Short Abstract: We propose an exhaustive computational protein-protein interaction screening system “MEGADOCK” that conducts docking and post-docking analysis on the protein tertiary structure data. MEGADOCK consists of two parts —a docking calculation part and interaction prediction part. The docking calculation part performs all-to-all docking and generates high-scoring decoys for all possible combinations of the given protein structures. Subsequently, the post-docking calculations are conducted to analyze the structural distributions of high-scoring decoys for each pair of proteins and decides if the given two proteins have potential to interact. Finally, we obtain a possible protein interaction network by connecting the positively predicted interactions.
MEGADOCK is designed to be fitted to large-scale parallelized computing environments. Performance of our prediction method was measured on a general docking benchmark dataset and a small bacterial signal transduction pathway data. In addition, some preliminary results of the application of our methods to a larger scale data set (such as 2,000 x 2,000 = 4,000,000 docking analysis) will be shown in the poster. It includes analysis of potential interactions among proteins that consist human epidermal growth factor signaling pathway and gene products of which expression changed upon drug induction.
The proposed system can be one of the promising mothodology that can go between structural study and system level analyses by utilizing cumulative protein structure data.

Short Abstract: Identifying similarities and differences in expression patterns across multiple time series can provide a better understanding of the relationships among various chemical treatments or the effects induced by a gene knockout/suppression. We consider the task of identifying sets of genes that have a high degree of similarity both in their (i) expression profiles within each treatment, and (ii) changes in expression responses across treatments. Previously, we developed an approach for aligning time series that computes clustered alignments. In this approach, an alignment represents the correspondences between two gene expression time series. Portions of one of the time series may be compressed or stretched to maximize the similarities between the two series. A clustered alignment groups genes such that the genes within a cluster share a common alignment, but each cluster is aligned independently of the others. Unlike standard gene-expression clustering, which groups genes according to the similarity of their expression profiles, the clustered-alignment approach clusters together genes that have similar changes in expression responses across treatments. We have now extended the clustered alignment approach to produce multi-level clusterings that identify subsets of genes that have a high degree of similarity both in their (i) expression profiles within each treatment, and (ii) changes in expression responses across treatments. We examine the validity of this multi-level clustering method by performing a GO-term enrichment analysis of the clusters. Additionally, we use permutation testing to determine if our clusters have alignment scores that are unlikely to occur by chance.

Short Abstract: This poster is based on Proceedings Submission 142.

Motivation: Biological networks are an attractive means for studying evolution. Evolutionary mechanics can be inferred from the topology of networks of extant organisms. One method for inferring evolutionary mechanics is to construct models which generate networks sharing topological characteristics with their empirical counterparts. However, selecting the most appropriate measures upon which to base comparisons remains ambiguous given the innumerable possibilities of topological properties. Furthermore, when a model needs to be improved, it is unclear which topological property should be addressed by new model mechanics.

Results: We introduce an optimization process which applies selective pressure to individual topological properties concurrently with network generation. Applying our method to protein interaction networks, we are better able to identify the relevance, volatility, and correlation among topological properties in a model of network evolution. We find that selective pressure on individual topological characteristics may drive additional properties towards empirical values. Other topological characteristics are found to be less influential and vacillate under different selection regimens. These findings suggest that an evolutionary process that directly impacts one of these influential topological characteristics may be sufficient to explain multiple observed characteristics. Our framework also provides the capability to discern the interrelationships among topological properties in network evolution and can help to focus the search for biologically plausible and relevant processes important to network evolution.

Short Abstract: Comprehensive protein-protein interaction (PPI) maps are a powerful resource for uncovering the molecular basis of genetic interactions and providing mechanistic insights. Coaffinity purification combined with tandem mass spectrometry (coAP-MS) has been used to generate PPI maps at proteome scale, but results are confounded by both high false positive and false negative rates.

To address these issues, several methods have been developed to post-process coAP-MS datasets. These generally fall in two classes: spoke and matrix models. Spoke models produce confidence scores on directly observed interaction data, whereas matrix models additionally infer interactions that are not directly observed and hence have broader coverage at the expense of increased false positives.

Recent literature has shown promising results from matrix model methods. However, with few exceptions, these methods only consider binary experimental data (where each possible interaction is deemed either observed or unobserved), throwing away any quantitative information from the experiment such as spectral counts.

We propose a novel approach to incorporating quantitative interaction information into coAP-MS PPI prediction. Our methodology introduces a probabilistic framework that addresses the uncertainty of observed interactions, for example interactions with low spectral counts. Using a sampling-based approach, we model this uncertainty with an ensemble of possible alternative experimental outcomes. Importantly, this procedure allows us to directly harness previous methods without modification, thus extending previous methods to use quantitative information. We demonstrate that our approach improves interaction prediction performance on the recently published Drosophila Protein interaction Map (DPiM), the largest Drosophila coAP-MS dataset to date, which includes nearly 5000 proteins.

