Academia.eduAcademia.edu

Figure 34 – uploaded by Björn Olsson

See full PDFdownloadDownload figure
Fig. 1. The EnsEMBL and TIGR-XML entities model protein-coding genes on the genomic DNA as fixed hierarchical tree structures The EnsEMBL and TIGR-XML Annotation. EnsEMBL [2] and TIGR-XML [3] model protein-coding genes on the genomic DNA as fixed hierarchical tree structures as shown in Fig. 1. A gene locus may have one or more splice isoforms (mRNAs). Each mRNA splits into the protein coding CDS region and the two untranslated regions 5’UTR (upstream) and 3’UTR (downstream). This topology cannot deal with alternative start codons for the same mRNA or other alternative mRNAs with the same splicing pattern, such as mRNAs with alternative transcription start sites or alternative polyadenylation sites. Moreover, there are no crosslinks given between the different CDSs of a gene, although these would be instructive, since often alternatively spliced mRNAs differ only in the UTR regions but lead to the same CDS and therefore code for the same protein.

Figure 1 The EnsEMBL and TIGR-XML entities model protein-coding genes on the genomic DNA as fixed hierarchical tree structures The EnsEMBL and TIGR-XML Annotation. EnsEMBL [2] and TIGR-XML [3] model protein-coding genes on the genomic DNA as fixed hierarchical tree structures as shown in Fig. 1. A gene locus may have one or more splice isoforms (mRNAs). Each mRNA splits into the protein coding CDS region and the two untranslated regions 5’UTR (upstream) and 3’UTR (downstream). This topology cannot deal with alternative start codons for the same mRNA or other alternative mRNAs with the same splicing pattern, such as mRNAs with alternative transcription start sites or alternative polyadenylation sites. Moreover, there are no crosslinks given between the different CDSs of a gene, although these would be instructive, since often alternatively spliced mRNAs differ only in the UTR regions but lead to the same CDS and therefore code for the same protein.

Related Figures (102)

Fig. 1. Data integration pipeline
Fig. 1. Data integration pipeline
Table 4. Mapping of OBO and OWL terms. NDY stands for not defined yet.
Table 4. Mapping of OBO and OWL terms. NDY stands for not defined yet.
Table 6. CCO ID and GO ID for the sample shown in Figure [4 Conclusions and Future Work
Table 6. CCO ID and GO ID for the sample shown in Figure [4 Conclusions and Future Work
Fig. 4. Comparison of two sample class sub-hierarchies. Let O be an ontology struc- ture, where {s1,82,83,84} C S and {pi,p2} C P and O’, where {s2,s3} C S’ and {p1, p2,p3, pa} C P’. The ontology O, on the left, has an inconsistent relation s4. On the right, the ontology O’ is shown with a different, supposedly correct semantics.
Fig. 4. Comparison of two sample class sub-hierarchies. Let O be an ontology struc- ture, where {s1,82,83,84} C S and {pi,p2} C P and O’, where {s2,s3} C S’ and {p1, p2,p3, pa} C P’. The ontology O, on the left, has an inconsistent relation s4. On the right, the ontology O’ is shown with a different, supposedly correct semantics.
Source database Number of vertices Mean degree *OMIM entries correspond to phenotype descriptions and gene loci.
Source database Number of vertices Mean degree *OMIM entries correspond to phenotype descriptions and gene loci.
Fig.1. The graph induced by 20 best paths from gene AD6 to Alzheimer disease. The terminal vertices are rectangular. The edges are labelled with their probabilities (a = 0.25,rq = 0.8). Gene AD6 (Entrez Gene entry 64851) is linked to the locus description (OMIM entry 605526) by a direct edge and via three articles. The locus description is in turn linked to another locus description, insulin-degrading enzyme (OMIM entry 146680), via two articles, and, finally, to Alzheimer disease (OMIM entry 104300) via two proteins and two UniGene clusters. to eliminate the trivial links, we removed all paths shorter than three edges from this set. Figure [shows an induced graph for AD6 (but for clarity only 20 best paths). We used the goodness value of the best of the remaining paths, and the two-terminal network reliability of the graph induced by the remaining paths as test statistics. Two-terminal network reliability was estimated using Monte Carlo algorithm with 100,000 iterations; standard deviation of the estimate is less than 0.0064. Based on these two statistics, we then estimated two p values—one for the best path and another for the connection subgraph—for each candidate gene.
Fig.1. The graph induced by 20 best paths from gene AD6 to Alzheimer disease. The terminal vertices are rectangular. The edges are labelled with their probabilities (a = 0.25,rq = 0.8). Gene AD6 (Entrez Gene entry 64851) is linked to the locus description (OMIM entry 605526) by a direct edge and via three articles. The locus description is in turn linked to another locus description, insulin-degrading enzyme (OMIM entry 146680), via two articles, and, finally, to Alzheimer disease (OMIM entry 104300) via two proteins and two UniGene clusters. to eliminate the trivial links, we removed all paths shorter than three edges from this set. Figure [shows an induced graph for AD6 (but for clarity only 20 best paths). We used the goodness value of the best of the remaining paths, and the two-terminal network reliability of the graph induced by the remaining paths as test statistics. Two-terminal network reliability was estimated using Monte Carlo algorithm with 100,000 iterations; standard deviation of the estimate is less than 0.0064. Based on these two statistics, we then estimated two p values—one for the best path and another for the connection subgraph—for each candidate gene.
Table 2. Results: Alzheimer disease (a = 0.25, rg = 0.8)
Table 2. Results: Alzheimer disease (a = 0.25, rg = 0.8)
In a second, more challenging experiment, we evaluated the strength of link between APP and five genes whose protein products interact with the APP protein: HADH2, APBA1, CHRNA7, APOA1, and SHC1. The interactions were obtained from the Int Act-databasd4, The experiments were carried out the same way as with Alzheimer disease, except that we used the first, symmetric null hypothesis (i.e., vertices at both ends were randomized). In the results, two genes show significant linkage to APP (Table[/). The other three genes get non-significant p values despite relatively high values of the test statistics (com- pared to the Alzheimer experiment), suggesting that pairs of genes are generally
In a second, more challenging experiment, we evaluated the strength of link between APP and five genes whose protein products interact with the APP protein: HADH2, APBA1, CHRNA7, APOA1, and SHC1. The interactions were obtained from the Int Act-databasd4, The experiments were carried out the same way as with Alzheimer disease, except that we used the first, symmetric null hypothesis (i.e., vertices at both ends were randomized). In the results, two genes show significant linkage to APP (Table[/). The other three genes get non-significant p values despite relatively high values of the test statistics (com- pared to the Alzheimer experiment), suggesting that pairs of genes are generally
In our research, we want to investigate to which extent a system relying mostly o1 seneral linguistic knowledge, and to a minimal degree on specialized ontologies cat ye successful in extracting relations from biological texts. We recently proposed « srammar-based method for extraction of biological relations from scientific texts [23. 26]. The method uses an algorithm that searches through the syntactic trees producec y a linguistic parser, identifies relations mentioned in the sentence, and classifie: hem with respect to their semantic class and epistemic status (facts, counterfactuals xr hypotheses). The semantic categories used in the classification are based on the elation set used in KEGG, so that pathway maps following the same notational con. vention as KEGG can be automatically generated, and even other relations involvins ological objects (coocurrence, part-whole relations) may be extracted. Subse juently, we added several extensions and improvements of the method, such as at mproved named entity recognition component and the addition of a method for dis. inguishing between text describing previous and current work, thereby making i sossible to avoid extracting relations from text sections which merely report finding: rom previous work or common knowledge, rather than new findings [27]. In our work, we consider text analysis as a method to extract candidate pathways, which must be further evaluated for biological plausibility. As shown in Figure 1, a set of candidate pathways is extracted from a corpus consisting of a set of selected biomedical articles from PubMed, which allows the user to find all articles containing particular words or combinations of words. The text analysis process uses lexical databases and a grammar to achieve sufficient comprehension of the text for high- precision extraction of pathways corresponding to the textual description. However, even if the candidate pathways extracted by our approach are generally correct with respect to the text (as demonstrated in [24] and [27]), not all of them are necessarily biologically plausible. Therefore, each candidate is evaluated by alignment to and comparison with currently known pathways collected from KEGG and other model databases. As described in detail in [28] the plausibility score is derived by calculat- ing the semantic similarity between the set of gene products that have been paired in
