Attributes of living systems are constrained in evolution. An alternative to the analysis of conserved attributes ('characters') is analysis of functional interactions ('couplings') that cause conservation. In one application, evolutionary couplings in proteins in the form of amino acid pairwise covariation across a protein family (EVCouplings), can be used to computationally fold proteins, to predict oligomerization, functional sites and paths, and functionally distinct conformational states. In this talk I will present (i) the application of EVCouplings to drug target and drug response trans-membrane proteins, leading to the identification residues involved in interactions, ligand binding and disease. And (ii) the prediction of co-evolved residues across interacting protein and the 3D structure of protein complexes. We anticipate these types of global statistical algorithms will enable exciting new progress in a number of fields including protein design and human genetic variation, to name but two. A web service at www.EVfold.org provides a tool for the analysis of co-variation in proteins with respect to functional interactions and structural distance constraints

Despite great strides in de novo prediction of protein structure from amino acid sequence over the past decade, there seem to be rapidly diminishing returns in applying these methods to real problems. At the domain level, certainly, de novo prediction is of very limited use, particularly for globular proteins. It’s now rare to find interesting protein domains without any homologues of known 3-D structure, and even in cases where no templates can be found, template-free modelling results turn out to be little better than they were 10 years ago. More worryingly, there are no reliable ways of knowing if de novo methods have even been successful without actually solving the structure experimentally. In the last few years, developments in contact prediction have generated a lot of excitement, but the reality here is that for globular proteins, structural genomics projects have already targeted the largest domain families for which co-evolution methods work well enough, and in practice, it is now very hard to find suitable new targets for these approaches. Strangely enough, the ever present problems that have plagued 3-D modelling are still with us e.g. generating accurate structure-sequence alignments for very distant homologues or analogues, domain parsing and in particular the so-called modelling “endgame” i.e. protein model refinement. In my talk I will address some of these issues, drawing in particular on my experiences as a long-time CASP predictor and more recently, an assessor of model refinement in CASP10.

Immunoglobulins are key players of the immune response and their overall structure is reasonably well conserved. They are composed of two heavy and two light chains that contain four and two domains with a similar fold, respectively. Antibodies bind their cognate antigen using the tip of the first domains of each chain (VL and VH). From a structural point of view, the antigen binding site is formed by six loops, three from the light (L1, L2, L3) and three from the heavy chain (H1, H2, H3) named according to their order of appearance in the amino acid sequence. The structure of the framework, of the main chain of five of these loops and of part of the sixth can be predicted quite accurately. However a high quality of the details of the structure of antibodies is necessary to open the road to the possibility of understanding the molecular basis of their specificity, of redesigning them for therapeutical purposes and of unravelling the aetiology of important such as chronic lymphocytic leukaemia.

I will discuss the state of the art in this area.

Protein structure homology models are used routinely in a wide range of a applications when no direct experimental structures are available. Validation of models is essential to estimate the accuracy and reliability of computationally predicted protein structures. The quality of a model ultimately determines its usability for specific applications in life science research. The average performance of a prediction method can be assessed in retrospective experiments such as CASP and CAMEO by comparing the prediction to the experimental reference structure. In practical modeling applications, however, the reliability has to be predicted for each specific model using model quality estimation tools.

The Disease Ontology (DO) database (http://disease-ontology.org) represents a comprehensive knowledge base of >800 human diseases, including inherited, developmental and acquired. It semantically integrates disease and medical vocabularies through cross mapping and integration of diverse disease-specific terms and identifiers, including Medical Subject Headings (MeSH), International Classification of Diseases (ICD), and Online Mendelian Inheritance of Man (OMIM). The DO has been used to connect gene annotations with curated diseases and with disease phenotypes. The DO incorporates gene connections by mining Gene Wiki using the NCBO Annotator. The DO is utilized for disease annotation by major biomedical databases, as a standard representation of human disease in biomedical ontologies, and as an ontological cross mappings resource. The DO web browser has been designed for speed, efficiency and robustness through the use of a graph database. Full-text contextual searching functionality using Lucene allows the querying of name, synonym, definition, DOID and cross-references with complex Boolean search strings. We have been working to couple the DO with the Human Phenotype Ontology to understand the overlapping connections between diseases, disease phenotypes, and genetic variation. I will present paths that enable the exploration and exploitation of DO and its linked data structures.

Online Mendelian Inheritance in Man (OMIM®) is the premier resource for phenotypic annotation of the human genome.  While our primary focus is single gene Mendelian disorders, we also include diseases with complex etiology where there is significant contribution of a single gene to the phenotype (e.g., CFH and age-related macular degeneration).  Fundamental to building OMIM is defining a phenotype or disease.  There are many challenges to phenotyping:  cultural considerations, advances in diagnostic modalities and therapeutics, changes in diagnostic criteria, differences in medical care, medical subspecialty bias (the blind men and the elephant), ascertainment of the phenotypic features of a disorder at different stages of development, phenotypic variability, and genetic heterogeneity.  These last two items are at the heart of lumping and splitting.  How many different phenotypes can be caused by mutation in a gene? And how many genes can cause a phenotype?  It is important to understand and accommodate the evolution of knowledge in medicine and molecular biology when defining and naming a disorder.  With over 40 years of experience, OMIM takes the lead in considering these issues when naming disorders, assigning acronyms, and organizing and representing the complexity of phenotype-gene relationships.

