Membrane proteins are unique in that they interact with lipid bilayers, making them indispensable for transporting molecules and relaying signals between and across cells. Due to the significance of the protein's functions, mutations often have profound effects on the fitness of the host. This is apparent both from experimental studies, which implicated numerous missense variants in diseases, as well as from evolutionary signals that allow elucidating the physicochemical constraints that intermembrane and aqueous environments bring. In this review, we report on the current state of knowledge acquired on missense variants (referred to as to single amino acid variants) affecting membrane proteins as well as the insights that can be extrapolated from data already available. This includes an overview of the annotations for membrane protein variants that have been collated within databases dedicated to the topic, bioinformatics approaches that leverage evolutionary information in order to shed light on previously uncharacterized membrane protein structures or interaction interfaces, tools for predicting the effects of mutations tailored specifically towards the characteristics of membrane proteins as well as two clinically relevant case studies explaining the implications of mutated membrane proteins in cancer and cardiomyopathy.

SUMMARY: The ability of a T cell to recognize foreign peptides is defined by a single α and a single β hypervariable complementarity determining region (CDR3), which together form the T-cell receptor (TCR) heterodimer. In ∼30-35% of T cells, two α chains are expressed at the mRNA level but only one α chain is part of the functional TCR. This effect can also be observed for β chains, although it is less common. The identification of functional α/β chain pairs is instrumental in high-throughput characterization of therapeutic TCRs. TCRpair is the first method that predicts whether an α and β chain pair forms a functional, HLA-A*02:01 specific TCR without requiring the sequence of a recognized peptide. By taking additional amino acids flanking the CDR3 regions into account, TCRpair achieves an AUC of 0.71. AVAILABILITY AND IMPLEMENTATION: TCRpair is implemented in Python using TensorFlow 2.0 and is freely available at https://www.github.com/amoesch/TCRpair. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: This study is motivated by the following three considerations: a) the physico-chemical properties of transmembrane (TM) proteins are distinctly different from those of globular proteins, necessitating the development of specialized structure prediction techniques, b) for many structural features no specialized predictors for TM proteins are available at all, and c) deep learning algorithms allow to automate the feature engineering process and thus facilitate the development of multi-target methods for predicting several protein properties at once. RESULTS: We present AllesTM, an integrated tool to predict almost all structural features of transmembrane proteins that can be extracted from atomic coordinate data. It blends several machine learning algorithms: random forests and gradient boosting machines, convolutional neural networks in their original form as well as those enhanced by dilated convolutions and residual connections, and, finally, long short-term memory architectures. AllesTM outperforms other available methods in predicting residue depth in the membrane, flexibility, topology, relative solvent accessibility in its bound state, while in torsion angles, secondary structure and monomer relative solvent accessibility prediction it lags only slightly behind the currently leading technique SPOT-1D. High accuracy on a multitude of prediction targets and easy installation make AllesTM a one-stop shop for many typical problems in the structural bioinformatics of transmembrane proteins. CONCLUSIONS: In addition to presenting a highly accurate prediction method and eliminating the need to install and maintain many different software tools, we also provide a comprehensive overview of the impact of different machine learning algorithms and parameter choices on the prediction performance. AllesTM is freely available at https://github.com/phngs/allestm.

Motivation: Existing sources of experimental mutation data do not consider the structural environment of amino acid substitutions and distinguish between soluble and membrane proteins. They also suffer from a number of further limitations, including data redundancy, lack of disease classification, incompatible information content, and ambiguous annotations (e.g. the same mutation being annotated as disease and benign). Results: We have developed a novel database, MutHTP, which contains information on 183 395 disease-associated and 17 827 neutral mutations in human transmembrane proteins. For each mutation site MutHTP provides a description of its location with respect to the membrane protein topology, structural environment (if available) and functional features. Comprehensive visualization, search, display and download options are available. Availability and implementation: The database is publicly available at http://www.iitm.ac.in/bioinfo/MutHTP/. The website is implemented using HTML, PHP and javascript and supports recent versions of all major browsers, such as Firefox, Chrome and Opera. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: The V3 loop of the gp120 glycoprotein of the Human Immunodeficiency Virus 1 (HIV-1) is considered to be responsible for viral coreceptor tropism. gp120 interacts with the CD4 receptor of the host cell and subsequently V3 binds either CCR5 or CXCR4. Due to the fact that the CCR5 coreceptor is targeted by entry inhibitors, a reliable prediction of the coreceptor usage of HIV-1 is of great interest for antiretroviral therapy. Although several methods for the prediction of coreceptor tropism are available, almost all of them have been developed based on only subtype B sequences, and it has been shown in several studies that the prediction of non-B sequences, in particular subtype A sequences, are less reliable. Thus, the aim of the current study was to develop a reliable prediction model for subtype A viruses. Results: Our new model SCOTCH is based on a stacking approach of classifier ensembles and shows a significantly better performance for subtype A sequences compared to other available models. In particular for low false positive rates (between 0.05 and 0.2, i.e. recommendation in the German and European Guidelines for tropism prediction), SCOTCH shows significantly better prediction performances in terms of partial area under the curves and diagnostic odds ratios compared to existing tools, and thus can be used to reliably predict coreceptor tropism for subtype A sequences. Availability and implementation: SCOTCH can be downloaded/accessed at http://www.heiderlab.de.