In our research, we want to investigate to which extent a system relying mostly o1 seneral linguistic knowledge, and to a minimal degree on specialized ontologies cat ye successful in extracting relations from biological texts. We recently proposed « srammar-based method for extraction of biological relations from scientific texts [23. 26]. The method uses an algorithm that searches through the syntactic trees producec y a linguistic parser, identifies relations mentioned in the sentence, and classifie: hem with respect to their semantic class and epistemic status (facts, counterfactuals xr hypotheses). The semantic categories used in the classification are based on the elation set used in KEGG, so that pathway maps following the same notational con. vention as KEGG can be automatically generated, and even other relations involvins ological objects (coocurrence, part-whole relations) may be extracted. Subse juently, we added several extensions and improvements of the method, such as at mproved named entity recognition component and the addition of a method for dis. inguishing between text describing previous and current work, thereby making i sossible to avoid extracting relations from text sections which merely report finding: rom previous work or common knowledge, rather than new findings [27]. In our work, we consider text analysis as a method to extract candidate pathways, which must be further evaluated for biological plausibility. As shown in Figure 1, a set of candidate pathways is extracted from a corpus consisting of a set of selected biomedical articles from PubMed, which allows the user to find all articles containing particular words or combinations of words. The text analysis process uses lexical databases and a grammar to achieve sufficient comprehension of the text for high- precision extraction of pathways corresponding to the textual description. However, even if the candidate pathways extracted by our approach are generally correct with respect to the text (as demonstrated in [24] and [27]), not all of them are necessarily biologically plausible. Therefore, each candidate is evaluated by alignment to and comparison with currently known pathways collected from KEGG and other model databases. As described in detail in [28] the plausibility score is derived by calculat- ing the semantic similarity between the set of gene products that have been paired in
2 Named Entity Recognition and Tagging — Domain Adaptation and Testing Since the precondition for in-depth syntactic and semantic text analysis is appropriate tagging, we started by domain-oriented training and development of the Named Entity Identification (NER) procedure and the semantico-syntactic tagger. The original NER-algorithm and the tagger had been developed for the purpose of processing news reports [29, 30]. The tagger utilized parts of WordNet (version 1.6 [31]), a list of frequent closed-class words and an internally developed lexicon of verbs denoting the alignment, using the Gene Ontology structured vocabulary of annotation term By this technique it is possible to identify similarities between gene products thi have annotations that seem to differ at a surface-level comparison, but which actuall perform highly similar molecular functions or participate in similar biological proc esses. Using the derived plausibility scores, the candidate pathways are ranked so th: the biologist using the system can select the most biologically relevant pathways thi have been derived by the text analysis module. For details on the evaluation of cand date pathways, the reader is referred to [28]. Figure 2 shows the detailed architectur of the text processing system aimed at path extraction. The syntactic analysis and the algorithm deriving KEGG-like pathways from syr
2 Named Entity Recognition and Tagging — Domain Adaptation and Testing Since the precondition for in-depth syntactic and semantic text analysis is appropriate tagging, we started by domain-oriented training and development of the Named Entity Identification (NER) procedure and the semantico-syntactic tagger. The original NER-algorithm and the tagger had been developed for the purpose of processing news reports [29, 30]. The tagger utilized parts of WordNet (version 1.6 [31]), a list of frequent closed-class words and an internally developed lexicon of verbs denoting the alignment, using the Gene Ontology structured vocabulary of annotation term By this technique it is possible to identify similarities between gene products thi have annotations that seem to differ at a surface-level comparison, but which actuall perform highly similar molecular functions or participate in similar biological proc esses. Using the derived plausibility scores, the candidate pathways are ranked so th: the biologist using the system can select the most biologically relevant pathways thi have been derived by the text analysis module. For details on the evaluation of cand date pathways, the reader is referred to [28]. Figure 2 shows the detailed architectur of the text processing system aimed at path extraction. The syntactic analysis and the algorithm deriving KEGG-like pathways from syr
Tagging of the training corpus resulted in 2500 unique tagged strings (after exclusion of tags identified as numbers). The distribution of morphosyntactic categories among those unique entries is shown in figure 4. It can easily be seen that proper nouns (i.e. names of genes, proteins etc.) and common nouns are the dominating categories. As much as 66% of the unique words (types) found in our corpus were nouns or proper nouns. This fits the observation, made in [23], that biomedical texts are extremely “noun-heavy” (74% of the word occurrences in a 15 000 corpus of biological texts investigated by the authors of [23] belonged to noun phrases). WordNet was the primary knowledge source responsible for noun identification (95% of the tagged common nouns were found in WordNet). All proper names found in the corpus were identified by the NER-procedure. Verbs, however, were to a con- siderable degree (38% of all content verb occurrences) identified on the basis of the small internal verb lexicons: the one of cognition and communication verbs (79 lex- emes) and the yet smaller lexicon of frequent “bioverbs” (55 lexemes). This shows that the repertoire of verbs in this domain is highly repetitive. The results obtained from the test set confirm this observation. MTF oa Pes cep Se: eee 8 cee A 2 Pe, Bec ee ec ees eee: cee SS “Besse else ee oe See DS eee athe ieee
Tagging of the training corpus resulted in 2500 unique tagged strings (after exclusion of tags identified as numbers). The distribution of morphosyntactic categories among those unique entries is shown in figure 4. It can easily be seen that proper nouns (i.e. names of genes, proteins etc.) and common nouns are the dominating categories. As much as 66% of the unique words (types) found in our corpus were nouns or proper nouns. This fits the observation, made in [23], that biomedical texts are extremely “noun-heavy” (74% of the word occurrences in a 15 000 corpus of biological texts investigated by the authors of [23] belonged to noun phrases). WordNet was the primary knowledge source responsible for noun identification (95% of the tagged common nouns were found in WordNet). All proper names found in the corpus were identified by the NER-procedure. Verbs, however, were to a con- siderable degree (38% of all content verb occurrences) identified on the basis of the small internal verb lexicons: the one of cognition and communication verbs (79 lex- emes) and the yet smaller lexicon of frequent “bioverbs” (55 lexemes). This shows that the repertoire of verbs in this domain is highly repetitive. The results obtained from the test set confirm this observation. MTF oa Pes cep Se: eee 8 cee A 2 Pe, Bec ee ec ees eee: cee SS “Besse else ee oe See DS eee athe ieee
Fig. 5. Tags (occurrences) in the test set in relation to knowledge sources 95.2%). Figure 5 shows the impact of information learned from the training corpus ot different part-of-speech classes. As could be expected, almost all occurrences o closed class words obtained their POS tags from the tag memory database. The use o WordNet was considerably reduced. Only 18.6% of the common nouns in the tes material required search in WordNet, compared to 95% of the common nouns in the training set. 87% of all “bioverb” occurrences, 46% of the cognition and communica tion verbs, and 67% of other content verb occurrences were covered by the Tas Memory Database (TMD). This means that searching WordNet’s verb part wa: needed only for 19% of all verb occurrences in the test corpus, compared to 62% it the training corpus. As expected, the category that was least covered by the TMD wa: the group of proper nouns. The time needed for preprocessing and tagging decreasec by almost 40%.
Fig. 5. Tags (occurrences) in the test set in relation to knowledge sources 95.2%). Figure 5 shows the impact of information learned from the training corpus ot different part-of-speech classes. As could be expected, almost all occurrences o closed class words obtained their POS tags from the tag memory database. The use o WordNet was considerably reduced. Only 18.6% of the common nouns in the tes material required search in WordNet, compared to 95% of the common nouns in the training set. 87% of all “bioverb” occurrences, 46% of the cognition and communica tion verbs, and 67% of other content verb occurrences were covered by the Tas Memory Database (TMD). This means that searching WordNet’s verb part wa: needed only for 19% of all verb occurrences in the test corpus, compared to 62% it the training corpus. As expected, the category that was least covered by the TMD wa: the group of proper nouns. The time needed for preprocessing and tagging decreasec by almost 40%.
Fig. 7. The current format of the output from the information extraction procedure thus implemented 1n an additional piece of code (in C7). In Figure 7, we show the present format of the output from the information extrac- tion procedure. The relation/processes identified in a sentence are presented as a rela- tively simple graph, where the node “subj’(ect) refers to the biological entity actively involved in the process, nodes “obj’”(ect) or” prepobj(ect)” refer to other entities par- ticipating in the relations, and “circumstances” contain the information about place, conditions, etc., that may be of relevance. The main relation, including the circum- stances, is presented to the user, and other circumstance nodes may be expanded by the Weer
Fig. 7. The current format of the output from the information extraction procedure thus implemented 1n an additional piece of code (in C7). In Figure 7, we show the present format of the output from the information extrac- tion procedure. The relation/processes identified in a sentence are presented as a rela- tively simple graph, where the node “subj’(ect) refers to the biological entity actively involved in the process, nodes “obj’”(ect) or” prepobj(ect)” refer to other entities par- ticipating in the relations, and “circumstances” contain the information about place, conditions, etc., that may be of relevance. The main relation, including the circum- stances, is presented to the user, and other circumstance nodes may be expanded by the Weer
The quotes indicate the beginning and the end of the sentence and the box in the middle is used to create the content. If the user clicks on it, then a menu is displayed that shows the different options for beginning a sentence. Since we want to talk about the protein Bub1 we first insert ‘Bubl’ as a proper name. This looks as follows.
