes of Honzo sapiens and Escherichia coli. Here we also noted a drastic reduction of false classifications compared to TMHMM/SignalP, suggesting that Phobius is well suited for whole-genome annotation of signal peptides and transmembrane regions. The method is available at http://phobius.cgb.ki.se/ as well as at http://phobius.binf.ku.dk/., QC 20111124

Transmembrane prediction methods are generally benchmarked on a set of proteins with experimentally verified topology. We have investigated if the accuracy measured on such datasets can be expected in an unbiased genomic analysis, or if there is a bias towards 'easily predictable' proteins in the benchmark datasets. As a measurement of accuracy, the concordance of the results from five different prediction methods was used (TMHMM, PHD, HMMTOP, MEMSAT, and TOPPRED). The benchmark dataset showed significantly higher levels (up to five times) of agreement between different methods than in 10 tested genomes. We have also analyzed which programs are most prone to make mispredictions by measuring the frequency of one-out-of-five disagreeing predictions., QC 20111124