Translation of consecutive prolines causes ribosome stalling, which is alleviated but cannot be fully compensated by the elongation factor P. However, the presence of polyproline motifs in about one third of the E. coli proteins underlines their potential functional importance, which remains largely unexplored. We conducted an evolutionary analysis of polyproline motifs in the proteomes of 43 E. coli strains and found evidence of evolutionary selection against translational stalling, which is especially pronounced in proteins with high translational efficiency. Against the overall trend of polyproline motif loss in evolution, we observed their enrichment in the vicinity of translational start sites, in the inter-domain regions of multi-domain proteins, and downstream of transmembrane helices. Our analysis demonstrates that the time gain caused by ribosome pausing at polyproline motifs might be advantageous in protein regions bracketing domains and transmembrane helices. Polyproline motifs might therefore be crucial for co-translational folding and membrane insertion.

MOTIVATION: Cross-reactivity (CR) or invocation of autoimmune side effects in various tissues has important safety implications in adoptive immunotherapy directed against selected antigens. The ability to predict CR (on-target and off-target toxicities) may help in the early selection of safer therapeutically relevant target antigens. RESULTS: We developed a methodology for the calculation of quantitative CR for any defined peptide epitope. Using this approach, we performed assessment of 4 groups of 283 currently known human MHC-class-I epitopes including differentiation antigens, overexpressed proteins, cancer-testis antigens and mutations displayed by tumor cells. In addition, 89 epitopes originating from viral sources were investigated. The natural occurrence of these epitopes in human tissues was assessed based on proteomics abundance data, while the probability of their presentation by MHC-class-I molecules was modelled by the method of Keşmir et al. which combines proteasomal cleavage, TAP affinity and MHC-binding predictions. The results of these analyses for many previously defined peptides are presented as CR indices and tissue profiles. The methodology thus allows for quantitative comparisons of epitopes and is suggested to be suited for the assessment of epitopes of candidate antigens in an early stage of development of adoptive immunotherapy. AVAILABILITY AND IMPLEMENTATION: Our method is implemented as a Java program, with curated datasets stored in a MySQL database. It predicts all naturally possible self-antigens for a given sequence of a therapeutic antigen (or epitope) and after filtering for predicted immunogenicity outputs results as an index and profile of CR to the self-antigens in 22 human tissues. The program is implemented as part of the iCrossR webserver, which is publicly available at http://webclu.bio.wzw.tum.de/icrossr/ CONTACT: [email protected] information: Supplementary data are available at Bioinformatics online.

MOTIVATION: Adoptive T cell therapies based on introduction of new T cell receptors (TCRs) into patient recipient T cells is a promising new treatment for various kinds of cancers. A major challenge, however, is the choice of target antigens. If an engineered TCR can cross-react with self-antigens in healthy tissue, the side-effects can be devastating. We present the first web server for assessing epitope sharing when designing new potential lead targets. We enable the users to find all known proteins containing their peptide of interest. The web server returns not only exact matches, but also approximate ones, allowing a number of mismatches of the users choice. For the identified candidate proteins the expression values in various healthy tissues, representing all vital human organs, are extracted from RNA Sequencing (RNA-Seq) data as well as from some cancer tissues as control. All results are returned to the user sorted by a score, which is calculated using well-established methods and tools for immunological predictions. It depends on the probability that the epitope is created by proteasomal cleavage and its affinities to the transporter associated with antigen processing and the major histocompatibility complex class I alleles. With this framework, we hope to provide a helpful tool to exclude potential cross-reactivity in the early stage of TCR selection for use in design of adoptive T cell immunotherapy. AVAILABILITY AND IMPLEMENTATION: The Expitope web server can be accessed via http://webclu.bio.wzw.tum.de/expitope.

BACKGROUND: Viruses are the most abundant and genetically diverse biological entities on earth, yet the repertoire of viral proteins remains poorly explored. As the number of sequenced virus genomes grows into the thousands, and the number of viral proteins into the hundreds of thousands, we report a systematic computational analysis of the point of first-contact between viruses and their hosts, namely viral transmembrane (TM) proteins. RESULTS: The complement of α-helical TM proteins in double-stranded DNA viruses infecting bacteria and archaea reveals large-scale trends that differ from those of their hosts. Viruses typically encode a substantially lower fraction of TM proteins than archaea or bacteria, with the notable exception of viruses with virions containing a lipid component such as a lipid envelope, internal lipid core, or inner membrane vesicle. Compared to bacteriophages, archaeal viruses are substantially enriched in membrane proteins. However, this feature is not always stable throughout the evolution of a viral lineage; for example, TM proteins are not part of the common heritage shared between Lipothrixviridae and Rudiviridae. In contrast to bacteria and archaea, viruses almost completely lack proteins with complicated membrane topologies composed of more than 4 TM segments, with the few detected exceptions being obvious cases of relatively recent horizontal transfer from the host. CONCLUSIONS: The dramatic differences between the membrane proteomes of cells and viruses stem from the fact that viruses do not depend on essential membranes for energy transformation, ion homeostasis, nutrient transport and signaling.