The quotes indicate the beginning and the end of the sentence and the box in the middle is used to create the content. If the user clicks on it, then a menu is displayed that shows the different options for beginning a sentence. Since we want to talk about the protein Bub1 we first insert ‘Bubl’ as a proper name. This looks as follows.
where the proper name ‘Bub1’ is now fixed as the beginning of the sentence, and we have a new menu with different entries. We want to express the interaction with another protein, and thus we choose the verb ‘interacts-with’. Like proper names, verbs are hierarchically structured and we can navigate through this hierarchy. In the next step we get
where the proper name ‘Bub1’ is now fixed as the beginning of the sentence, and we have a new menu with different entries. We want to express the interaction with another protein, and thus we choose the verb ‘interacts-with’. Like proper names, verbs are hierarchically structured and we can navigate through this hierarchy. In the next step we get
(HGVS). According to this proposed standard, our variant should be described by the following expression: NC_000019.8:g.11087877G>T, where NC_000019.8 is the unique accession number (in the NCBI RefSeq database) of the sequence used to po- sition the variant, the letter g means that the sequence is genomic, by opposition to p for instance which is used for a protein sequence, 11087877 corresponds to the posi- tion in the referred sequence, and G>T describes the substitution itself (http:// www.hegvs.org/mutnomen/recs.html) [14]. However this nomenclature has not been universally adopted yet. Previous nomenclatures sometimes subsist for historical rea- sons. For example our variant is still found in OMIM as the “FH NAPLES” or “Gly544Val”, that is to say with denominations related to the historical context of its discovery. In addition, private and disease- or locus-specific databases continue using non-conventional representations that enlarge the set of possible nomenclatures. Fig- ure 2 illustrates the numerous alternative manners of designating a unique genome variant in private and public databases. It is worth noting that some of the non- conventional notations (c) are ambiguous: the first one does not mention the reference nucleotide, the third and fourth ones refer to two different versions of the same protein. Finding intersection between several genomic variation databases is a critical issue for genetic diagnosis and “variome” exploration [15, 16]. However, as shown above. this task is not easy because of the amount of alternative and equivalent representa- tions. Thus a system capable of establishing equivalence i.e. aligning between the different representations of a given variant is needed for investigating genome varia- tions, and for being a basis for further pharmacogenomic studies.
(HGVS). According to this proposed standard, our variant should be described by the following expression: NC_000019.8:g.11087877G>T, where NC_000019.8 is the unique accession number (in the NCBI RefSeq database) of the sequence used to po- sition the variant, the letter g means that the sequence is genomic, by opposition to p for instance which is used for a protein sequence, 11087877 corresponds to the posi- tion in the referred sequence, and G>T describes the substitution itself (http:// www.hegvs.org/mutnomen/recs.html) [14]. However this nomenclature has not been universally adopted yet. Previous nomenclatures sometimes subsist for historical rea- sons. For example our variant is still found in OMIM as the “FH NAPLES” or “Gly544Val”, that is to say with denominations related to the historical context of its discovery. In addition, private and disease- or locus-specific databases continue using non-conventional representations that enlarge the set of possible nomenclatures. Fig- ure 2 illustrates the numerous alternative manners of designating a unique genome variant in private and public databases. It is worth noting that some of the non- conventional notations (c) are ambiguous: the first one does not mention the reference nucleotide, the third and fourth ones refer to two different versions of the same protein. Finding intersection between several genomic variation databases is a critical issue for genetic diagnosis and “variome” exploration [15, 16]. However, as shown above. this task is not easy because of the amount of alternative and equivalent representa- tions. Thus a system capable of establishing equivalence i.e. aligning between the different representations of a given variant is needed for investigating genome varia- tions, and for being a basis for further pharmacogenomic studies.
(1) The data preparation step consists in extracting the four input features from in- put data and depends on each specific source format. This preparation step also de- pends on whether the variant description is explicit (e.g. Genome-browser-like syntax or HGVS syntax) or implicit (e.g. database identifier). (1a) When the description is explicit, the four input features can be directly extracted by parsing the description according to a format-specific scheme. (1b) When the description is implicit, input data should first be completed in view of extracting input features. For example if the input data is a dbSNP identifier, it can be used to query the database and extract from the variant record the explicit data composing the input features. (2) Pivot features consist in the particular set of features obtained for a given vari-
(1) The data preparation step consists in extracting the four input features from in- put data and depends on each specific source format. This preparation step also de- pends on whether the variant description is explicit (e.g. Genome-browser-like syntax or HGVS syntax) or implicit (e.g. database identifier). (1a) When the description is explicit, the four input features can be directly extracted by parsing the description according to a format-specific scheme. (1b) When the description is implicit, input data should first be completed in view of extracting input features. For example if the input data is a dbSNP identifier, it can be used to query the database and extract from the variant record the explicit data composing the input features. (2) Pivot features consist in the particular set of features obtained for a given vari-
Figure 4 also shows the result of an experience carried on variants of a specific gene (named here geneA). Three sets of data were processed by the SNP-Converter application : 274 and 55 variants from private databases snp_base_| and snp_base_2 respectively, and 377 variants from dbSNP. Among the 706 assertions created by the wrapper, 671 could be qualified as original individuals, and 35 were found equivalent to existing individuals.
Figure 4 also shows the result of an experience carried on variants of a specific gene (named here geneA). Three sets of data were processed by the SNP-Converter application : 274 and 55 variants from private databases snp_base_| and snp_base_2 respectively, and 377 variants from dbSNP. Among the 706 assertions created by the wrapper, 671 could be qualified as original individuals, and 35 were found equivalent to existing individuals.
Fig. 1. Population, content and schematic relation of SABIO and SABIO-RK. SABIO contains general information about biochemical pathways and reactions in different organisms, including details about corresponding enzymes and reactants. Most of these data are collected from other databases like KEGG and UniProt. SABIO-RK extends SABIO by storing information about the reactions’ kinetic properties, such as the kinetic laws with their corresponding parameters and environmental conditions under which they were determined. manually extracted from scientific articles and verified by curators. At the moment it is very difficult to extract this information automatically, such as by the use of text mining technologies, given that most of the data are highly scattered through various publications and are frequently found in tables, formulas or graphs. However, we are working on the development of support tools, one of them for the identification of synonyms of chemical compound names, as we will describe in section 3.
Fig. 1. Population, content and schematic relation of SABIO and SABIO-RK. SABIO contains general information about biochemical pathways and reactions in different organisms, including details about corresponding enzymes and reactants. Most of these data are collected from other databases like KEGG and UniProt. SABIO-RK extends SABIO by storing information about the reactions’ kinetic properties, such as the kinetic laws with their corresponding parameters and environmental conditions under which they were determined. manually extracted from scientific articles and verified by curators. At the moment it is very difficult to extract this information automatically, such as by the use of text mining technologies, given that most of the data are highly scattered through various publications and are frequently found in tables, formulas or graphs. However, we are working on the development of support tools, one of them for the identification of synonyms of chemical compound names, as we will describe in section 3.
For consistency and to avoid duplicate entries, lists of compounds, organisms, tissues, compartments, kinetic law types and parameter units already existing in the SABIO-RK database are provided for selection at the input interface. These lists also contain synonyms referring to the same content to enable the search for alternative names of compounds, tissues etc. These may mean that the information presented to the database user is not exactly that included in the paper because the entries are presented with recommended names, however the user can always obtain the synonyms of the entries with multiple names. Already existing reactions can be searched in the database by defining one or two reaction participants. Enzymes variants catalysing the reactions are distinguished by a description of their subform, like wildtype or mutant protein species. They are named as wildtype or mutant followed by their name. Different isoforms of an enzyme are also named as wildtype followed by the name or abbreviation of the isoform. Furthermore, by including the specification of one or more accession number(s) of the UniProt database (if available), a direct link is provided to the properties of the enzyme. Additionally, the curators are confronted with missing or only partial information in the literature. For example a reaction definition can be incomplete supposing that only substrates of reactions are named without a definition of the reaction products. SF ste ah Ve: ye Be ee Derm enue mee: wel Dibecs sspreweme md aesttodiabcem Kaa etmaawesdlowwan ‘aiseawtkih tesn
For consistency and to avoid duplicate entries, lists of compounds, organisms, tissues, compartments, kinetic law types and parameter units already existing in the SABIO-RK database are provided for selection at the input interface. These lists also contain synonyms referring to the same content to enable the search for alternative names of compounds, tissues etc. These may mean that the information presented to the database user is not exactly that included in the paper because the entries are presented with recommended names, however the user can always obtain the synonyms of the entries with multiple names. Already existing reactions can be searched in the database by defining one or two reaction participants. Enzymes variants catalysing the reactions are distinguished by a description of their subform, like wildtype or mutant protein species. They are named as wildtype or mutant followed by their name. Different isoforms of an enzyme are also named as wildtype followed by the name or abbreviation of the isoform. Furthermore, by including the specification of one or more accession number(s) of the UniProt database (if available), a direct link is provided to the properties of the enzyme. Additionally, the curators are confronted with missing or only partial information in the literature. For example a reaction definition can be incomplete supposing that only substrates of reactions are named without a definition of the reaction products. SF ste ah Ve: ye Be ee Derm enue mee: wel Dibecs sspreweme md aesttodiabcem Kaa etmaawesdlowwan ‘aiseawtkih tesn
The SABIO-RK database has been designed to meet the Systems Biology community equirements. It aims to support modelers with high quality data in setting-up in-silico nodels describing biochemical reaction networks. The database enables complex earches for reactions, parameters, etc. and uses existing or defines new standards fot lata formats. Selected kinetic data can be exported in SBML format to build models or the simulation of complex biochemical processes. SABIO-RK also bundles nformation for researchers interested in comparing reaction kinetic data originating rom different sources.
