The life of individuals are strongly in uenced by their health. An example concerns salinity resistant plants, an invention which may alleviate issues of climate change and rising sea levels. A di erent issue conserns drug discovery for humans, such as accurate and inexpensive cures available for the poor, personalized drugs, etc. In drug discovery the applied strategy is to combine domain experts with data made accessible through o -the-shelf software, and from the latter expect to identify new drugs. While computational drug-discovery is known to be working when number of candidate-factors are su ciently small, the established methods and software are unfeasible for mining in big-data knowledge bases. In this paper we address the above issue. We present an holistic approach for searches in big-data with complex relations. We demonstrate how our novel strategies for integration of large heterogeneous datasets results in knowledge discovery. In our work we address issues of: semantics, entity similarity, clustering, data-engine, hypothesis testing, and user-interfaces. To verify our approach we implement data from 37 external data-resources, resulting in a database with more than 30 million bio-medical relationships. When we compare our ndings with existing literature we observe how our holistic approach for big-data mining discover 1000+ novel candidates for drug interaction. To address keyissues in knowledge discovery we have constructed 10+ new softwareapproaches for data-mining, tools which enable the development of a new method for mining of big-data. To enable reuse of our approaches, they are available from: http://www.knittingTools.org/, http://www. knittingTools.org/gui_lib_mine.cgi, https://bitbucket.org/oekseth/ mine-data-analysis/downloads/, and https://bitbucket.org/oekseth/ hplysis-cluster-analysis-software.
In life-science a recurring task is to understand how and why entities relate: to
construct a hypothesis which translates discrete observations into a conceptual
gure capturing core-traits of an evaluated subject, as exempli ed in Fig. 3 for
the research of [
Copyright held by the author(s). NOBIDS 2017
Alzheimer [
In contrast to established approaches for data-mining, an understanding of
drug-interactions require the analysis of possible interactions, as exempli ed in
Fig. 1. While the \PubMed" database [
The high cost of drug development discourage the development of drugs for
the poor [
In this paper we relate the above perspectives, demonstrating how a new holistic approach for mining of big data enable user-interactive drug discovery. In the method and associated software we unify the approaches of user-centric and software-centric approaches for data-mining, as depicted in Fig. 5. What we assert is that an holistic approach which increase accuracy and performance of data from disparate sources, software for mining, and tacit understanding, is su cient to address major issues in drug discovery, a view supported by [15]. \R&D e ciency represents the ability of an R&D system to translate inputs (for example, ideas, investments, e ort) into de ned outputs (for example, internal milestones that represent resolved uncertainty for a given project or product launches), generally over a de ned period of time" [15]. The ensemble of methods and software, summarized in Fig. 5, address challenges which have prevented established semantic data-bases from knowledge discovery, e.g., as observed with respect to the issues encountered by [17,?,19].
In the work we have identi ed and addressed the issues of: 1. Disparate data: automatic approaches to unify distinctively di erent data, where results are exempli ed in Fig. 1. 2. Execution-time: high-performance software for accurate and large-scale datamining, as exempli ed in Fig. 4; 3. User searches: interactive real-time data-mining which stimulate use of tacit knowledge, as exempli ed in Fig. 3 and Fig. 6.
The remainder of the paper is organised as follows. In section 2 we brie y survey related approaches, before we in section 3 describe the approach. In the result-section 4 we identify evaluate/discuss how the holistic approach address 10,000 s r i a )p8,000 X = e c n ta6,000 s i d , e t a ic4,000 d e r p ( f to2,000 n u o C 0 1
The semantic inference distance for predicates activates state change regulates state change active in pathway through casuality state transition cites research state transition for protein–protein 2 3 4 5 6 7 8
Semantic inference distance. current issues in big-data mining. This paper ends with a brief summary of observations in section 6. 2
A challenge in data-mining conserns the slow performance of software, as observed for [20] in Fig. 4. A possible explanation of the latter is an unawareness of high-performance software implementation strategies [21]. To exemplify, the major e orts in \systems biology is on developing fundamental computational and informatics tools" [22], an assertion motivated by how \a concerted e ort to bring all the useful tools for pathway analysis in a common platform is still missing" [23]. When combining the observations of ([22,?]) and Fig. 4 we realize how poor-performing software represents a hurdle in knowledge discovery. To summarize, we observe that established approaches for data-mining su ers from: 1 1. Disparate data: insu cient data-coverage and prediction, e.g., in [23,24,25,26,17]; 2. Execution-time: high query response-time, e.g., in [20,?];
warnings (ie, if any).