MOTIVATION: Computer-assisted studies of structure, function and evolution of viruses remains a neglected area of research. The attention of bioinformaticians to this interesting and challenging field is far from commensurate with its medical and biotechnological importance. It is telling that out of >200 talks held at ISMB 2013, the largest international bioinformatics conference, only one presentation explicitly dealt with viruses. In contrast to many broad, established and well-organized bioinformatics communities (e.g. structural genomics, ontologies, next-generation sequencing, expression analysis), research groups focusing on viruses can probably be counted on the fingers of two hands. RESULTS: The purpose of this review is to increase awareness among bioinformatics researchers about the pressing needs and unsolved problems of computational virology. We focus primarily on RNA viruses that pose problems to many standard bioinformatics analyses owing to their compact genome organization, fast mutation rate and low evolutionary conservation. We provide an overview of tools and algorithms for handling viral sequencing data, detecting functionally important RNA structures, classifying viral proteins into families and investigating the origin and evolution of viruses.

Phosphorylation at specific residues can activate a protein, lead to its localization to particular compartments, be a trigger for protein degradation and fulfill many other biological functions. Protein phosphorylation is increasingly being studied at a large scale and in a quantitative manner that includes a temporal dimension. By contrast, structural properties of identified phosphorylation sites have so far been investigated in a static, non-quantitative way. Here we combine for the first time dynamic properties of the phosphoproteome with protein structural features. At six time points of the cell division cycle we investigate how the variation of the amount of phosphorylation correlates with the protein structure in the vicinity of the modified site. We find two distinct phosphorylation site groups: intrinsically disordered regions tend to contain sites with dynamically varying levels, whereas regions with predominantly regular secondary structures retain more constant phosphorylation levels. The two groups show preferences for different amino acids in their kinase recognition motifs - proline and other disorder-associated residues are enriched in the former group and charged residues in the latter. Furthermore, these preferences scale with the degree of disorderedness, from regular to irregular and to disordered structures. Our results suggest that the structural organization of the region in which a phosphorylation site resides may serve as an additional control mechanism. They also imply that phosphorylation sites are associated with different time scales that serve different functional needs.