The SABIO-RK database has been designed to meet the Systems Biology community equirements. It aims to support modelers with high quality data in setting-up in-silico nodels describing biochemical reaction networks. The database enables complex earches for reactions, parameters, etc. and uses existing or defines new standards fot lata formats. Selected kinetic data can be exported in SBML format to build models or the simulation of complex biochemical processes. SABIO-RK also bundles nformation for researchers interested in comparing reaction kinetic data originating rom different sources.
The ontology design and supporting applications described in the following sections are advancements to the SIBIOS system described in [13, 15, 18]. These enhancements aim into facilitating the ontology maintenance as new services are added to the system. In addition the ontology is deployed to support new components added to SIBIOS system.
The ontology design and supporting applications described in the following sections are advancements to the SIBIOS system described in [13, 15, 18]. These enhancements aim into facilitating the ontology maintenance as new services are added to the system. In addition the ontology is deployed to support new components added to SIBIOS system.
With the wide array of knowledge levels among biologists and bioinformaticians, it is necessary to allow a flexible system for service selection. Service browsing, the most common exploratory method supports classification of services by properties such as input, output, and task [17]. This approach provides a process that is adaptable to users of all levels. However, in order to provide a more robust querying interface, additional service descriptions and querying capabilities are necessary such as those supported by [12].
With the wide array of knowledge levels among biologists and bioinformaticians, it is necessary to allow a flexible system for service selection. Service browsing, the most common exploratory method supports classification of services by properties such as input, output, and task [17]. This approach provides a process that is adaptable to users of all levels. However, in order to provide a more robust querying interface, additional service descriptions and querying capabilities are necessary such as those supported by [12].
Service selection in SIBIOS utilizes both service browsing and advanced querying. RACERPro [22] with its reasoning capabilities is used to retrieve information from the ontology. Using browsing capabilities users can browse through the existing services categorized using either one of the available service properties. Figure 3.a shows part of the services classified using “has_input’ property. For each property a hierarchical structure is built based on the list of biological and bioinformatics terms
Service selection in SIBIOS utilizes both service browsing and advanced querying. RACERPro [22] with its reasoning capabilities is used to retrieve information from the ontology. Using browsing capabilities users can browse through the existing services categorized using either one of the available service properties. Figure 3.a shows part of the services classified using “has_input’ property. For each property a hierarchical structure is built based on the list of biological and bioinformatics terms
Fig. 4. An example of workflow composition in SIBIOS Service composition is the process of connecting services into a meaningfu workflow. SIBIOS ontology is deployed each time a service s is proposed to b scheduled after pervious services. Such composition is allowed only if a subset of thi input parameters of s matches a subset of the output parameters of each targetec previous service. To leverage on the hierarchical structure that characterizes th description of bioinformatics and biological terms, we utilize reasoning capabilities o Racerpro [22] to support an inclusive matching rather than an exact matching betwee! input/output parameters. For example if a previous service to s has sequence as outpu and one of the input parameters of service s is protein sequence then the compositio! between the services can occur. Similarly if the input of service s is sequence, then 1 should compose with a previous service that outputs protein sequence.
Fig. 4. An example of workflow composition in SIBIOS Service composition is the process of connecting services into a meaningfu workflow. SIBIOS ontology is deployed each time a service s is proposed to b scheduled after pervious services. Such composition is allowed only if a subset of thi input parameters of s matches a subset of the output parameters of each targetec previous service. To leverage on the hierarchical structure that characterizes th description of bioinformatics and biological terms, we utilize reasoning capabilities o Racerpro [22] to support an inclusive matching rather than an exact matching betwee! input/output parameters. For example if a previous service to s has sequence as outpu and one of the input parameters of service s is protein sequence then the compositio! between the services can occur. Similarly if the input of service s is sequence, then 1 should compose with a previous service that outputs protein sequence.
The present situation motivated us to develop a better model or data structure. It provides a network graph instead of a tree and offers direct access to problems like “which mRNAs lead to the same CDS” or “which mRNAs are likely to be disease- related due to ruptures or extensions in UTR regions” or “which Arabidopsis splice template belongs to a specific barley EST.”
The present situation motivated us to develop a better model or data structure. It provides a network graph instead of a tree and offers direct access to problems like “which mRNAs lead to the same CDS” or “which mRNAs are likely to be disease- related due to ruptures or extensions in UTR regions” or “which Arabidopsis splice template belongs to a specific barley EST.”
Fig. 3. Every sTMP has a unique splicing pattern, but the length of the first and the last exon remains undefined. The data structucture makes it possible to model different gene isoforms with the same splicing pattern. Crosslinks between genes refer to homologous genes. Noncoding and regulatory RNAs easily fit into the data structure. Qdld suuClure, Additionally, we connect to orthologous and paralogous genes via similarity or identity links, depending on the level of similarity. Combined with an ontology database, this settles our gene network data structure for complex queries. The Plant Data Warehouse ontology database fosters the investigation of gene families and subnets of interest to the users.
Fig. 3. Every sTMP has a unique splicing pattern, but the length of the first and the last exon remains undefined. The data structucture makes it possible to model different gene isoforms with the same splicing pattern. Crosslinks between genes refer to homologous genes. Noncoding and regulatory RNAs easily fit into the data structure. Qdld suuClure, Additionally, we connect to orthologous and paralogous genes via similarity or identity links, depending on the level of similarity. Combined with an ontology database, this settles our gene network data structure for complex queries. The Plant Data Warehouse ontology database fosters the investigation of gene families and subnets of interest to the users.
Table 1. RefSeq genes and isoforms in the human genome RNA genes are only sparsely available in the current RefSeq annotation. For example, all 79 snoRNAs annotated in the human genome are on chromosome 15. Since from other species (e.g. Arabidopsis) a spread over all chromosomes is known, we speculate (oder expect) that there are many more such genes in the human genome, but that up to now only the team that was responsible for the annotation of chromosome 15 considered them worthwhile. These RNAs are involved in processing and modification of other RNAs, such as ribosomal and small nuclear spliceosomal RNAs. SnoRNAs form a large family of relatively well-characterised non-coding RNAs (ncRNAs) [6].
Table 1. RefSeq genes and isoforms in the human genome RNA genes are only sparsely available in the current RefSeq annotation. For example, all 79 snoRNAs annotated in the human genome are on chromosome 15. Since from other species (e.g. Arabidopsis) a spread over all chromosomes is known, we speculate (oder expect) that there are many more such genes in the human genome, but that up to now only the team that was responsible for the annotation of chromosome 15 considered them worthwhile. These RNAs are involved in processing and modification of other RNAs, such as ribosomal and small nuclear spliceosomal RNAs. SnoRNAs form a large family of relatively well-characterised non-coding RNAs (ncRNAs) [6].