Construct result-representation. y it x e l p m o C l a c i g o L : d n e g e L
Caption: Above cartoon/figure depicts/illustrate the ‘flow of data’ initiated when a user requests the computation of a query: how a large number of ‘code/logical modules’ are involved in an
operation which from the outside (ie, by a non-expert) may be seen as a trivial task. Of interest is to observe the modules involved in the handling ‘of the user-request’: for the “web-interface”
CSS FandigHT.M2L:areC‘reospmonspiblue’ tfoarttheiolaynouat,lwhcileoJQmuepry,lJeaxvaiStcryipt avndePresrl-uCGsI huansdleesrt-hei nustere-inrpautc;ttheiousner-inipnut is krenceoivewd/hlaenddlegd eby thseePaerrl-CcGhI-ebass.ed T“wehb-seerver”
(currently running on an “Apache” ‘framework) which construct a valid KNIT-TSV query; the KNIT-TSV query is received by the “data-server::client” written in “C” (where the
prograambmoinvg-leanguagge uofr“Ce” isrcheolsaenttoessimpulifystheerin-tsereacatiorncwhithethse “Atpoachet” hweeb-secrvoerm),wphicuhitniatiatteisoancoamlmuniincaftioenrweithnKcniettingp-ToroolscSKeBs-sse.rveTr(ahnditshererafete-r sending the
query in question to the SKB-server); Knitting-Tools “data-server::computer” (which is written in the programming-languages of “C” and “C++”) accesses the ‘already loaded
SKB-daetac-otbjsect’t,bhefeore evoaluwatingothfe qiunerfyoanrd mretaurntinigotnhe,resaulnttodthec“doatma-seprvleer::xcliietnty”, whoicfh asgaoinfrtewturnathreer,esufltotortheuPserel-CrG-Iq-buaseedr“iweebs-sesrvuerp”.Tpheo“wretbe-sderver” (a)
relate the provenance-relationships to each of the relationships (thereby simplifying the viewing of the result), (b) construct “JavaScript” and HTML data-structures (for simplifying and
‘beauatiftyingw’thwewus.erkexnpeiritentceianndgusTero-inotelracstio.n)o,brefgor.e (c) Winithiatiinlge/conmstruactijnogdrifferreentsdeataa-rrecprhesenetatioonsr/latysoutsa(erge,tabilens,vheeast-tmeapds, raidnial viiemws, petcr).oTvheirnesgult is then
seen by the user (though the web-interface), where the JQeury and JavaScript support allow user-interaction (ie, without the data to be reloaded for each/every request).
performance of software categorized in the upper part of the gure, their return on
investment is limited, i.e., as captured from the gure. In contrast the majority of our
e orts are invested on improving the software modules depicted in the bottommost
part in the above gure, as exempli ed in Fig. 5.
3. User searches: user-interfaces which limits domain-experts from accurate
data-searches and result-interpretation, e.g., in [
In below we brie y examine the above issues, focusing on issues concerned with disparate data and execution-time. 2.1
There are more than 106 research-works concerned with data-mining . An
exam1
ple concerns the k-means cluster-algorithm, where new permutations are
published every year, e.g., with respect to [
Observation from searching on \Google Scholar" for terms such as PCA, k-means,
Sum of Squared Error (SSE), spearman, Euclid, correlation, similarity, etc.
with respect to issues in missing data and similarity-metrics. In order to
evaluate accuracy of cluster-algorithms, application of feature-selection, and
manydimensional hypothesis-testing, metrics for cluster-consistency are used [
A major challenge in big-data analytics concerns the slow performance of
databaseengines [
Method: A holistic method for knowledge discovery
In the integration of real-time user access to 30 million bio-medical relationships
we have faced the challenges described in research, as exempli ed in section 2.