Contact: [email protected] Supplementary information: Supplementary data are available at Bioinformatics online. Data mining is a cornerstone of modern bioinformatics. Techniques such as feature selection or data-driven model (classifier) building are used broadly in many fields including gene expression data analysis (Wood et al., 2007), proteomics (Barla et al., 2008), secondary and tertiary protein structure prediction (Heringa, 2000; Kryshtafovych et al., 2007), prediction of experimental behavior of proteins or other molecules (Liao et al., 2006; Smialowski et al., 2007) and medical science (Patel and Goyal, 2007). It is widely accepted that any model building requires stringent evaluation of its performance and generalization capabilities using well-established methods such as k-fold cross-validation, leave-one-out cross-validation, bootstrap and others (Frank et al., 2004). The growing complexity of the data and the urge to improve available methods have led to the development of procedures combining feature selection and model building. Feature selection is often used to limit the amount and dimensionality of the data or to select features that correlate well with the target class. Feature selection methods can be subdivided into those that are unsupervised, i.e. unaware of class attributes [e.g. removal of a feature with the same constant values throughout the whole dataset, PCA (principal component analysis), MF (matrix factorization)] and those that are supervised, i.e. driven by class information. The latter group includes filter methods using, e.g. information gain as well as the Wrapper approach (Witten and Frank, 2005). Special consideration is required when supervised feature selection is used to construct input data for model building (classification). In order to correctly evaluate classifiers built on such projected data, the entire procedure including feature selection and model training has to be evaluated against independent data. In other words, the test set must not be used for supervised feature selection (or more generally, supervised preprocessing). Otherwise, estimates of the classifiers’ performance will be over optimistic (Ambroise and McLachlan, 2002; Efron, 2005; Kohavi and John, 1997; Molinaro et al., 2005; Reunane, 2004). Any model building method integrated with feature selection must be externally evaluated. Evaluation must include both: supervised feature selection and classification and should not be limited to classification only (Fig. 1). The correct [(A) followed (C)] and the incorrect [(B) followed by (C)] procedure for combining supervised feature selection and learning a classifier. In the figure, processes and products are depicted by ellipses and rectangles, respectively. Training and test sets consist of features X and a target attribute Y (to be predicted). X′ is a subset of features reduced by supervised feature selection, f() is a classifier and Ŷ contains the prediction of Y values by this function. (A and B) show the workflows for the correct and incorrect application of supervised feature selection, and (C) holds the evaluation workflow (more description in the text). The correct procedure is depicted in Figure 1 and consists of the workflow from Figure 1A for training and Figure 1C for testing (evaluation). In a preprocessing step, supervised feature selection reduces the set of features X to a subset X′ (Y is the target attribute). Subsequently, the reduced training set is used to infer a classifier f(). During testing (Fig. 1C), the trained classifier f() is evaluated using an independent test set with the feature space reduced to X′ according to the feature selection derived in the previous step (Fig. 1A). The classifier predicts Ŷ for each instance. Various performance measures can then be calculated by comparing the predictions Ŷ with the true values for Y. In contrast with this procedure, the most common mistake (according to our observation) of machine learning applications in the life sciences is illustrated in Figure 1B: both training and test sets are used for supervised feature selection. After that, a classifier f() is learned from the reduced training set as before. During testing, this classifier is applied to the test set as described above (Fig. 1C). However, as evident from the illustration, information from the test set has already been used for the inference of the classifier, by choosing an appropriate subset of features for the learning algorithm (Fig. 1B). Therefore, class information from the test set has leaked to the training phase. This analysis remains true for classification and regression and regardless whether training and test sets are created as part of a k-fold cross-validation or any other method. Recently, it has repeatedly come to our attention while reading or reviewing manuscripts submitted to Bioinformatics and other journals that often no proper external evaluations of the whole method consisting of supervised feature selection and subsequent classification is provided. In this brief note we would like to demonstrate the consequences of this type of a mistake for classification performance estimation using data where attribute values are generated randomly as well as data with random class assignments. We also estimate the consequences of such erroneous overoptimism as a function of the dataset size. For simplicity, we focus only on classification and leave out regression methods. Three types of datasets were used in this study. All have 21 attributes: 20 frequencies of amino acids and one class attribute assigning each instance to one of two classes. Datasets of the first type (randomly generated attributes) contain real random attribute values resembling natural frequency distributions of 20 amino acids. To generate randomly perturbated frequencies of the 20 amino acids, we used Gaussian random numbers with peak maximums at their natural occurrence frequencies (as provided at http://prowl.rockefeller.edu/aainfo/struct.htm) and standard deviation (SD) equal to 0.25 of this value. All negative values were set to 0 and the sum of 20 frequencies for a given instance was not allowed to be greater than 100. Datasets of the second type (randomize class data) consist of instances picked randomly from a dataset of 1000 proteins of which a half showed good solubility upon heterologous expression in Escherichia coli, while the other half was notoriously insoluble. Classes of these instances—soluble or non-soluble—were assigned randomly. Third type of datasets (real data) was same as second except that real class labels were preserved. Quantum Random Bit Generator Service (QRBG; Stevanovic et al., 2008) was used as a source of seeds for random numbers. We maintained even class distribution at all times. For selecting the best features, the following methods were used: Wrapper (Kohavi and John, 1997), Relief Attribute Evaluation with Ranker (Witten and Frank, 2005) and PCA (Pearson, 1901). The Wrapper method takes into account class information by evaluating feature sets based on the performance of the classifier. Hence, the resulting feature set is tailored to a given classification method. In our comparison, the Wrapper method is the most ‘aggressive’ feature selection method. It was setup to use Naive Bayes for classification and Best First for attribute space search (Witten and Frank, 2005). The Relief method is also supervised, but does not optimize feature sets directly for classifier performance. Thus, it takes into account class information in a ‘less aggressive’ manner than the Wrapper method. The threshold of Ranker coupled to the Relief Attribute Evaluation method was set to zero. PCA (Jolliffe, 2002) is an unsupervised feature selection method and hence does not take into account class information at all. PCA dimensionality reduction was accomplished by the following steps: data normalization, calculation of orthonormal vectors [PC (principal components)], sorting PC according to decreasing variance. We save PC accounting cumulatively for 95% of the data variance (Jolliffe, 2002). Naive Bayes and nearest neighbor IB1 (Aha and Kibler, 1991) algorithms were used for classification, because they are among the simplest and most fundamental classification methods. Classifiers were trained and evaluated using 10-fold cross-validation. MCC (Matthew's correlation coefficient) and AUROC (area under receiver operating curve) were calculated to measure classifier performance. For each dataset size, feature selection and classification algorithm the whole procedure starting from data construction was repeated 30 times. Overfitting (overoptimism) was measured as the difference in classification performance (denoted by ΔAUROC in the following) on data after and before feature selection, averaged over 30 trials. Information density is defined as the number of instances per number of attributes. For datasets resulting from feature selection, we calculate the information density ratio (ID ratio, the factor by which the number of features is reduced) to measure loss of information. We generated a range of datasets containing between 10 and 1000 instances. For both types of randomize datasets, half of the data were tagged with the ‘no’ class and the rest with ‘yes’ to simulate a two class problem with an even class distribution. We found that using supervised feature selection with dataset containing both training and test sets leads to a significant overoptimistic assessment of classifier performance. This holds true for datasets with randomly generated attribute values, randomized class data (Fig. 2) and real data (Supplementary Fig. 1, Supplementary Table 1). The extent of overfitting depends on the dataset size. The strongest overfitting (AUROC increase ∼0.5) was observed for the smallest real datasets with random class labels using the Wrapper method. Classification on randomly generated attribute data exhibits a slightly smaller increase (∼0.4) (data with real class labels also reached 0.4). These values are extremely high considering that the AUROC ranges from 0.5 for random guessing to 1.0 for the best theoretically possible model. Improper use of feature selection also distorts classifier performance when larger datasets are used. Even for datasets with 1000 instances we observed slight albeit statistically significant increase in AUROC values (0.04) (whiskers in Fig. 2 mark 95% confidence intervals). As seen in Figure 2, the use of both the Wrapper and the Relief feature selection algorithms results in falsified measures of classifier performance. Application of PCA which is not supervised did not lead to substantial overfitting on any tested dataset. Relation between the number of instances and the extent of overfitting caused by feature selection as measured by AUROC growth. (A) Randomly generated attribute values, (B) randomly tagged real data. Three different feature selection algorithms were used: Wrapper (hashed bars), Relief Attribute Evaluation (white bars), PCA (black bars). Whiskers mark 95% confidence intervals. Feature selection reduces the amount of information by lowering the number of attributes and thus leads to an increase of the ratio between the number of instances and the number of attributes. We assessed the extent of information loss by calculating the ID ratio between datasets before and after feature selection as described above. A higher ID ratio means that a higher percentage of features was removed. For example, ID ratio = 2, means that after feature selection the number of attributes is reduced by two. For each of the feature selection method, we found that the more features got removed, the stronger were the effects of overfitting (Fig. 3). Interestingly, there are also differences between attribute selection algorithms. Wrapper seems to be less prone to overfitting compared with the Relief method. By almost the same magnitude of overfitting (increase in AUROC 0.25), it reduces the information content more effectively (4.5 times) compared with Relief (2.5 times). Wrappers reduce the number of features stronger than Relief method possibly because it comprehensively evaluates also combinations of features, selecting features based on not only their own relevancy but also redundancy (Kohavi and John, 1997). Moreover, Wrapper efficiency in feature selection depends on internally employed classifier and search algorithms. Data with randomly generated attribute values showed higher resistance against classification bias permitting a greater decrease in the feature number compared with real data with randomized class tags and real data. PCA did not cause overfitting while still being able to slightly reduce the number of attributes of all data types. Similar results were obtained when MCC was used instead of AUROC and also when the whole analysis was repeated using the IB1 nearest neighbor classification method (Supplementary Table 1, Data not shown). Relation between information loss and overfitting measured by ΔAUROC growth. (A) Randomly generated attribute values, (B) randomly tagged real data. Three feature selection methods were examined: Wrapper (black circles), Relief Attribute Evaluation (open squares) and PCA (gray triangles). Lines were fitted by linear regression: solid lines to Wrapper and dashed to Relief Attribute Evaluation data points. ID ratio is the information density ratio. In summary, we show that supervised feature selection coupled with or integral to model building (classification) requires external evaluation. Failure in setting up evaluation with external data (Fig. 1B) leads to overoptimism in the assessment of classifier performance. It is reasonable to expect that overfitting will be at least in the same order of magnitude as what is observed on randomly generated data for most of the real life classification tasks, where amino acid composition is used as input. Real data may be more prone to overfitting because they are likely to contain more patterns than randomly generated data. Even if patterns are orthogonal to the problem under consideration overfitting can still occur. For this reason, whenever model building (classification) is integrated with supervised attribute selection, it is crucial to evaluate classifiers with the data not used for attribute selection (Fig. 1 A and C). Conflict of Interest: none declared.