eS Se ee ee eS eee ee ee eee ee eee — version tiling path and can be seen as a revised and improved version for many genes. Therefore, we compare the RefSeq and the AltSplice annotation. One of the standard techniques for gene annotation is a spliced alignment of ESTs and cDNAs [9], [10] against the genomic template as shown in Fig. 4. It is slow and often unreliable in finding short exons, long introns, the true start codon, or the correct strand of single exon genes. In contrast, the sTMP data structure allows a simple query that checks existing gene annotations by in-silico splicing and aligning the resulting sTMPs with full-length cDNA. It allows to quickly find all low-error matching mRNA-cDNA pairs for whole genomes by calculating seed triggered [11] branch-and-bound k-band alignments. This was done with the separately downloaded RefSeq transcripts as cDNA database against the sTMPs derived from the RefSeq genome or the AltSplice annotation. As shown in Table 3, 6,614 conflicts exist between the RefSeq and the AltSplice annotation. Most of them consider only a few nucleotides at the boundaries of the first or the last exons, but 1,660 are severe exon count conflicts, meaning that the spliced alignment solution differs in the number of exons. Generally, the average quality of
eS Se ee ee eS eee ee ee eee ee eee — version tiling path and can be seen as a revised and improved version for many genes. Therefore, we compare the RefSeq and the AltSplice annotation. One of the standard techniques for gene annotation is a spliced alignment of ESTs and cDNAs [9], [10] against the genomic template as shown in Fig. 4. It is slow and often unreliable in finding short exons, long introns, the true start codon, or the correct strand of single exon genes. In contrast, the sTMP data structure allows a simple query that checks existing gene annotations by in-silico splicing and aligning the resulting sTMPs with full-length cDNA. It allows to quickly find all low-error matching mRNA-cDNA pairs for whole genomes by calculating seed triggered [11] branch-and-bound k-band alignments. This was done with the separately downloaded RefSeq transcripts as cDNA database against the sTMPs derived from the RefSeq genome or the AltSplice annotation. As shown in Table 3, 6,614 conflicts exist between the RefSeq and the AltSplice annotation. Most of them consider only a few nucleotides at the boundaries of the first or the last exons, but 1,660 are severe exon count conflicts, meaning that the spliced alignment solution differs in the number of exons. Generally, the average quality of
To process data and mappings, iFuice and thus BioFuice offer a set of high-level op- erators which can be combined within script programs. Table | shows a selection of these operators that are relevant for the examples in this paper; the full definitions are given in [Ra05]. The operators typically operate on a set of input objects, e.g. an entire LDS, and generate a set of output objects which can be used as the input of further op- erators. In Table 1, OI denotes a set of object instances from one object type; objects are identified by their ids which are assumed to also identify the LDS the objects belong to. To solve the EST classification problem posed in the beginning of this section we can use the following simple script determining three sets of EST sequences: Mappings can often be represented by sets of cross-references between ob- jects/instances of different LDS. For instance, the mapping between Gene @Ensembl and Protein@SwissProt can be derived from the existing SwissProt references in the Ensembl] gene instances. Alternatively, mappings can be derived on demand by exe- cuting queries or a program (script). In our example, the private source MyEstSet contains ESTs that are only described by a sequence. Hence, neither the PDS MyEst- Set nor the public PDS Ensembl provide correspondences between the instances they offer. In this case, a BLAST'-like tool can be used to determine for a set of EST se- quences the most similar DNA sequences within Ensembl. This creates a mapping between EstSequence @MyEstSet and SequenceRegion@EnsembI thereby integrating the local source into the peer-to-peer network represented by the SMM. Special map- ping types are so-called same-mappings interrelating semantically equivalent in- stances of the same object type. In Figure la there is one same-mapping on genes between Ensembl and NetAffx.
To process data and mappings, iFuice and thus BioFuice offer a set of high-level op- erators which can be combined within script programs. Table | shows a selection of these operators that are relevant for the examples in this paper; the full definitions are given in [Ra05]. The operators typically operate on a set of input objects, e.g. an entire LDS, and generate a set of output objects which can be used as the input of further op- erators. In Table 1, OI denotes a set of object instances from one object type; objects are identified by their ids which are assumed to also identify the LDS the objects belong to. To solve the EST classification problem posed in the beginning of this section we can use the following simple script determining three sets of EST sequences: Mappings can often be represented by sets of cross-references between ob- jects/instances of different LDS. For instance, the mapping between Gene @Ensembl and Protein@SwissProt can be derived from the existing SwissProt references in the Ensembl] gene instances. Alternatively, mappings can be derived on demand by exe- cuting queries or a program (script). In our example, the private source MyEstSet contains ESTs that are only described by a sequence. Hence, neither the PDS MyEst- Set nor the public PDS Ensembl provide correspondences between the instances they offer. In this case, a BLAST'-like tool can be used to determine for a set of EST se- quences the most similar DNA sequences within Ensembl. This creates a mapping between EstSequence @MyEstSet and SequenceRegion@EnsembI thereby integrating the local source into the peer-to-peer network represented by the SMM. Special map- ping types are so-called same-mappings interrelating semantically equivalent in- stances of the same object type. In Figure la there is one same-mapping on genes between Ensembl and NetAffx.
Figure 2 gives an overview of the BioFuice system architecture. It consists of twc main components, the iFuice core and the BioFuice query component. BioFuice util. izes the generic iFuice core to execute script programs and fuse data from severa sources. The BioFuice query component provides interactive query functionality supports local storage of analysis results and meets specific bioinformatics require: ments, e.g. to export genomic sequences in specific formats for later use in othe analysis tools. BioFuice query and iFuice core may run on the same machine or dif: ferent machines. The BioFuice query component can also be run in stand-alone mode independent of iFuice, e.g. offline on a laptop. In this case analysis and query process: ing are restricted to local data sources. BioFuice is already in use for different applica: tions, in particular for gene expression analysis, protein interaction analysis anc analysis of non-coding RNAs.
Figure 2 gives an overview of the BioFuice system architecture. It consists of twc main components, the iFuice core and the BioFuice query component. BioFuice util. izes the generic iFuice core to execute script programs and fuse data from severa sources. The BioFuice query component provides interactive query functionality supports local storage of analysis results and meets specific bioinformatics require: ments, e.g. to export genomic sequences in specific formats for later use in othe analysis tools. BioFuice query and iFuice core may run on the same machine or dif: ferent machines. The BioFuice query component can also be run in stand-alone mode independent of iFuice, e.g. offline on a laptop. In this case analysis and query process: ing are restricted to local data sources. BioFuice is already in use for different applica: tions, in particular for gene expression analysis, protein interaction analysis anc analysis of non-coding RNAs.
Objects of target LDS sharing the same object type can be aggregated, i.e. attribute sets of corresponding instances from these LDS are merged. BioFuice automatically determines available mapping paths from the source mapping model connecting the selected LDS and query targets; the paths are visualized on the right hand side for user selection. Before the query is executed the user can also specify whether the original sources or the local snapshots should be used to answer the specified query. Fig. 3. Selected BioFuice query capabilities
Objects of target LDS sharing the same object type can be aggregated, i.e. attribute sets of corresponding instances from these LDS are merged. BioFuice automatically determines available mapping paths from the source mapping model connecting the selected LDS and query targets; the paths are visualized on the right hand side for user selection. Before the query is executed the user can also specify whether the original sources or the local snapshots should be used to answer the specified query. Fig. 3. Selected BioFuice query capabilities
in this section we describe a method that supports similarity-based grouping ol biological data and that enables the development of grouping procedures. The components and the main steps of the method are illustrated in figure [I] The method uses as input the data source on which the grouping is performed. Further, it uses similarity functions that can compute similarity values between data values, and grouping attributes on which we base the computation of the similarity of data entries in the data source. There may also be external sources to support the grouping task. Based on this input the method can generate groupings of data. In addition to this, the method also allows the evaluation and analysis of the grouping results. For this purpose we use a library of known
in this section we describe a method that supports similarity-based grouping ol biological data and that enables the development of grouping procedures. The components and the main steps of the method are illustrated in figure [I] The method uses as input the data source on which the grouping is performed. Further, it uses similarity functions that can compute similarity values between data values, and grouping attributes on which we base the computation of the similarity of data entries in the data source. There may also be external sources to support the grouping task. Based on this input the method can generate groupings of data. In addition to this, the method also allows the evaluation and analysis of the grouping results. For this purpose we use a library of known
Table 3. Test case 1. Data entries in group 2. data entries Q01813, NP_000280, NP_002618, P10515, NP_000275, NP_005114 and NP_002603 to be similar to P08559 since the only GO term in the data entry P08559, GO:0005739 - the component term “Mitochondrion” , has high similarity with other component terms in the other data entries (marked by bold in table 3). Similarly, the data entries P11177 and NP_000275 are similar to P29802 because of the included GO:0004739. The comparison algorithm ignores the fact that some data entries have more GO terms assigned than the others. 2) The test case uses a grouping approach that assumes that similarity between data entries is transitive. As a result, two groups of similar data entries identified previously are connected by NP_000275 into a single group of similar data entries. 3) The fact that some GO terms in the annotation were general and that some of the data entries contained very few GO terms also caused the classes to be grouped into a single group. Bswamnila 9 A C'lace Tiictrihutad Amano Savaral Crniine Tn tact race