From the works of others we realize that it is not feasible to follow the
established strategies. To exemplify, major e orts by [17,?,19] are placed on
translating data-formats into RDF. However, their approaches have not resulted in
knowledge discovery. In the uni cation of data-resources we have addressed
issues in assimilation of the graph-structured \BioPax" [16] formats and
evidenceannotations in \OBO" [
A di erent aspect concerns the execution-time of user-queries, exempli ed
in Fig. 2. To address the high time-cost of translating external data-bases into
RDF, and searching RDF data-stores, we have designed a data-engine which
accepts semantic relationships. When measuring the response-time of queries we
observe how our new data-engine address issues in execution-time, as described
in our [
The above described strategy exemplify approaches which reduce searchtime without introducing erroneous heuristics, e.g., in contrast to [18]. Fig. 5 presents a summary of the approaches undertaken to optimize the performance aspects of bio-medical knowledge discovery, hence a holistic method for datamining. To exemplify, we from Fig. 5 observe how the holistic approach address issues in disparate data through a combination of manually curated rules (to address quality issues in data-resources, e.g., the \HDR" use-case), application of clustering to unify entities both with respect to their database-resource (e.g., \uniprot"), etc.
1,000
800 ] s [ e 600 m i T r e s U 400 200 0
Time-cost of pariwise simliarity-metrics in established software-libraries
Kendall: “cluster-C” Kendall: “hpLysis” 100 200 300 400 500 600
Time to compute pairwise simliarity for a squared matrix The holistic approach for drug-discovery, introduced in this paper, is constructed to relate domain-experts to accurate interactions minted from big data. From below sub-sections we assert that the approach manages to correctly address the issues described in section 2.
There are more than 220 di erent knowledge sources in the bio-medical domain
[66]. Correctness and usability are characteristics which describe the most
popular tools for knowledge integration. An example tool is cPath, written by [66],
which is built around a MySQL databas1e [67]. From the performance
measurements of data-structures in sub-section 4.4 we observe how semantic searches
through MySQL results in a 10,000,000x+ performance-delay. The best tools
provide access to data which has been manually curated by eld experts, such
as the Reactome tool [26] or the BioGrid tool [25]. The back-bone of the tools
is often the Gene Ontology (GO) [68], which is used to de ne the lexicographic
order of the genes and proteins. Translating compartmentalized knowledge into
an ontology for reasoning, such as the RDF format, is seen in [
In data-mining the execution-time of software may render high-quality analytical
approaches useless, e.g., as inferred for large data-sets in Fig. 4. The application
of established implementation-strategies results in under-performing code due
to the challenges of compilers to identify strategies for performance tuning (as
it otherwise would not have been a time-di erence between di erent software
implementations). In below we list a subset of observations from the holistic
optimization of approaches for data-mining:
1. Search-time: the test-cases listed in sub-section 4.4 relate the time-cost of
semantic searches in our novel database-engine to the established use of
Btrees ([
The above observations indicates that the application of low-level optimization strategies is a central part in e orts for mining of big-data, re ecting observations in section 2, hence the importance of an holistic optimization strategy. 4.3
A sound interaction between software-tools and human domain-experts is seen as an essential part of knowledge discovery. In below we exemplify a subset of the strategies we have applied in the holistic approach (Fig. 5): 1. Semantic user-interactions: Fig. 6 exemplify how users are provided with support for both semantic queries (topmost sub- gures), interactive exploration (sub- gures in the middle) and signature queries (bottom-right sub- gure); 2. Testing hypothesis: in Fig. 3 we observe how a combination of the \MINE" metric [69] and our web-based framework for data-mining facilitate knowledge discovery on ltered query-subsets. Pool of KBs and DBs Translating id’s into
Normalized column-based N(7) Standardization (sanitation)