BACKGROUND: Virtually all currently available microRNA target site prediction algorithms require the presence of a (conserved) seed match to the 5' end of the microRNA. Recently however, it has been shown that this requirement might be too stringent, leading to a substantial number of missed target sites. RESULTS: We developed TargetSpy, a novel computational approach for predicting target sites regardless of the presence of a seed match. It is based on machine learning and automatic feature selection using a wide spectrum of compositional, structural, and base pairing features covering current biological knowledge. Our model does not rely on evolutionary conservation, which allows the detection of species-specific interactions and makes TargetSpy suitable for analyzing unconserved genomic sequences.In order to allow for an unbiased comparison of TargetSpy to other methods, we classified all algorithms into three groups: I) no seed match requirement, II) seed match requirement, and III) conserved seed match requirement. TargetSpy predictions for classes II and III are generated by appropriate postfiltering. On a human dataset revealing fold-change in protein production for five selected microRNAs our method shows superior performance in all classes. In Drosophila melanogaster not only our class II and III predictions are on par with other algorithms, but notably the class I (no-seed) predictions are just marginally less accurate. We estimate that TargetSpy predicts between 26 and 112 functional target sites without a seed match per microRNA that are missed by all other currently available algorithms. CONCLUSION: Only a few algorithms can predict target sites without demanding a seed match and TargetSpy demonstrates a substantial improvement in prediction accuracy in that class. Furthermore, when conservation and the presence of a seed match are required, the performance is comparable with state-of-the-art algorithms. TargetSpy was trained on mouse and performs well in human and drosophila, suggesting that it may be applicable to a broad range of species. Moreover, we have demonstrated that the application of machine learning techniques in combination with upcoming deep sequencing data results in a powerful microRNA target site prediction tool http://www.targetspy.org.