Table 3. Test case 1. Data entries in group 2. data entries Q01813, NP_000280, NP_002618, P10515, NP_000275, NP_005114 and NP_002603 to be similar to P08559 since the only GO term in the data entry P08559, GO:0005739 - the component term “Mitochondrion” , has high similarity with other component terms in the other data entries (marked by bold in table 3). Similarly, the data entries P11177 and NP_000275 are similar to P29802 because of the included GO:0004739. The comparison algorithm ignores the fact that some data entries have more GO terms assigned than the others. 2) The test case uses a grouping approach that assumes that similarity between data entries is transitive. As a result, two groups of similar data entries identified previously are connected by NP_000275 into a single group of similar data entries. 3) The fact that some GO terms in the annotation were general and that some of the data entries contained very few GO terms also caused the classes to be grouped into a single group. Bswamnila 9 A C'lace Tiictrihutad Amano Savaral Crniine Tn tact race
belonging to the complex vary in their function. For instance, FE, has the ma jor catalytic function, namely the pyruvate dehydrogenase function [BTS02]. Ir our case, P11177, NP_000275, P08559 and P29803 describe E,, while P1051! describes EH. The difference between functions is also reflected in the avail able GO annotations: P11177, NP_000275, P08559 and P29803 are annotatec with “pyruvate dehydrogenase (acetyl-transferring) activity”, while P10515 ha: “dihydrolipoyllysine-residue acetyltransferase activity”. These two terms are fa: from each other in GO. Test case 1 is the only one that organizes all the dat: entries into a single group. This illustrates how the knowledge from the com ponent and process GO ontologies may positively contribute to the groupings on function. For the other test cases, where only GO terms from the functio1 ontology were used, the available GO annotations are too specific to identify the whole enzyme complex. Example 3. Clique-Based Grouping. In test case 11 we used a clique-based algorithm for grouping similar data entries, which requires that all data entries in a group are similar to each other. The grouping result had a few cases where data entries in a class are distributed between 2 groups. For instance, the data entries of class 8, which describes Phosphoglycerate mutase, were distributed between 2 overlapping groups: one group contained P07738, Q8NOY7 and P15259, while the other group included P07738, Q8NOY7 and P18669 (figure [2). P15259 and P18669 are not found to be similar as the GO annotations differ from each other by GO:0042803 and GO:0004083, respectively. As a result, the data entries are moved to separate groups. P07738 and Q8NOY7 are included in both of the eroups as they are found to be similar to P15259 and P18669. We have checked that lowering the threshold, e.g. to 0.8, would combine all four data entries to a single group. This example illustrates the high impact of the different aspects in the grouping procedures on the grouping result; in this case the threshold, the
belonging to the complex vary in their function. For instance, FE, has the ma jor catalytic function, namely the pyruvate dehydrogenase function [BTS02]. Ir our case, P11177, NP_000275, P08559 and P29803 describe E,, while P1051! describes EH. The difference between functions is also reflected in the avail able GO annotations: P11177, NP_000275, P08559 and P29803 are annotatec with “pyruvate dehydrogenase (acetyl-transferring) activity”, while P10515 ha: “dihydrolipoyllysine-residue acetyltransferase activity”. These two terms are fa: from each other in GO. Test case 1 is the only one that organizes all the dat: entries into a single group. This illustrates how the knowledge from the com ponent and process GO ontologies may positively contribute to the groupings on function. For the other test cases, where only GO terms from the functio1 ontology were used, the available GO annotations are too specific to identify the whole enzyme complex. Example 3. Clique-Based Grouping. In test case 11 we used a clique-based algorithm for grouping similar data entries, which requires that all data entries in a group are similar to each other. The grouping result had a few cases where data entries in a class are distributed between 2 groups. For instance, the data entries of class 8, which describes Phosphoglycerate mutase, were distributed between 2 overlapping groups: one group contained P07738, Q8NOY7 and P15259, while the other group included P07738, Q8NOY7 and P18669 (figure [2). P15259 and P18669 are not found to be similar as the GO annotations differ from each other by GO:0042803 and GO:0004083, respectively. As a result, the data entries are moved to separate groups. P07738 and Q8NOY7 are included in both of the eroups as they are found to be similar to P15259 and P18669. We have checked that lowering the threshold, e.g. to 0.8, would combine all four data entries to a single group. This example illustrates the high impact of the different aspects in the grouping procedures on the grouping result; in this case the threshold, the
These cases are illustrated in Figure [4 The grammar of the language is Note that the brackets { and } have been placed in quotes when they ap- pear as terminals to avoid confusion with the repetition operator of the gram- mar. Condition have higher precedence that the structural operators, and — has precedence over |. Parentheses can be used whenever necessary. We use the short- cut # = [-t] * — (or, equivalently, —[t—]*), making the symbol # the notional equivalent of // in XPath. It is important to point out a few distinctive aspects of this DAG pattern
These cases are illustrated in Figure [4 The grammar of the language is Note that the brackets { and } have been placed in quotes when they ap- pear as terminals to avoid confusion with the repetition operator of the gram- mar. Condition have higher precedence that the structural operators, and — has precedence over |. Parentheses can be used whenever necessary. We use the short- cut # = [-t] * — (or, equivalently, —[t—]*), making the symbol # the notional equivalent of // in XPath. It is important to point out a few distinctive aspects of this DAG pattern
Fig. 6. Left: Adjacency list for undirected graph in Fig 3. Right: Adjacency list for the directed sraph in Fig 3.
Fig. 6. Left: Adjacency list for undirected graph in Fig 3. Right: Adjacency list for the directed sraph in Fig 3.
5.3 Single Pass Algorithm for Indexing the Database
5.3 Single Pass Algorithm for Indexing the Database
Fig. 11. Common edges after Step 1 for the example in Fig 10 Step 3. From the above computed common subgraphs, find a maximal subgraph and use it to compute the similarity measure based on Eq 2 (MCS similarity). Use the set of common edges to compute the similarity measure based on Eq 1 (Cosine The words edges and terms are used synonymously.
Fig. 11. Common edges after Step 1 for the example in Fig 10 Step 3. From the above computed common subgraphs, find a maximal subgraph and use it to compute the similarity measure based on Eq 2 (MCS similarity). Use the set of common edges to compute the similarity measure based on Eq 1 (Cosine The words edges and terms are used synonymously.
Similarity). Rank the pathways in descending order of similarity based on the similarity measure chosen by the user. Time Complexity Analysis for the Steps Described Above: Step 1.1: For the i’th edge in the query graph, let n; be the number of pathways that have the edge. Using the index associated with this edge, all pathways having this edge can be obtained. Thus, all edges in common between the database pathways and the query can be obtained in time O( (sum over all edges in the query) n;) = O(n), where n is the number of such common edges.
Similarity). Rank the pathways in descending order of similarity based on the similarity measure chosen by the user. Time Complexity Analysis for the Steps Described Above: Step 1.1: For the i’th edge in the query graph, let n; be the number of pathways that have the edge. Using the index associated with this edge, all pathways having this edge can be obtained. Thus, all edges in common between the database pathways and the query can be obtained in time O( (sum over all edges in the query) n;) = O(n), where n is the number of such common edges.
Fig. 1. Supported information flow in the Arevir system Clinical information systems can support decision-making by providing relevant and valid information at the right time and place [14]. These requirements are generally not met with systems used in the management of HIV-infected patients. Firstly, resis- tance testing is usually performed in specialized virologic laboratories. Secondly, clinical data management systems — if at all existent in electronic form — are not pre- pared to handle and interpret genotypic data. Thus, test results remain separated from the electronic patient record. This situation is unsatisfying not only with regard to rou- tine clinical decision-making, but also to research on optimizing therapies by means of incorporating genomic data.
Fig. 1. Supported information flow in the Arevir system Clinical information systems can support decision-making by providing relevant and valid information at the right time and place [14]. These requirements are generally not met with systems used in the management of HIV-infected patients. Firstly, resis- tance testing is usually performed in specialized virologic laboratories. Secondly, clinical data management systems — if at all existent in electronic form — are not pre- pared to handle and interpret genotypic data. Thus, test results remain separated from the electronic patient record. This situation is unsatisfying not only with regard to rou- tine clinical decision-making, but also to research on optimizing therapies by means of incorporating genomic data.
Fig. 2. Simplified entity-relationship diagram of the six main database modules comprising a total of 36 tables
Fig. 2. Simplified entity-relationship diagram of the six main database modules comprising a total of 36 tables
Fig. 2. A screenshot of the Ensembl ProtView and SPICE. A DAS track is available in Ensembl that shows which protein structures are aligned to the protein. By clicking on the DAS track the SPICE browser can be launched. SPICE shows the protein structure, the UniProt sequence, and the Ensembl predicted protein (ENSP) sequence. One of the exons is selected and its position can be projected onto the 3D structure.
Fig. 2. A screenshot of the Ensembl ProtView and SPICE. A DAS track is available in Ensembl that shows which protein structures are aligned to the protein. By clicking on the DAS track the SPICE browser can be launched. SPICE shows the protein structure, the UniProt sequence, and the Ensembl predicted protein (ENSP) sequence. One of the exons is selected and its position can be projected onto the 3D structure.
Fig. 2. Local user interface. The main window consists of two frames. The left one displays the list of available workflow modules on each server machine. The right bigger frame shows how to construct the workflow. We illustrate how to select the output port for the workflow module called as max.