Methodological advances have in recent years given us unprecedented information on the molecular details of living cells. However, it remains an unsolved challenge to systematically link molecular-level data to their phenotypic consequences at cellular level, e.g. diseases. One reason for this is that biology is facing the limitations of reductionism: most diseases cannot be attributed to a single gene – they can only be understood at the systems level. Networks have proven to be a very useful abstraction for bridging single-gene and systems-level analysis. Mapping all disease-related information about genes onto as comprehensive as possible a protein interaction network is thus an important step towards systems-level analysis of diseases. To this end, the STRING database (http://string-db.org) scores and integrates evidence from a diverse range of curated databases, raw data repositories, text-mining methods, and computational prediction methods to provide a comprehensive protein association network. We have recently developed three new web-based resources that use similar techniques to associate the proteins in the STRING network with their respective cellular compartments, tissues, and diseases to enable systems biology studies of diseases, taking into account both interactions and spatial localization of the proteins. These are available at http://compartments.jensenlab.org, http://tissues.jensenlab.org, and http://diseases.jensenlab.org.

Performing proper clinical diagnostics necessitates combining observation of specific gene sequence variants with a particular disease phenotype. The Leiden Open source Variation Database (LOVD) is a freely available Web-based software for the collection, display, and curation of DNA variants in locus-specific databases (LSDBs). The LOVD software was developed for the web-based collection and display of sequence variants and their functional consequences. The basic LOVD-view is a gene, listing all variants known (published and unpublished, submitted directly to the database) and links to other views and resources. To cope with all user demands LOVD has several access levels; website visitor, submitter, collaborator, curator and database manager. The data collected can be both public and non-public (i.e. requiring registration and login). LOVD facilitates the querying of non-public data, not returning the data but the number of records fulfilling the query. The LOVD software follows existing standards from organizations like HGVS, HVP and Gen2Phen regarding variant description, database content and data sharing. Lately many existing gene variant databases have changed to the LOVD-format, facilitating the use of sophisticated tools for data querying. The possibilities are essential to allow both import of exome/genome data into LOVD as well as their automated annotation.

The visualization of gene regulatory networks (GRNs) poses many challenges. These directed networks have three strong structural characteristics: out-degrees with a scale-free distribution, in-degrees bound by a low maximum and few and small cycles. Besides these global structural properties, subnetworks of a specific structure, called motifs, provide important knowledge about gene regulatory networks to domain experts. The talk provides an overview of visualization techniques that are currently in use for GRNs. We will look at visualization of gene expression from time series in the network context, and approaches to extract and visualize differentially expressed subnetworks (hotspots) from GRNs. Such a hotspot-based visualization approach enables interactive exploration of regulatory interactions in differentially expressed gene sets, and it allows a researcher to explore gene expression in direct relation to the affected cellular gene network. We will also look at novel depictions of GRNs that do not rely on node-link diagrams, which often result in complex hairballs. The recently proposed Compressed Adjacency Matrix provides an alternative depiction, which is space efficient and promotes the visual detection of patterns such as regulatory motifs. 

Changes in the cell microenvironment are mediated by signals from the cell surface to the nucleus. Such signals are transmitted by sequential protein interactions, represented as an interaction network. These networks are decomposed into signaling pathways, that are a consensus expert description of the function of a set of interactions. Interactions in these curated pathways are described in published articles. However, identifying signaling pathways from experimental data is difficult, due to the large number of published studies that need to be surveyed, in order to find support for interactions detected in the data. For that reason, reliable pathway reconstructions from literature evidence are indispensable.  We propose a novel approach to characterize, compare and build signaling pathways by combining the interactome with literature evidence. Our approach is based on Latent Semantic Indexing (LSI), a method combining text mining and dimensionality reduction to identify key concepts in published abstracts. We used LSI to identify and characterize pathways by measuring textual similarities between articles used as evidences for individual protein-protein interactions. We will present our framework as well as the techniques and visualization used to evaluate the performances of our method. Our approach is a useful tool for experimentalists, since it adds value to experimental results by providing a knowledge-based context.

Improvements in biological data acquisition and genomes sequencing now allow to reconstruct entire metabolic networks of many living organisms. Size and complexity of these data prohibit manual drawing and thereby urge the need of dedicated visualization techniques. In this talk, we will present an overview of metabolism representation at different scales from a single pathway to an entire network. The automatic drawing of metabolic pathways has been widely studied and provided efficient solutions. On the other hand, when interested in the automatic generation of entire metabolic networks representations, respecting the network topology, its decomposition into pathways, and biological drawing conventions raises many graph drawing and theoretical issues.