Input to
Column-based
N(7) Translating axioms into
KT-axiom-spec KEnfSoTewtrrcaultenicvdsteuglareDeteasintato uSnyifnicoantyiomn
Knit ing-Tools -CORE(KT)
Legends:
Disparate data Execution-time User-searches - > Update KB-mappings. <- get normalized KB-identity Mapping between equal KB-identifiers Clustering into disjoint KB-entries
Knit ing-Tools -SKB quReeryala-tcimceess CoCmopraerilsaotinonand
Formats: OBO, Bel , CSV/TSV
Local and client-server access GUI-query-API
Visualisation API
Similarity
Clustering
Feature Selection and Group Assignment
Test hypothesis www
Terminal
Selection Criterias
Fig. 5: An holistic approach for data-mining. The above gure depicts a collection of labelled boxes, such as Pools of KBs and DBs and Visualization. The legend-text (topConstSruKcBtion of right) describe the classi cation of the di erent background rectangles, as discussed in section 3. An example of a classi cation concerns the process of handling disparate data, where tasks for data-parsing, entity optimization, and format-uni cation, are combined into an automated approach. The size of the background-boxes re ect their computational complexity, as discussed in Fig. 2. The uniqueness of our approach concerns how we relate the existing strategies into one uni ed model, thereby avoiding overheads associated with generalised approaches (such as RDF centered integrationstrategies). To exemplify, when approaches such as [19] apply Standardization they use standardized rules for all of the integrated data-resources, hence entities from di erent data-resources are syntactically correct while semantically inaccurate.
The results, summarized of our software, as listed in in this paper, may below: to validate and elaborate our be re-produced through application 4.4 1. 2. 3.
Semantics and data: MINE data-mining: MINE high-perform downloads/; Software for
The holistic approach represented in this paper manages to address issues in bigdata analysis. The term big data depends on the complexity of data, algorithms, and use-cases to be evaluated, e.g., where [70] asserts that a study of 857 proteins implies a large-scale analysis. This section therefore seeks to address strategies for: http://www.knittingTools.org/; http://www.knittingTools.org/gui_lib_mine.cgi; ance software: https://bitbucket.org/oekseth/mine- data- analysis/ data-analysis: https://bitbucket.org/oekseth/hplysis- cluster- analysis- software. how the holistic approach improves drug
Formats: OBO, Bel , CSV/TSV
Local and client-server access - A weakness of our approach, is the complexity of it, a complexity whic maintain several (??) isolated parts. - A strength of our layout, is the (relative) disjointness of the component makes it easier to verify cor ectness (and to some automatical y extend fi in naming of entities).
Concatenation of reasoning-res ults
Clustering
Proximity Centrality Interactive www.kn Filtering and Analysis Local and client-server access graph-based visualization (??) ANN combined with PCA Analysis GUI-query-API signature searches Paral el-plots
Heatmaps Dendograms
Legend-plot Circle-plots
Dynamic fetching
of node-information Visualization through www.knittingTools.org Fig. 6: A user-interface for semantic queries. The above sub- gures depict the userinterface for semantic searches, a user-interface designed for domain-experts without interest in programming. The two uppermost sub- gures depicts the query-form for submitting conditional questions. The drop-down menu observed in the top-left gure exempli es the support for auto-complete. In the middle-right sub- gure the queryresults is seen. To handle the frequent issue of 1000+ identi ed relationships, the table include lter-options. When a user lters a subset of the query-result, it is visualized in the middle-left sub- gure. 1. Disparate data: algorithms and data which may be accurately queried; 2. Execution-time: why the enabled performance-increase is important in drug discovery; 3. User searches: how domain-experts may identify accurate prediction from complex data. 5.1
The \Knitting-Tools" web-server includes a number of pre-computed use-cases (http://knittingtools.org/examples.cgi). The results have been manually investigated and veri ed. Accuracy of predictions depends on data (Fig. 1) and accessibility (Fig. 6). Below use-cases provides a brief introduction to how complex searches may be applied on big-data.