Fig. 2. Local user interface. The main window consists of two frames. The left one displays the list of available workflow modules on each server machine. The right bigger frame shows how to construct the workflow. We illustrate how to select the output port for the workflow module called as max.
Fig. 3. REST and SOAP client interactions with the wavex web service Our results show (Fig. B) that the SOAP client requests are processed on average of 16% more slowly than requests from the REST client. The worse performance of the SOAP client in comparison with that of the REST one is evidently caused by requirement of the SOAP protocol to use XML for the data encapsulation. Conversion into XML format increases data volumes sent to server and thus slows down the performance. We believe that the difference in the SOAP and REST clients’ performance will be even more pronounced should slow running client computer or Internet connection be used.
Fig. 3. REST and SOAP client interactions with the wavex web service Our results show (Fig. B) that the SOAP client requests are processed on average of 16% more slowly than requests from the REST client. The worse performance of the SOAP client in comparison with that of the REST one is evidently caused by requirement of the SOAP protocol to use XML for the data encapsulation. Conversion into XML format increases data volumes sent to server and thus slows down the performance. We believe that the difference in the SOAP and REST clients’ performance will be even more pronounced should slow running client computer or Internet connection be used.
Fig. 4. The complexity evolution of MF ontology. All configurations are the same as figure 2.
Fig. 4. The complexity evolution of MF ontology. All configurations are the same as figure 2.
browsing data objects, services and namespaces (associated with data containers). The list of services is displayed as a browsable tree from which the user gains access to procedures. In the same way automatic interfaces are built foequested, the system provides notification about the progress status of services, -including historical re- cords of executed tasks-, together with the relationships between input data, service applied and output data. Frequently output data become the input for new services. The GUI provides a specific list of suitable services that can be applied.
browsing data objects, services and namespaces (associated with data containers). The list of services is displayed as a browsable tree from which the user gains access to procedures. In the same way automatic interfaces are built foequested, the system provides notification about the progress status of services, -including historical re- cords of executed tasks-, together with the relationships between input data, service applied and output data. Frequently output data become the input for new services. The GUI provides a specific list of suitable services that can be applied.
Fig. 2. Monitoring the Execution of a Workflow. Each row contains a service (processor) in the workflow and its state. Finished services (2397 and 2396) include links to their results. Then, the services in execution can be executed several times if an error occurs. Finally, the other services indicate which service is being waited for. Thus, the execution process of a workflow shows a set of services that are changing their status right up until the end of the execution, when results can be analyzed.
Fig. 2. Monitoring the Execution of a Workflow. Each row contains a service (processor) in the workflow and its state. Finished services (2397 and 2396) include links to their results. Then, the services in execution can be executed several times if an error occurs. Finally, the other services indicate which service is being waited for. Thus, the execution process of a workflow shows a set of services that are changing their status right up until the end of the execution, when results can be analyzed.
Fig. 3. Homology search and phylogenetic study workflow and its output showing the rela- tioships among the protein sequences related to SMN_HUMAN with the parameter ‘thresh- old’from runCreateTreeFromClustalw service fixed to one. The SMN proteins from different organisms appear to be related to human ones, and other proteins containing tudor domain and old related SMN proteins appear more similar to fish SMN (SMN_BRARE). Interesting issues are observed in this result regarding maternal tudor protein from Drosophila melanogaster (see text).
Fig. 3. Homology search and phylogenetic study workflow and its output showing the rela- tioships among the protein sequences related to SMN_HUMAN with the parameter ‘thresh- old’from runCreateTreeFromClustalw service fixed to one. The SMN proteins from different organisms appear to be related to human ones, and other proteins containing tudor domain and old related SMN proteins appear more similar to fish SMN (SMN_BRARE). Interesting issues are observed in this result regarding maternal tudor protein from Drosophila melanogaster (see text).
related to one of 57 representative predisposition genes (Vogelstein and Kinzler, 2004). With 100% recall, all 57 genes and 57 cancer types were represented in the BioLT dictionaries for genes/proteins and diseases. Ninety-five percent (95%) of the relationships were ranked in the top three results of up to thousands of hits. For the remaining three genes, the corresponding diseases were found in the fourth and fifth positions. If, for example, one compares the relationships between the protein PDGFRA and 81 co-occurring disease terms from the BioLT disease dictionary with the disease relation “familiar gastrointestinal stromal tumors” noted in the publication above, the following four co-occurrences are presented at the top of the results in BioLT: hypereosinophilic syndrome (hes) (13 hits), gastrointestinal stromal tumors (19 hits), systemic mast cell disease (7 hits), and systemic mastocytosis (5 hits). After the qualitative assessment of text-mining results, it is necessary to provide the possibility to manually validate and annotate results further. To base a decision- Table 1. The following example shows the different spellings and meanings the gene symbol “psp” has in the MEDLINE corpus (table threshold is set to 10 occurrences). The intended gene meanings are in bold font.
related to one of 57 representative predisposition genes (Vogelstein and Kinzler, 2004). With 100% recall, all 57 genes and 57 cancer types were represented in the BioLT dictionaries for genes/proteins and diseases. Ninety-five percent (95%) of the relationships were ranked in the top three results of up to thousands of hits. For the remaining three genes, the corresponding diseases were found in the fourth and fifth positions. If, for example, one compares the relationships between the protein PDGFRA and 81 co-occurring disease terms from the BioLT disease dictionary with the disease relation “familiar gastrointestinal stromal tumors” noted in the publication above, the following four co-occurrences are presented at the top of the results in BioLT: hypereosinophilic syndrome (hes) (13 hits), gastrointestinal stromal tumors (19 hits), systemic mast cell disease (7 hits), and systemic mastocytosis (5 hits). After the qualitative assessment of text-mining results, it is necessary to provide the possibility to manually validate and annotate results further. To base a decision- Table 1. The following example shows the different spellings and meanings the gene symbol “psp” has in the MEDLINE corpus (table threshold is set to 10 occurrences). The intended gene meanings are in bold font.
Fig. 1. Visualization of cancer terms and their classification using the “NCI thesaurus” ontology: PDGFRA was found to be associated with the disease “EGIST”, which is a synonymous term for the ontology concept “Extragastrointestinal Gastrointestinal Stromal Tumor’. That concept is derived from “Gastrointestinal Stromal Tumor”, which itself has synonymous terms that were found to be associated with the gene PDGFRA. The manually annotated evidence codes [8] for the automatically generated relations are partly displayed. The cancer terms were mapped to the “NCI thesaurus” ontology [5] (Fig. 1) to enable inferred searches for disease concepts. Gene-specific information was added by mapping the gene names dictionary to the EntrezGene database, integrated with the BioRS system. Knowledge provided by the KEGG and PubChem databases adds information about metabolic processes. The core data model of ““Gene-Disease” and “Gene-Compound” relationships was extended with relations to functional ontologies
Fig. 1. Visualization of cancer terms and their classification using the “NCI thesaurus” ontology: PDGFRA was found to be associated with the disease “EGIST”, which is a synonymous term for the ontology concept “Extragastrointestinal Gastrointestinal Stromal Tumor’. That concept is derived from “Gastrointestinal Stromal Tumor”, which itself has synonymous terms that were found to be associated with the gene PDGFRA. The manually annotated evidence codes [8] for the automatically generated relations are partly displayed. The cancer terms were mapped to the “NCI thesaurus” ontology [5] (Fig. 1) to enable inferred searches for disease concepts. Gene-specific information was added by mapping the gene names dictionary to the EntrezGene database, integrated with the BioRS system. Knowledge provided by the KEGG and PubChem databases adds information about metabolic processes. The core data model of ““Gene-Disease” and “Gene-Compound” relationships was extended with relations to functional ontologies
and test theories. Such scientific data integration procedures necessarily invoke scien- tific theories that cannot be inferred from schemas or data alone. For example, consider a systematist who wishes to use both genomic sequence data and morphological data in the process of inferring the evolutionary relationships among organisms. Instead of simply mapping DNA sequences and morphological data into a uniform data format, different processes may be applied to each data source to infer evolutionary (i.e., phy- logenetic) trees. The systematist then may use the assumption that the organisms have only one true set of evolutionary relationships, and that the phylogenetic trees inferred from genomic and morphological data approximate the true relationships. By employ- ing this theory, the researcher may “integrate” these distinct data sources by computing a consensus tree that reflects commonalities in the distinct phylogenetic trees inferred from the different data sources. These consensus trees (i.e., the resulting data product of integration) can then be analyzed further or applied in other studies. The challenge of integrating life-science data from multiple sources becomes even more daunting as disciplines become increasingly specialized and as more diverse types afiectantifie Asta are dacired Scientific worlkfawceveteme PTO ADD ODO OUI aim at fa-