Short Abstract: Autism is a neurodevelopmental disorder involving a large number of functionally diverse genes. Currently, it is not completely understood how these genes interact with each other and with other human genes. Even less is known about the influence of alternative splicing (AS) on protein-protein interactions. Here, we performed high-throughput splice isoform discovery for 191 autism candidate genes from fetal and adult human brain RNA, and then screened them for interactions with 15,000 human ORFs. We have identified 373 brain-expressed AS isoforms, 226 of which were novel. This increased the isoform space of autism candidate genes by 29%. We then build autism-centered “splice-actome” consisting of 630 isoform-level interactions. By incorporating isoform interactions into the network, we were able to expand it by 25%. The comparison between fetal and adult isoform networks demonstrated 59% overlap, emphasizing network-level similarities and differences between these two brain tissues. In order to evaluate the probability of isoforms to share partners, we have implemented Fraction of Shared Interactions (FSI) score. This score allowed identification of contrasting isoforms groups: those that share the majority of their interaction partners, and those that have unique partners. Furthermore, isoform interactions also influenced topological properties of the network such as its connectivity and modularity. This work clearly demonstrates that “splice-actome” adds another layer of complexity to autism network, and may be necessary step towards better understanding of other disease networks.

Short Abstract: Protein–Protein Interaction (PPI) networks are believed to be composed of modules that perform specific, crucial functions in biological systems. As such networks are important sources of information related to protein complexes and functional modules, a variety of community detection methods have been developed to exploit them. However, due to the nature of various regulating modes exists in biological networks, a single method may over-fit to a single graph property and is blind to communities highlighted by other network properties.
We propose an integration method to capture protein complex generated by different methods. The proposed integrated method is implemented into a web-based platform with a powerful interactive network analyzer called Spotlight. Several popular methods with different viewpoints for network community detection, including CPM, FastGreedy, HUNTER, MCL, SpinGlass, and WalkTrap, are also executable simultaneously. The integrated method in Spotlight successfully harvests biological machineries from the yeast interactome, evaluated by the enrichment of GO terms and recovery of MIPS complexes. This integrating graph clustering method/ platform will help discovering network community structures as well as novel complex components/ regulators, inferred by the close agglomeration to known biological complexes.
Availability: Spotlight is freely accessible at http://hub.iis.sinica.edu.tw/spotlight. Video clips for a quick view of usage are available in the website online help page.

Short Abstract: Mapping expression quantitative trait loci (eQTL) is the process of linking variation in gene expression levels to genomic variation. Typically, genome wide gene expression, measured using micro array technology, is combined with SNP data to identify cis or trans loci, providing insight in the mechanisms of gene regulation. In the last decade, a large number of eQTL mapping methods have been demonstrated. As it turns out, the results of different eQTL mapping methods vary considerably in the number and the location of loci found. Our basic assumption is that different methods yielding different eQTL are complementary. If this assumption holds, this may eventually lead to ensemble strategies for eQTL mapping. We have compared the results of three mapping methods: non-parametric regression, mixed models and elastic net regression. All three methods are evaluated on the well known yeast dataset of Brem and Kruglyak. The data includes genome wide gene expression and 2957 SNPs for 112 segregants of Saccharomyces cerevisiae. We analyzed the results of eQTL mapping quantitativly and qualitatively, identifying regions of consensus and disagreement, and variations in detection strength between the methods. We payed special attention to the way the different mapping methods handle the presence of large haplotype blocks in the SNP data. As a result, we found that the three methods indeed yield very different results. We finally attempt to combine the different results using a simple chromosome position based rule set.

Short Abstract: Much of the information about gene relationships is hidden in the text of the many papers in the biology research corpus. A variety of automated and manual techniques have been used to extract this information. We have developed GeneLabel, a crowdsourcing application designed to get people to help extract the information in these sentences. Using these manual labelings, we refine a statistical model to improve sentence and word selection for further manual curation.

We begin by extracting sentences from the body of research articles in open access journals, and identifying sentences that contain two or more genes of interest. Using part-of-speech tagging and several simple heuristics, we score every word in each sentence based on how important we think it is for describing gene relationships.

We use this information to build the GeneLabel web interface. GeneLabel is a webapp which presents users with a set of sentences. For each sentence, the user is asked to select the relationship demonstrated between two chosen genes. We begin by showing only a few words from each sentence. Every 5 seconds, we add additional words. Users are given a score based on how quickly and accurately they label the set of sentences.

We record which words were displayed when a user labels a sentence, and use this information to refine our model of word importance. Ongoing work involves the selection of sentences most likely to contain interesting gene relationships and the analysis of preliminary usage data.

The Genelabel prototype can be accessed at http://genelabel.appspot.com.

Short Abstract: Several factors regulate the interaction between proteins in an interactome. We are interested in factors that enable a protein to can carry out many interactions. To address part of the problem, we are trying to understand how the intrinsically disordered regions (IDRs) in the proteins participate in the protein-protein interaction. IDRs can be recognized in experimental structural data but have also been predicted from primary amino acid sequences. We have explored the correlation of bioinformatic predictions of IDRs from sequence data and experimental identification from structural data, namely B-factors, measurementes available for proteins structures derived from X-ray crystallography studies that indicate protein mobility at the individual atom resolution. We have written programs that correlate B-factors and IDRs predictions using the crystallography structures determined at 1.5-2.0 Å available in the PDB. The correlation between B-factors and IDRs was also studied using the annotation of Gene Ontology by cellular component for each protein in our database. We are also studying the correlation between predicted IDRs and the interaction sites known from experimental data to understand if there is a close relationship between the number of IDRs, the lenght of IDRs and the number of interactions that a protein can carry out.