Use-case(1): What is known for \notch2"? (http://knittingtools.org/ query.cgi?queryID=get_allRelations_forA_vertex). The use-case illustrates an exploratory search to identify all relationships, synonyms and provenance (e.g., the set of databases) describing a vertex of interest, e.g., the \notch2". Ampli es the use of basic search functionality to fetch relationships in KnittingTools, both for visual evaluation (Fig. 6), and as a prior data-gathering step before application of software for pattern identi cation (sub-section 5.2 and subsection 5.3).
Use-case(2): What is known for \cdk4", and how was this known? (http:// knittingtools.org/query.cgi?queryID=get_allRelations_forA_vertex_evidence_ and_synonyms). Extends use-case(1) with logic to fetch the synonymous vertices (for each vertex in the set of identi ed relations) and the provenance for each relationships. Provides insight into why the identi ed relationships were predicted, i.e., their provenance.
question(3): Identify the regulations associated to the important event of apoptosis (i.e., 'controlled cell death'). (http://knittingtools.org/query.cgi? queryID=intro_basic_bio_2). The query identi es relationships associated to pathways and regulations for chemical entities, proteins, genes, and pathways. In the result provenance is associated to each relationship, a provenance which becomes visible when clicking the green-plus button in the result-table (Fig. 6). 5.2
The MINE software combine an highly accurate algorithm for pattern
identi cation [69] with a web-interface for interactive data-exploration (Fig. 3).
To evaluate the applicability of the MINE web-based software (http://www.
knittingTools.org/gui_lib_mine.cgi) the data-sets of [71], [72], and [
For large data-sets the prediction accuracy relates directly to the execution-time of data-mining software, i.e., a users are otherwise needed to explore smaller data-samples and test fewer number of hypothesis. The below paragraphs exempli es how our proposed approach increase accuracy of large-scale data-analysis.
Application(1). Large-scale ontology engineering (https://bitbucket.org/ oekseth/hplysis-ontology/). We have developed a new hpLysis-onto software for high-performance engineering of bio-medical ontology. Ontologies are used in a large number of application, e.g., to identify similarities of gene products from experimental outcomes [73] The hpLysis-onto software address performance issues in computation of transitive closures and transitive reductions, an issue hampering analysis of large and complex data-sets. For the task to compute transitive closures for all vertices, the software of [74] consumes more than one day on the 24 MB \Gene Ontology" [68]. In contrast, the hpLysis-onto software manages to answer the latter query in less than one second, hence a signi cant improvement in performance.
Application(2). Large scale Semantic similarity (https://bitbucket.
org/oekseth/hplysis-cluster-analysis-software). The hpLysis software is
updated with a new high-performance library for computation of 20+
semantic similarity metrics. \Generally speaking, semantic similarity measures involve
the GO tree topology, information content of GO terms, or a combination of
both" [75]. The software proposed by [75] takes several hours to complete. The
hpLysis-semantic software improves the performance of established software
approaches by 1000x+, i.e., without reducing the prediction accuracy. The latter is
enabled through increased utilization of computer memory hardware. Semantic
similarity-metrics are used to identify important traits in data-sets [76], e.g., to
(1) relate hypothetical assumptions to gene-expression-levels [
Application(3). Many-dimensional data-analysis ([
We have presented both a method, a database, and 10+ software, for datamining. This paper argue that the holistic approach, which captures an ensemble of approaches and high performance software, manages to overcome the current hurdles in big-data drug-discovery. Fig. 6 exemplify how domain-experts may interact with our real-time support for querying 30+ million bio-medical relationships. In order to evaluate the quality of the approach we investigate the number of accurate and unique relationships identi ed in our approach, exempli ed in Fig. 1. The 1000+ novel candidate interactions which are identi ed highlight the ability of our approach to automatically identify relationship which are not known in literature. Through an optimized data-engine the relationships are accessible for users in real-time, exempli ed in Fig. 3.