Point mutations resulting in the substitution of a single amino acid can cause severe functional consequences, but can also be completely harmless. Understanding what determines the phenotypical impact is important both for planning targeted mutation experiments in the laboratory and for analyzing naturally occurring mutations found in patients. Common wisdom suggests using the extent of evolutionary conservation of a residue or a sequence motif as an indicator of its functional importance and thus vulnerability in case of mutation. In this work, we put forward the hypothesis that in addition to conservation, co-evolution of residues in a protein influences the likelihood of a residue to be functionally important and thus associated with disease. While the basic idea of a relation between co-evolution and functional sites has been explored before, we have conducted the first systematic and comprehensive analysis of point mutations causing disease in humans with respect to correlated mutations. We included 14,211 distinct positions with known disease-causing point mutations in 1,153 human proteins in our analysis. Our data show that (1) correlated positions are significantly more likely to be disease-associated than expected by chance, and that (2) this signal cannot be explained by conservation patterns of individual sequence positions. Although correlated residues have primarily been used to predict contact sites, our data are in agreement with previous observations that (3) many such correlations do not relate to physical contacts between amino acid residues. Access to our analysis results are provided at http://webclu.bio.wzw.tum.de/~pagel/supplements/correlated-positions/.

MOTIVATION: The analysis of co-evolving residues has been exhaustively evaluated for the prediction of intramolecular amino acid contacts in soluble proteins. Although a variety of different methods for the detection of these co-evolving residues have been developed, the fraction of correctly predicted contacts remained insufficient for their reliable application in the construction of structural models. Membrane proteins, which constitute between one-fourth and one-third of all proteins in an organism, were only considered in few individual case studies. RESULTS: We present the first general study of correlated mutations in alpha-helical membrane proteins. Using seven different prediction algorithms, we extracted co-evolving residues for 14 membrane proteins having a solved 3D structure. On average, distances between correlated pairs of residues lying on different transmembrane segments were found to be significantly smaller compared to a random prediction. Covariation of residues was frequently found in direct sequence neighborhood to helix-helix contacts. Based on the results obtained from individual prediction methods, we constructed a consensus prediction for every protein in the dataset that combines obtained correlations from different prediction algorithms and simultaneously removes likely false positives. Using this consensus prediction, 53% of all predicted residue pairs were found within one helix turn of an observed helix-helix contact. Based on the combination of co-evolving residues detected with the four best prediction algorithms, interacting helices could be predicted with a specificity of 83% and sensitivity of 42%. AVAILABILITY: http://webclu.bio.wzw.tum.de/helixcorr/

MOTIVATION: Obtaining soluble proteins in sufficient concentrations is a recurring limiting factor in various experimental studies. Solubility is an individual trait of proteins which, under a given set of experimental conditions, is determined by their amino acid sequence. Accurate theoretical prediction of solubility from sequence is instrumental for setting priorities on targets in large-scale proteomics projects. RESULTS: We present a machine-learning approach called PROSO to assess the chance of a protein to be soluble upon heterologous expression in Escherichia coli based on its amino acid composition. The classification algorithm is organized as a two-layered structure in which the output of primary support vector machine (SVM) classifiers serves as input for a secondary Naive Bayes classifier. Experimental progress information from the TargetDB database as well as previously published datasets were used as the source of training data. In comparison with previously published methods our classification algorithm possesses improved discriminatory capacity characterized by the Matthews Correlation Coefficient (MCC) of 0.434 between predicted and known solubility states and the overall prediction accuracy of 72% (75 and 68% for positive and negative class, respectively). We also provide experimental verification of our predictions using solubility measurements for 31 mutational variants of two different proteins.

BACKGROUND: Unsupervised annotation of proteins by software pipelines suffers from very high error rates. Spurious functional assignments are usually caused by unwarranted homology-based transfer of information from existing database entries to the new target sequences. We have previously demonstrated that data mining in large sequence annotation databanks can help identify annotation items that are strongly associated with each other, and that exceptions from strong positive association rules often point to potential annotation errors. Here we investigate the applicability of negative association rule mining to revealing erroneously assigned annotation items. RESULTS: Almost all exceptions from strong negative association rules are connected to at least one wrong attribute in the feature combination making up the rule. The fraction of annotation features flagged by this approach as suspicious is strongly enriched in errors and constitutes about 0.6% of the whole body of the similarity-transferred annotation in the PEDANT genome database. Positive rule mining does not identify two thirds of these errors. The approach based on exceptions from negative rules is much more specific than positive rule mining, but its coverage is significantly lower. CONCLUSION: Mining of both negative and positive association rules is a potent tool for finding significant trends in protein annotation and flagging doubtful features for further inspection.