The signal transduction networks regulate cell proliferation, growth, adaptation and death; and integrate information from the environment with cellular status. However, the combinatorial complexity makes these networks extremely difficult to visualise or model accurately. Furthermore, the granularity difference between empirical observations and the specific states used in many theoretical models makes the relationship between them ambiguous, decreasing model quality and reliability. To address this, we have developed the reaction-contingency (rxncon) framework to describe cellular signal transduction networks at the same granularity as empirical data. This minimises the combinatorial complexity and support automatic visualisation and mathematical modelling. The key feature is strict separation of elemental reactions from contingencies, which define contextual constrains on these reactions. The rxncon software tool support the framework by automating both the visualisation and model creation from the network definition. This system integrates the three levels of network analysis; definition, visualisation and mathematical modelling; an important step towards a common framework to bridge different standards as well as experimental and theoretical systems biology efforts.

Genetic interaction screens have been applied to great success in S. cerevisiae to study gene function and the genetic architecture of the cell. However, most previous studies have been performed under optimal growth conditions while many functional interactions are known to be condition specific. In this study we have performed a large scale genetic interaction analysis in the presence of five different stress conditions (e.g. osmotic, oxidative, cell-wall altering) resulting in approximately 257,000 differential measurements. We find an extensive number of conditional genetic interactions that recapitulate known stress specific functional associations. Novel findings include a role for Bud14, the histone methylase COMPASS and some membrane trafficking complexes in modulating the CWI pathway. Finally, the conditional genetic interactions were compared with conditional changes in phosphorylation and gene expression status to dissect the functional importance of the different response mechanisms of the cell (i.e. early vs. late responses).

Post-translational modifications (PTMs) regulate protein activity, stability and interaction profiles and are critical for cellular functioning. Further regulation is gained through PTM interplay whereby modifications modulate the occurrence of other PTMs or act in combination. Integration of global acetylation, ubiquitination and tyrosine or serine/threonine phosphorylation datasets with protein interaction data identified hundreds of protein complexes that selectively accumulate each PTM, indicating coordinated targeting of specific molecular functions. A second layer of PTM coordination exists in these complexes, mediated by PTM integration (PTMi) spots. PTMi spots represent very dense modification patterns in disordered protein regions and showed an equally high mutation rate as functional protein domains in cancer, inferring equivocal importance for cellular functioning. Systematic PTMi spot identification highlighted more than 300 candidate proteins for combinatorial PTM regulation. This study reveals two global PTM coordination mechanisms and emphasizes dataset integration as requisite in proteomic PTM studies to better predict modification impact on cellular signaling.

Protein interaction prediction methods provide orthogonal information to the one obtained from high-throughput experiments and help to contextualize functional data. As such computational and experimentally derived interactions are a common component of genome analysis pipelines. Still, much remains to be done in terms of improving the quality and specificity of predicted interactions.

Recent publications in the area of protein folding have revitalized the interest in the use of co-evolution based methods for the prediction of protein interactions. In this presentation, I will first review the evolution, possibilities and limitations of the methods based on co-evolution for the prediction of protein interactions. In the second part of the talk, I will present the application of a new co-evolution based approach to the prediction of specie specific protein interaction networks in a large set of bacterial species.

References
de Juan D, Pazos F, Valencia A. (2013) Emerging methods in protein co-evolution. Nat Rev Genet. 14:249-261.

Key disease proteins have hundreds of known interactions. For example, EGFR has over 400. However, at any one time, a single protein is likely bound to only a few of its possible partners; its many known interactions likely correspond to numerous distinct interaction contexts, such as different cell types. Identifying these contexts is important for understanding drug mechanism of action; for a therapeutic effect, a drug may need to regulate specific interactions in a specific context. First, we identify the context of interactions by integrating PPI networks with gene expression data from multiple tissues and diseases. Second, we overlay these context-annotated networks with drug-regulated gene expression data; thus, linking drug mechanism of action to specific interaction contexts.

High-throughput interaction discovery initiatives are providing thousands of novel protein interactions which are unveiling many unexpected links between apparently unrelated biological processes. In particular, analyses of the first draft human interactomes highlight a strong association between protein network connectivity and disease. Indeed, recent exciting studies have exploited the information contained within protein networks to disclose some of the molecular mechanisms underlying complex pathological processes. These findings suggest that both protein-protein interactions and the networks themselves could emerge as a new class of targetable entities, boosting the quest for novel therapeutic strategies. In this talk, I will summarize our work towards the characterization and modelling of the protein-interaction network underlying Alzheimer´s disease, together with our most recent attempts to decipher complex cell networks to the point of being able to predict how the perturbation of a node might affect the system as a whole.

The speakers of the session and the organisers invite to a discussion on current trends.