This paper explains how to evaluate and analyse inconsistent Resource Description Framework (RDF) data by using EMAGE semantic (RDF) dataset as its use case. The author exploits the sub graph matching powers and mathematical functions of SPARQL query in evaluating inconsistent RDF data in a semantic dataset. He also proposes a mathematical method for calculating the amount of inconsistency in RDF data through a graph search approach. Finally, He analyzed the evaluated inconsistent RDF data.
EMAGE is a database of in situ gene expression data in the mouse embryo and an
accompanying suite of tools to search and analyze the data
(http://www.emouseatlas.org/emage/). EMAGE publishes in situ gene expression data
for the developmental mouse. Its data is collected through a scrutinized process which
involves assessing and tabulating of Biologist‟s experimental reports. These data
include reports on gene expressions in mouse experiments which are reported elsewhere
[
EMAGE‟s dataset can serve as a platform for Biologists to find solutions to the
causes of abnormalities in organisms. Biologists can suggest answers to the cause of
abnormalities in organisms through comparing the data indicating the strength of
expressed gene in a healthy organism with that of unhealthy organism [
Information in semantic dataset is represented by RDF data in the form of triples and stored in a triple store. A triple consists of subject, predicate and an object. An illustration of a RDF triple is as shown in figure 1 below. <http://www.cubist_project.eu/HWU#tissue_EMAP_42> <http://www.w3.org/2001/01/rdf-schema#label> “embryo” .
Each triple in RDF dataset represents a statement of a relationship between the entities denoted by the nodes that it links. RDF data can contain one or more triples. Each triple is composed of a subject, predict and an object. In RDF data, each subject of a triple is represented by a Universal Resource Identifier (URI) or blank node, each predicate is represented by a URI and each object node is represented by a URI, a blank node or a literal. For example in figure 1, the subject of the triple is a URI “http://www.cubist_project.eu/hwu#tissue_EMAP_42”, the predicate is a URI “http://www.w3.org/2000/rdf-schema#label”, and the object node is a literal “embryo”. The author adopted turtle serialization format (http://www.w3.org/TR/turtle/) in this example. RDF data has other serialization formats for representing its data such as N-Triple, N3, RDF/XML and RDFa.
Language Inconsistency exists in RDF data when the data does not conform to the rules governing their design. This is evident when there is a contradiction in the RDF data such that the RDF data contains both A and ⌐A.
Inconsistency in EMAGE dataset is identified through identifying data which do
not conform to EMAGE‟s textual annotation rules. These rules include the general
“detected somewhere in” and “not detected everywhere in” rules which are used to
propagate gene expression levels up and down the hierarchical structure of a
particular EMAP anatomy. In addition, the expression level of a gene in a particular structure
of a given Theiler stage in EMAGE dataset is reasoned through propagation approach.
Through propagation approach, the associated level of gene expression in tissues that
exhibit “is_part_of” relationship with other tissue(s) within a particular structure are
propagated up or down the given structure in line with the chosen level of gene
expression of that structure. As a consequence, gene expression levels could be
inconsistent. This can be as a result of positive propagation (expressions propagated up the
anatomy) that contradicts with an experimental result or negative propagation
(expressions propagated down the anatomy) that contradicts with an experimental result.
Also, gene expression can be completely contradictory (two experiments on the same
tissue in which a gene is stated as detected in one experiment and not detected in the
second experiment) or partly contradictory (two experiments on the same tissue in
which the genes detected have different expression levels). Also, Inconsistency in
EMAGE datasets has been categorized and defined [
In other to identify inconsistent RDF data, a subset of EMAGE RDF model dataset was stored in OWLIM-SE triple store (http://www.ontotext.com/owlim). The investigated dataset has 1,216,277 triples. The author applied appropriate SPARQL queries as to retrieve inconsistent data from the stored RDF dataset. He was able to detect binary inconsistency in the investigated dataset in some tissues which have “is_part_of” relationship with other tissues of the same hierarchical annotation structure. In these tissues, a gene is specified as “detected” and also specified as “not detected” in their related tissue. An example of EMAGE hierarchical annotation structure is shown in the figure 2 below. The SPARQL query in figure 3 identifies RDF data with binary inconsistency from Theiler stage 15 of the investigated HWU RDF model dataset. It can also be applied to any other Theiler stage by changing the Theiler stage number in the statement under label #3 of the query. Table 1 displays the result set. The author used the hash key (#) together with a unique number in the SPARQL query to identify comments that explain the SPARQL statement(s).