UNLABELLED: Conserved domains represent essential building blocks of most known proteins. Owing to their role as modular components carrying out specific functions they form a network based both on functional relations and direct physical interactions. We have previously shown that domain interaction networks provide substantially novel information with respect to networks built on full-length protein chains. In this work we present a comprehensive web resource for exploring the Domain Interaction MAp (DIMA), interactively. The tool aims at integration of multiple data sources and prediction techniques, two of which have been implemented so far: domain phylogenetic profiling and experimentally demonstrated domain contacts from known three-dimensional structures. A powerful yet simple user interface enables the user to compute, visualize, navigate and download domain networks based on specific search criteria. AVAILABILITY: http://mips.gsf.de/genre/proj/dima

BACKGROUND: Comparison of large protein datasets has become a standard task in bioinformatics. Typically researchers wish to know whether one group of proteins is significantly enriched in certain annotation attributes or sequence properties compared to another group, and whether this enrichment is statistically significant. In order to conduct such comparisons it is often required to integrate molecular sequence data and experimental information from disparate incompatible sources. While many specialized programs exist for comparisons of this kind in individual problem domains, such as expression data analysis, no generic software solution capable of addressing a wide spectrum of routine tasks in comparative proteomics is currently available. RESULTS: PROMPT is a comprehensive bioinformatics software environment which enables the user to compare arbitrary protein sequence sets, revealing statistically significant differences in their annotation features. It allows automatic retrieval and integration of data from a multitude of molecular biological databases as well as from a custom XML format. Similarity-based mapping of sequence IDs makes it possible to link experimental information obtained from different sources despite discrepancies in gene identifiers and minor sequence variation. PROMPT provides a full set of statistical procedures to address the following four use cases: i) comparison of the frequencies of categorical annotations between two sets, ii) enrichment of nominal features in one set with respect to another one, iii) comparison of numeric distributions, and iv) correlation of numeric variables. Analysis results can be visualized in the form of plots and spreadsheets and exported in various formats, including Microsoft Excel. CONCLUSION: PROMPT is a versatile, platform-independent, easily expandable, stand-alone application designed to be a practical workhorse in analysing and mining protein sequences and associated annotation. The availability of the Java Application Programming Interface and scripting capabilities on one hand, and the intuitive Graphical User Interface with context-sensitive help system on the other, make it equally accessible to professional bioinformaticians and biologically-oriented users. PROMPT is freely available for academic users from http://webclu.bio.wzw.tum.de/prompt/.

Fold designability has been estimated by the number of families contained in that fold. Here, we show that among orthologous proteins, sequence divergence is higher for folds with greater numbers of families. Folds with greater numbers of families also tend to have families that appear more often in the proteome and greater promiscuity (the number of unique "partner" folds that the fold is found with within the same protein). We also find that many disease-related proteins have folds with relatively few families. In particular, a number of these proteins are associated with diseases occurring at high frequency. These results suggest that family counts reflect how certain structures are distributed in nature and is an important characteristic associated with many human diseases.

MOTIVATION: Millions of protein sequences currently being deposited to sequence databanks will never be annotated manually. Similarity-based annotation generated by automatic software pipelines unavoidably contains spurious assignments due to the imperfection of bioinformatics methods. Examples of such annotation errors include over- and underpredictions caused by the use of fixed recognition thresholds and incorrect annotations caused by transitivity based information transfer to unrelated proteins or transfer of errors already accumulated in databases. One of the most difficult and timely challenges in bioinformatics is the development of intelligent systems aimed at improving the quality of automatically generated annotation. A possible approach to this problem is to detect anomalies in annotation items based on association rule mining. RESULTS: We present the first large-scale analysis of association rules derived from two large protein annotation databases-Swiss-Prot and PEDANT-and reveal novel, previously unknown tendencies of rule strength distributions. Most of the rules are either very strong or very weak, with rules in the medium strength range being relatively infrequent. Based on dynamics of error correction in subsequent Swiss-Prot releases and on our own manual analysis we demonstrate that exceptions from strong rules are, indeed, significantly enriched in annotation errors and can be used to automatically flag them. We identify different strength dependencies of rules derived from different fields in Swiss-Prot. A compositional breakdown of association rules generated from PEDANT in terms of their constituent items indicates that most of the errors that can be corrected are related to gene functional roles. Swiss-Prot errors are usually caused by under-annotation owing to its conservative approach, whereas automatically generated PEDANT annotation suffers from over-annotation. AVAILABILITY: All data generated in this study are available for download and browsing at http://pedant.gsf.de/ARIA/index.htm.

Mark Borodovsky, Andrey Rzhetsky, Dmitrij Frishman; The Fifth Georgia Tech—Oak Ridge National Laboratory International Conference in Bioinformatics: in sil

Abstract Summary: The MIPS mammalian protein–protein interaction database (MPPI) is a new resource of high-quality experimental protein interaction data in mammals. The content is based on published experimental evidence that has been processed by human expert curators. We provide the full dataset for download and a flexible and powerful web interface for users with various requirements. Availability: The MPPI database is located at http://mips.gsf.de/proj/ppi/ Contact: [email protected]