In this paper we have described an approach to unify our semantic interface (www.knittingTools.org) with our high-performance software application (e.g., https://bitbucket.org/oekseth/hplysis-cluster-analysis-software). Through concrete use-cases we have exempli ed how the approach address issues in disparate data, execution-time, and user searches, ie, parameters which are critical in discovery of knowledge. From the examples we observe how our method and 10+ novel software approaches address issues in big-data drugdiscovery. Therefore, we assert that our novel holistic approach may in uence strategies for mining of big-data. We plan to address the weakness of the user-interfaces and the unknown quality of our knowledge inferences. In order to improve our user-interfaces we are now initiating e orts in usability testing for di erent target groups. Similarily, we have initiated e orts to evaluate the drug-impact of our putative knowledge discoveries. Both of the issues require year-long lab-experiments, hence the importance of quality and performance enabled through our novel method and software.
The authors would like to thank MD K.I. Ekseth at UIO, Dr. O.V. Solberg at SINTEF, Dr. S.A. Aase at GE Healthcare, MD B.H. Helleberg at NTNU{ medical, Dr. Y. Dahl, Dr. T. Aalberg, Dr. J.C. Meyer, and K.T. Dragland at NTNU, and the High Performance Computing Group at NTNU for their support. 11. Herper, M.: The Truly Staggering Cost Of Inventing New Drugs. https://www.forbes.com/sites/matthewherper/2012/02/10/ the-truly-staggering-cost-of-inventing-new-drugs/#1a0129244a94 (2012) 12. DiMasi, J.A., Grabowski, H.G., Hansen, R.W.: The cost of drug development. New
England Journal of Medicine 372(20), 1972{1972 (2015) 13. Association, C.B.R., et al.: Fact sheet: New drug development process. FDA Special
Consumer Report 14. Thomas, K.: The price of health: the cost of developing new medicines. https://www.theguardian.com/healthcare-network/2016/mar/30/ new-drugs-development-costs-pharma (2016) 15. Paul, S.M., Mytelka, D.S., Dunwiddie, C.T., Persinger, C.C., Munos, B.H., Lindborg, S.R., Schacht, A.L.: How to improve r&d productivity: the pharmaceutical industry's grand challenge. Nature reviews. Drug discovery 9(3), 203 (2010) 16. Demir, E., Cary, M.P., Paley, S., Fukuda, K., Lemer, C., Vastrik, I., Wu, G., D'Eustachio, P., Schaefer, C., Luciano, J., et al.: The biopax community standard for pathway data sharing. Nature biotechnology 28(9), 935{942 (2010) 17. Blonde, W.: Metarel, an ontology facilitating advanced querying of biomedical knowledge. PhD thesis, Department of Mathematical Modelling, Statistics and Bioinformatics, Ghent University, Ghent, Belgium (2012) 18. Antezana, E., Blond, W., Egan~a, M., Rutherford, A., Stevens, R., De Baets, B., Mironov, V., Kuiper, M.: Biogateway: a semantic systems biology tool for the life sciences. BMC bioinformatics 10(10), 11 (2009) 19. Belleau, F., Nolin, M.-A., Tourigny, N., Rigault, P., Morissette, J.: Bio2rdf: towards a mashup to build bioinformatics knowledge systems. Journal of biomedical informatics 41(5), 706{716 (2008) 20. de Hoon, M.J., Imoto, S., Nolan, J., Miyano, S.: Open source clustering software.