To illustrate the different types of inconsistent data in the investigated dataset, the author used instances from Theiler stage 15. Figure 2 shows a subset of EMAP anatomy of Theiler stage 15. #1 Declare URI namespace prefix hwu: <http://www.cubist_project.eu/HWU#> prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> prefix rdf: http://www.w3.org/1999/02/22-rdf-syntax-ns# #2 Select variables whose bindings are returned as solutions of the query SELECT DISTINCT ?gene_label ?t_label ?t_Experiment_label ?gene_strength ?t2_label ?t2_Experiment_label ?gene_strength2 where { { #3 Select a set of triple pattern that depicts the investigated RDF data: Set ‘A’ ?x rdf:type hwu:Textual_Annotation ; hwu:belongs_to_experiment ?y ; hwu:in_tissue ?z ; hwu:has_involved_gene ?g ; hwu:has_strength ?gene_strength . ?z hwu:has_theiler_stage hwu:theiler_stage_15 ; rdfs:label ?t_label . ?y rdfs:label ?t_Experiment_label . ?g rdfs:label ?gene_label } OPTIONAL #4 SPARQL key word which enables optional match { #5 Select optional variables contradicting set ‘A’ in another set: set ‘B’ ?b rdf:type hwu:Textual_Annotation ; hwu:belongs_to_experiment ?y2 ; hwu:in_tissue ?ztissue2 ; hwu:has_involved_gene ?g ; hwu:has_strength ?gene_strength2 . ?ztissue2 rdfs:label ?t2_label . ?y2 rdfs:label ?t2_Experiment_label . #6 Stipulate the relationship between set ‘A’ and set ‘B’ ?z hwu:is_part_of ?ztissue2 . #7 Stipulate the necessary condition that can ascertain any #7 possible contradictory values between set ‘A’ and set ‘B’ Filter(?gene_strength = hwu:level_detected && ?gene_strength2 = hwu:level_not_detected ) } } #8 Aggregate values of variables to be returned group by ?gene_label ?t_label ?t_Experiment_label ?gene_strength ?t2_label ?t2_Experiment_label ?gene_strength2 #9 Restrict expected results to allow only the output of contradictory values having ((round((count(?t2_label))/(count(?t_label))*100)) > 0) #10 Establish the order for the result set order by ?gene_label The result set in table 1 above, shows identified binary inconsistent RDF data in Theiler stage 15. As an example, some tissues (future midbrain and future rhombencephalon of experiments EMAGE:3530 and EMAGE:3879 respectively) with involved gene “Pax2” whose expression level are specified as “level_detected” were identified. Future midbrain and Future rhombencephalon have the same involved gene “Pax2” and a “is_part_of” relationship with the tissue “Future brain” whose expression level is specified as “level_not_detected” in EMAGE:984. These expression levels of Pax2 as specified in these experiments contradict each other and do not abide with the semantics of the word “is_part_of” as utilized by EMAP. In addition, analogue inconsistency was detected in the investigated dataset in some tissues which have “is_part_of” relationship with other tissues. The identified analogue inconsistent data involve a gene with varied strength levels such as “strong” and “moderate” in tissues that have “is_part_of” relationship with other tissues. Analogue inconsistency in RDF data from Theiler stages 15 of the investigated dataset was identified by substituting the filter condition under label #7 of figure 3 with the below filter condition: Filter(?gene_strength = hwu:level_strong && ?gene_strength2 = hwu:level_weak || ?gene_strength = hwu:level_moderate && ?gene_strength2 = hwu:level_weak || ?gene_strength = hwu:level_strong && ?gene_strength2 = hwu:level_moderate)
The result set in table 2, shows the identified analogue inconsistent RDF data in Theiler stage 15. As an example from the table, some tissues (Branchial arch and limb of experiment EMAGE:5349) with involved gene “Fkbp3” whose expression levels are specified as “level_strong” have been identified from the investigated dataset. Branchial arch and Limb have “is_part_of” relationship with the tissue Embryo. Yet, Fkbp3 has a level of expression “level_weak” in Embryo in the same experiment. These expression levels of “Fkbp3” as specified in the experiment contradict each other and do not abide with the semantics of the word “is_part_of” as utilized by EMAP. Examples from other EMAGE inconsistency types include the inconsistency from positive propagation: Gene “Pax2” was “detected” in Future midbrain in EMAGE:3879 and “not detected” in Future brain in EMAGE:984 (Table 1). Future midbrain is part of future brain and it is located at a lower part to future brain in the anatomy structure of Theiler stage 15 (figure 2). Future brain should unavoidably have the same gene expression as future midbrain if gene expression is to be propagated up the anatomy. The strength level of future brain is therefore contradicted by not fully propagating Future midbrain’s gene expression level up the anatomy and this result to „an inconsistency of positive propagation‟. On the other hand, Future midbrain should unavoidably have the same gene expression level as future brain if gene expression is to be propagated down the anatomy. The strength level of future midbrain was contradicted by not fully propagating the gene expression level in future brain down the anatomy and this result to „an inconsistency of negative propagation‟. Figure 2 shows the tree illustrating the hierarchical structure of future midbrain and future brain in Theiler stage 15.