BACKGROUND: Alternative splicing is a major mechanism of generating protein diversity in higher eukaryotes. Although at least half, and probably more, of mammalian genes are alternatively spliced, it was not clear, whether the frequency of alternative splicing is the same in different functional categories. The problem is obscured by uneven coverage of genes by ESTs and a large number of artifacts in the EST data. RESULTS: We have developed a method that generates possible mRNA isoforms for human genes contained in the EDAS database, taking into account the effects of nonsense-mediated decay and translation initiation rules, and a procedure for offsetting the effects of uneven EST coverage. Then we computed the number of mRNA isoforms for genes from different functional categories. Genes encoding ribosomal proteins and genes in the category "Small GTPase-mediated signal transduction" tend to have fewer isoforms than the average, whereas the genes in the category "DNA replication and chromosome cycle" have more isoforms than the average. Genes encoding proteins involved in protein-protein interactions tend to be alternatively spliced more often than genes encoding non-interacting proteins, although there is no significant difference in the number of isoforms of alternatively spliced genes. CONCLUSION: Filtering for functional isoforms satisfying biological constraints and accounting for uneven EST coverage allowed us to describe differences in alternative splicing of genes from different functional categories. The observations seem to be consistent with expectations based on current biological knowledge: less isoforms for ribosomal and signal transduction proteins, and more alternative splicing of interacting and cell cycle proteins.

BACKGROUND: Alternative splicing is an efficient mechanism for increasing the variety of functions fulfilled by proteins in a living cell. It has been previously demonstrated that alternatively spliced regions often comprise functionally important and conserved sequence motifs. The objective of this work was to test the hypothesis that alternative splicing is correlated with contact regions of protein-protein interactions. RESULTS: Protein sequence spans involved in contacts with an interaction partner were delineated from atomic structures of transient interaction complexes and juxtaposed with the location of alternatively spliced regions detected by comparative genome analysis and spliced alignment. The total of 42 alternatively spliced isoforms were identified in 21 amino acid chains involved in biomolecular interactions. Using this limited dataset and a variety of sophisticated counting procedures we were not able to establish a statistically significant correlation between the positions of protein interaction sites and alternatively spliced regions. CONCLUSIONS: This finding contradicts a naïve hypothesis that alternatively spliced regions would correlate with points of contact. One possible explanation for that could be that all alternative splicing events change the spatial structure of the interacting domain to a sufficient degree to preclude interaction. This is indirectly supported by the observed lack of difference in the behaviour of relatively short regions affected by alternative splicing and cases when large portions of proteins are removed. More structural data on complexes of interacting proteins, including structures of alternative isoforms, are needed to test this conjecture.

Abstract Summary: Phylogenetic Web Profiler (PWP) is a web-based service designed to perform phylogenetic profiling of proteins against genomes. The current version offers a selection of 63 completed genomes and available plasmids as annotated in the PEDANT genome database. Unlike currently available applications, this tool offers several choices of ortholog prediction parameters including E-value cutoff, percent length difference tolerance, and annotation similarity. Additional features include tight integration with the PEDANT database and tools to analyze properties of predicted proteins. PWP should prove very useful for the analysis of functional-linkage between proteins. Availability: The PWP server is available at http://pedant.gsf.de/pwp/. The PWP functional predictions are also offered as part of the PEDANT genome analysis server (http://pedant.gsf.de). Contact: [email protected] * To whom correspondence should be addressed.

Abstract Summary: SNAPper is a network service for predicting gene function based on the conservation of gene order. Availability: The SNAPper server is available at http://pedant.gsf.de/snapper. SNAPper-based functional predictions will soon be offered as part of the PEDANT genome analysis server http://pedant.gsf.de. Contact: [email protected] * To whom correspondence should be addressed.

MOTIVATION: Enormous demand for fast and accurate analysis of biological sequences is fuelled by the pace of genome analysis efforts. There is also an acute need in reliable up-to-date genomic databases integrating both functional and structural information. Here we describe the current status of the PEDANT software system for high-throughput analysis of large biological sequence sets and the genome analysis server associated with it. RESULTS: The principal features of PEDANT are: (i) completely automatic processing of data using a wide range of bioinformatics methods, (ii) manual refinement of annotation, (iii) automatic and manual assignment of gene products to a number of functional and structural categories, (iv) extensive hyperlinked protein reports, and (v) advanced DNA and protein viewers. The system is easily extensible and allows to include custom methods, databases, and categories with minimal or no programming effort. PEDANT is actively used as a collaborative environment to support several on-going genome sequencing projects. The main purpose of the PEDANT genome database is to quickly disseminate well-organized information on completely sequenced and unfinished genomes. It currently includes 80 genomic sequences and in many cases serves as the only source of exhaustive information on a given genome. The database also acts as a vehicle for a number of research projects in bioinformatics. Using SQL queries, it is possible to correlate a large variety of pre-computed properties of gene products encoded in complete genomes with each other and compare them with data sets of special scientific interest. In particular, the availability of structural predictions for over 300 000 genomic proteins makes PEDANT the most extensive structural genomics resource available on the web.

MOTIVATION: It is only a matter of time until a user will see not many but one integrated database of information for molecular biology. Is this true? Is it a good thing? Why will it happen? Where are we now? What developments are fostering and what developments are impeding progress towards this end? SUPPLEMENTARY INFORMATION: A list of WWW resources devoted to database issues in molecular biology is available at http://www.mips.biochem.mpg.de CONTACT: [email protected]