Bioinformatics 20(9), 1453{1454 (2004) 21. Hucka, M., Finney, A., Sauro, H.M., Bolouri, H., Doyle, J.C., Kitano, H.: The erato systems biology workbench: enabling interaction and exchange between software tools for computational biology (2002) 22. Butcher, E.C., Berg, E.L., Kunkel, E.J.: Systems biology in drug discovery. Nature biotechnology 22(10), 1253 (2004) 23. Chowdhury, S., Sarkar, R.R.: Comparison of human cell signaling pathway databasesevolution, drawbacks and challenges. Database 2015 (2015) 24. Beisswanger, E., Lee, V., Kim, J.-J., Rebholz-Schuhmann, D., Splendiani, A., Dameron, O., Schulz, S., Hahn, U., et al.: Gene regulation ontology (gro): design principles and use cases. Studies in health technology and informatics 136, 9 (2008) 25. Stark, C., Breitkreutz, B.-J., Chatr-Aryamontri, A., Boucher, L., Oughtred, R., Livstone, M.S., Nixon, J., Van Auken, K., Wang, X., Shi, X., et al.: The biogrid interaction database: 2011 update. Nucleic acids research 39(suppl 1), 698{704 (2011) 26. Croft, D., OKelly, G., Wu, G., Haw, R., Gillespie, M., Matthews, L., Caudy, M., Garapati, P., Gopinath, G., Jassal, B., et al.: Reactome: a database of reactions, pathways and biological processes. Nucleic acids research 39(suppl 1), 691{697 (2011) 27. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al.: Scikit-learn: Machine learning in python. Journal of Machine Learning Research 12(Oct), 2825{2830 (2011) 63. Davis, L., Maizels, N.: Homology-directed repair of dna nicks via pathways distinct from canonical double-strand break repair. Proceedings of the National Academy of Sciences 111(10), 924{932 (2014) 64. Intel: SSE computer-hardware-low-level parallelism. https://software.intel.
com/sites/landingpage/IntrinsicsGuide/. Online; accessed 06. June 2017 65. Drepper, U.: What every programmer should know about memory. Red Hat, Inc 11, 2007 (2007) 66. Cerami, E., Bader, G., Gross, B., Sander, C.: cpath: open source software for collecting, storing, and querying biological pathways. BMC bioinformatics 7(1), 497 (2006) 67. MySQL: MySQL database engine. https://www.mysql.com/ (2017) 68. Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H., Cherry, J.M., Davis, A.P., Dolinski, K., Dwight, S.S., Eppig, J.T., et al.: Gene ontology: tool for the uni cation of biology. Nature genetics 25(1), 25{29 (2000) 69. Reshef, D.N., Reshef, Y.A., Finucane, H.K., Grossman, S.R., McVean, G., Turnbaugh, P.J., Lander, E.S., Mitzenmacher, M., Sabeti, P.C.: Detecting novel associations in large data sets. science 334(6062), 1518{1524 (2011) 70. Butland, G., Peregr n-Alvarez, J.M., Li, J., Yang, W., Yang, X., Canadien, V., Starostine, A., Richards, D., Beattie, B., Krogan, N., et al.: Interaction network containing conserved and essential protein complexes in escherichia coli. Nature 433(7025), 531{537 (2005) 71. Peaker, M., Taylor, E.: Sex ratio and litter size in the guinea-pig. Journal of reproduction and fertility 108(1), 63{67 (1996) 72. Wright, S., Chase, H.B.: On the genetics of the spotted pattern of the guinea pig.
Genetics 21(6), 758 (1936) 73. Pesquita, C., Faria, D., Falcao, A.O., Lord, P., Couto, F.M.: Semantic similarity in biomedical ontologies. PLoS computational biology 5(7), 1000443 (2009) 74. Antezana, E., Egana, M., Baets, B., Kuiper, M., Mironov, V.: Onto-perl: An api for supporting the development and analysis of bio-ontologies. Bioinformatics (2008) 75. Ehsani, R., Drabl s, F.: Topoicsim: a new semantic similarity measure based on gene ontology. BMC bioinformatics 17(1), 296 (2016) 76. Rada, R., Mili, H., Bicknell, E., Blettner, M.: Development and application of a metric on semantic nets. IEEE transactions on systems, man, and cybernetics 19(1), 17{30 (1989) 77. McInnes, B.T., Pedersen, T.: Evaluating measures of semantic similarity and relatedness to disambiguate terms in biomedical text. Journal of biomedical informatics 46(6), 1116{1124 (2013) 78. Aalberg, T., Zumer, M.: The value of marc data, or, challenges of frbrisation.
Journal of Documentation 69(6), 851{872 (2013)