4
There are two main methods of dealing with inconsistent data in a dataset: to diagnose
and repair it, and reasoning with the inconsistency [
Obviously, measuring inconsistency has been proven useful in analyzing diverse
range of information types such as news reports [
The amount of inconsistency in an investigated RDF data can be measured by evaluating the amount of contradiction in the RDF data against the likelihood of the contradiction to occur. This amount is assessed herein by calculating their ratio as a fraction of 100. The result educates us on how large/small the embedded contradiction in the RDF data is. As stated above, the amount of inconsistency in EMAGE‟s data from a graph based approach is herein assessed through adopting the mathematical and sub graph matching powers of SPARQL queries. This approach can be applied to all RDF dataset formats. It necessitates proper SPARQL query skills and adequate knowledge of the dataset by the dataset analyst. The amount of contradictions in the data under investigation against its total possibility to occur in the dataset is calculated as follows:
Xm =A RDF graph pattern in a RDF dataset Xk = Contradictory sub graph of Xm The interest is in calculating the amount of Xk in Xm such that = Total number of contradictions in Xk = Total number of occurrence of Xm in the dataset
Amount of Inconsistency in Xm = Xk Xm
In this investigation, the question “what amount of binary or analogue contradiction is present in the expression levels of the genes in each tissue experiment of Theiler stage 15” is answered. The amount of Binary or analogue inconsistency in RDF data from any of the Theiler stages of the investigated dataset is identified by adding the following SPARQL statement before label #2 of figure 3.
And also substituting the aggregation statement under the label #8 of the query with the below statement:
The result set of the administered query on Theiler stage 15 is as displayed in table 3 and 4 below. Table 3 above, gives a more clarifying result set of each inconsistent experiment in Theiler stage 15 of the dataset than table 1. Rather than listing inconsistent experiments singly (like in table 1), the amount of its occurrence in the RDF data with the stipulated pattern is measured. These measures inform us of the amount of inconsistent assays in each tissue experiment of a particular Theiler stage in the dataset. As an example in EMAGE:3530, it can reliably be stated that half (50%) of the assays are binary inconsistent. While in EMAGE:3879, less than half (33%) of the assays results are binary inconsistent. Consequently, decisions by Biologists to carry out further test or to remove existing experimental results from the dataset can be made.
A tissue experiment can have several assays. The author‟s approach identifies the amount of these inconsistent assay(s) in their corresponding tissue experiment. For example, in Theiler stage 15 of the investigated dataset, there are 6 assays on EMAGE:3879, and 2 of them are binary inconsistent thus the amount of inconsistency in the experiment is calculated by dividing 2 with 6 and multiplied the result by 100. The importance of identifying the amount of inconsistency in a tissue experiment is to identify how valid the assay results of a particular experiment are.
Evaluating and analyzing inconsistent RDF data of a RDF model dataset is a field yet to be explored. Interestingly, it has been shown in this paper that the measure and analysis of inconsistent RDF data gives an insight to the soundness of the information under investigation. Nevertheless, the author hopes to improve on this research by automating these processes of identifying, evaluating and analyzing inconsistent RDF data.
The author acknowledges the partners of CUBIST project especially Heriot-Watt University and Sheffield Hallam University for their support and provision of his research datasets. He also acknowledges his two PhD supervisors “Simon Andrews” and “Simon Polovina" for their invaluable contributions and review of this work.