and test theories. Such scientific data integration procedures necessarily invoke scien- tific theories that cannot be inferred from schemas or data alone. For example, consider a systematist who wishes to use both genomic sequence data and morphological data in the process of inferring the evolutionary relationships among organisms. Instead of simply mapping DNA sequences and morphological data into a uniform data format, different processes may be applied to each data source to infer evolutionary (i.e., phy- logenetic) trees. The systematist then may use the assumption that the organisms have only one true set of evolutionary relationships, and that the phylogenetic trees inferred from genomic and morphological data approximate the true relationships. By employ- ing this theory, the researcher may “integrate” these distinct data sources by computing a consensus tree that reflects commonalities in the distinct phylogenetic trees inferred from the different data sources. These consensus trees (i.e., the resulting data product of integration) can then be analyzed further or applied in other studies. The challenge of integrating life-science data from multiple sources becomes even more daunting as disciplines become increasingly specialized and as more diverse types afiectantifie Asta are dacired Scientific worlkfawceveteme PTO ADD ODO OUI aim at fa-
contain data tokens (labeled d; in Figure B), explicit metadata tokens (labeled m j in Figure[3), and other sub-collections (denoted using embedded control tokens, e.g., Dstart and beng). Metadata tokens are used to carry information that applies to the collections or data items that follow them in the stream. As shown in Figure[3] a series of actors may operate concurrently on the contents of collections. For example, in Figure] Actors 1- 4 all simultaneously process parts of collection a, Actors 2 and 3 each simultaneously process a part of collection c, and so on. Se Oy 1s ee, 0, a , , . ; a 2 ; a er ee
contain data tokens (labeled d; in Figure B), explicit metadata tokens (labeled m j in Figure[3), and other sub-collections (denoted using embedded control tokens, e.g., Dstart and beng). Metadata tokens are used to carry information that applies to the collections or data items that follow them in the stream. As shown in Figure[3] a series of actors may operate concurrently on the contents of collections. For example, in Figure] Actors 1- 4 all simultaneously process parts of collection a, Actors 2 and 3 each simultaneously process a part of collection c, and so on. Se Oy 1s ee, 0, a , , . ; a 2 ; a er ee
Fig. 4. A Kepver collection-oriented workflow for inferring phylogenetic trees
Fig. 4. A Kepver collection-oriented workflow for inferring phylogenetic trees
do not restrict the ordering of sub-items (collections or data items). A simple example of a collection schema and conforming instance are given i Fi gure[5] The schema, shown on the left, defines a PDB collection of interest as contain ing an optional header collection and one or more protein chain collections, where eac. protein chain contains one or more atoms having a name metadata value. A conformin: instance of the schema is shown on the right of Figure[5] The PDB collection instanc does not directly contain a protein chain, and instead contains multiple “molecule” col lections. Similarly, each protein chain does not directly contain an atom data item, an instead the atoms are nested within residue collections. Thus, unlike with XML Schem or XML DTDs, collection schemas allow instances to have additional items includin intermediate collections (e.g., matching PDBCollection//ProteinChain/ /Atom instea of PDBCollection/ProteinChain/Atom). ° With the caveat that metadata values, treated as attributes, “cascade” to nested items.
do not restrict the ordering of sub-items (collections or data items). A simple example of a collection schema and conforming instance are given i Fi gure[5] The schema, shown on the left, defines a PDB collection of interest as contain ing an optional header collection and one or more protein chain collections, where eac. protein chain contains one or more atoms having a name metadata value. A conformin: instance of the schema is shown on the right of Figure[5] The PDB collection instanc does not directly contain a protein chain, and instead contains multiple “molecule” col lections. Similarly, each protein chain does not directly contain an atom data item, an instead the atoms are nested within residue collections. Thus, unlike with XML Schem or XML DTDs, collection schemas allow instances to have additional items includin intermediate collections (e.g., matching PDBCollection//ProteinChain/ /Atom instea of PDBCollection/ProteinChain/Atom). ° With the caveat that metadata values, treated as attributes, “cascade” to nested items.
Fig. 8. The PaupHSearch, TreeReporter, and ComposeNexus collection-aware actors defined in terms of conventional actors nvocation, data values of the iteration-scope attributes are passed to the correspond- ng subworkflow input ports, the actor within the subworkflow operates on these data, ind the outputs written by the enclosed actor are accumulated by the subworkflow yutput port labeled String. The enclosing composite actor then inserts the output of he subworkflow back into the data stream. The labels on the ports specify the types ind quantity of data consumed or produced by the subworkflow, and provide anchors or mapping the iteration scope attributes to the ports. The PaupHSearch and TreeReporter composite actors are more sophisticated. Both
Fig. 8. The PaupHSearch, TreeReporter, and ComposeNexus collection-aware actors defined in terms of conventional actors nvocation, data values of the iteration-scope attributes are passed to the correspond- ng subworkflow input ports, the actor within the subworkflow operates on these data, ind the outputs written by the enclosed actor are accumulated by the subworkflow yutput port labeled String. The enclosing composite actor then inserts the output of he subworkflow back into the data stream. The labels on the ports specify the types ind quantity of data consumed or produced by the subworkflow, and provide anchors or mapping the iteration scope attributes to the ports. The PaupHSearch and TreeReporter composite actors are more sophisticated. Both
Fig. 1. Tree inference use case systematic biologists are attempting to develop a comprehensive history of life’s rigins by studying the phylogenetic relationships of the millions of earth species \ssembling these species and placing them on the “tree of life” requires increasec :mounts of information about each one as well as sophisticated analytical tools o build an understanding of the relationships among species. At present, the nfrastructure used to manage the flow of phylogenetic data lacks the querying ‘apabilities needed to address many important scientific challenges.
Fig. 1. Tree inference use case systematic biologists are attempting to develop a comprehensive history of life’s rigins by studying the phylogenetic relationships of the millions of earth species \ssembling these species and placing them on the “tree of life” requires increasec :mounts of information about each one as well as sophisticated analytical tools o build an understanding of the relationships among species. At present, the nfrastructure used to manage the flow of phylogenetic data lacks the querying ‘apabilities needed to address many important scientific challenges.
Fig. 3. Example of composite Step consider Fig. [3] In this workflow, Sc directly contains Sc, and transitively con- tains step classes S$; and Sz. The composite step class at the highest level, Sc, has input set {1,2} and output set {O,,O2}. Within Sc there is a composite step class Sc, which takes {J;} as input and produces {O,} as output; Sc also contains step class S3 which takes {Jz} as input and produces {O2} as output. Within Sc there is a step class S, which takes {7,} as input and produces {D} as output; {D} is then input to step class $2, which produces {O,} as output. Three examples of user classes for this workflow are:
Fig. 3. Example of composite Step consider Fig. [3] In this workflow, Sc directly contains Sc, and transitively con- tains step classes S$; and Sz. The composite step class at the highest level, Sc, has input set {1,2} and output set {O,,O2}. Within Sc there is a composite step class Sc, which takes {J;} as input and produces {O,} as output; Sc also contains step class S3 which takes {Jz} as input and produces {O2} as output. Within Sc there is a step class S, which takes {7,} as input and produces {D} as output; {D} is then input to step class $2, which produces {O,} as output. Three examples of user classes for this workflow are:
[able 1. Search fields for a GenBank record and corresponding XPath-expressions are related. For instance, a constraint could require that the organism should contain the string ‘Oncorhynchus’ while another query could require that the or- ganism should equal ‘Oncorhynchus mykiss’ and the tissue type equals ‘brain’. Clearly, the second constraint implies the first. So, we know that the first con- straint is true when the second is, and the second is false when the first is. Our optimization technique exploits these ideas to reduce the number of evaluations. More precisely, we define a graph structure that captures the relationships be- tween the constraints and consider two forms of propagation: false propagation and true propagation. We experimentally show that false propagation outper- forms true propagation and the pure streaming approach. Finally, we discuss how to incrementally maintain the containment graph. It
[able 1. Search fields for a GenBank record and corresponding XPath-expressions are related. For instance, a constraint could require that the organism should contain the string ‘Oncorhynchus’ while another query could require that the or- ganism should equal ‘Oncorhynchus mykiss’ and the tissue type equals ‘brain’. Clearly, the second constraint implies the first. So, we know that the first con- straint is true when the second is, and the second is false when the first is. Our optimization technique exploits these ideas to reduce the number of evaluations. More precisely, we define a graph structure that captures the relationships be- tween the constraints and consider two forms of propagation: false propagation and true propagation. We experimentally show that false propagation outper- forms true propagation and the pure streaming approach. Finally, we discuss how to incrementally maintain the containment graph. It
Table 2. Complex value representation of the GenBank record in Figure []
Table 2. Complex value representation of the GenBank record in Figure []
Insert query concerning GenBank XML Streaming Approach
Insert query concerning GenBank XML Streaming Approach
Related topics:
BioinformaticsText Mining and Information Retrieval
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved