Combining statistical techniques with semantic data representations holds the potential to enhance understandability of scienti c results. It can augment scienti c ndings with existing data sources in a reproducible manner through provenance capture, as well as enable further analysis and deduction through computer and human understandable de nitions of terms. We present a framework for semantically integrating and exploring numerical analyses. We call our work SemNExT for Semantic Numeric Exploration Technology. We apply our approach to data analysis aimed at improving understanding of human brain development that leverages the Cortecon RNA-Seq data repository. Our approach supports enrichment of Cortecon data through combinations with structured data sources available via SQL or SPARQL from the web to provide semantically enhanced analyses combined with statistical analyses. Our results are encoded as RDF graphs that may be used as input to reasoners and may drive provenance-aware visualizations. We introduce our infrastructure, describe its use on transcriptomic data analysis of a model of cerebral cortex development, and discuss some emerging suggestions for best practices and future research challenges.
Our work at the intersection of semantics and data analysis began after
discussions with successful data analysts detailed the level of human investigation
typically required. We hypothesized that some of the human interpretation and
vetting of results could be automated with semantic technologies and
structured web-available resources. We explore this hypothesis in the setting of brain
development data analysis. Recent work by [
Our Semantic Numeric Exploration Technology (SemNExT) framework is designed to facilitate statistical analyses that can be easily linked to external structured resources and published as linked data. This framework serves as an exemplar for complementing the work of statisticians with the use of provenance and linked data best practices. 2
BioGPS provides a pluggable workspace for interacting with di erent genomic
and proteomic datasets available on the World Wide Web [
Structured scienti c work ows such as Kepler [
Mathematicians have also produced modeling languages for mathematics,
such as MathML [
STATO is an ontology for modeling statistical tests and their applications,
as well as probability distributions, variables, spread, and variation metrics.6
Our modeling e orts complement the work done by the STATO authors, as
we introduce additional tests and provide an exemplar model that combines
statistical processes with the W3C's Provenance Model [
There is ongoing work at the Gene Ontology Consortium to develop a term enrichment protocol using REST principles and a JSON-based interchange format.7 The protocol allows for metadata exchange to describe the supported inputs and outputs supported by enrichment servers. There is also a JSON-LD context to enable interoperability with linked data clients. It lacks, however,
Analysis
Analysis
Factory Activity
Annotator
SemNExTAPI DataSource Factory
DataSource Entity
DatabaseSource
SPARQLSource a clear ability to expose provenance information to the client and a means of incorporating additional or alternative analyses.
Capadisli, Auer, and Riedl explore linked statistical data using RDF Data
Cube [
The SemNExT architecture is divided into six major components: (1) Data Source, (2) Analysis, (3) Annotator, (4) Provenance, (5) Web Service, and (6) Visualization. The interaction of these components is shown in Fig 1. Brie y, SemNExT instantiates a set of data sources and analyses using factories. Each data source provides one or more annotators that exploit the semantics of the underlying dataset to extract relevant attributes and links about entities. The act of annotating and analyzing entities are captured as activities and the results of these operations are cached and returned to the CHeM diagram interface. Data Source. SemNExT is an extensible framework for integrating datasets sourced from across the web. We describe the datasets used for our analyses and visualizations of transcriptomic data in Section 6.
Analysis. SemNExT can be set up to perform a variety of analyses on numeric
data provided by data sources via the rpy2 Python package. Section 4 describes
at a high level the di erent analyses used on the Cortecon genetic data.
Annotator. SemNExT data sources export Annotator objects that the
framework combines to extract information about entities across many datasets.
Annotators provide rules to help the framework order operations appropriately at
runtime. For example, UMLS provides relations and attributes for instances of
type umls-sn:Gene or Genome, i.a., so it contributes an Annotator that
annotates an entity in that class when an entity of that type is detected. Via
subclass relationships or inverse functional properties, the framework identi es
appropriate entities, extracts the associated relations and attributes, and returns
the enriched entity in the result. Subsection 6.5 presents an example.
Provenance. SemNExT heavily relies on the W3C recommended PROV
ontology [
Web Service. SemNExT provides a RESTful web services interface that is used for searching for diseases, genes, and KEGG pathways8 and for obtaining semantically enriched networks through the integration of datasets. Visualization. SemNExT provides a hybrid visualization called chord-heat map (CHeM) diagrams built using the JavaScript library D3. These diagrams will be discussed more in Section 7. 4
The Cortecon data set is an in-vitro stem cell model of human cerebral cortex
development [
To conduct the analysis, we rst normalize the data by gene into z-scores using the mean and sample standard deviation of each gene across time. We then performed SVD and FCM clustering. SVD, a form of dimensional reduction, reveals waves of gene transcription by plotting the genes in the space spanned by the rst two principal components. The angle of the genes within produces a temporal ordering of the genes. FCM extract clusters of related genes based on their temporal activity resulting in six clusters, which are correspond to stages of development Pluripotency, Ectoderm, Neural Di erentiation, Cortical Specication, Deep Layers, and Upper Layers. The analysis reveals a \transcriptomic 8 http://www.genome.jp/kegg/pathway.html sn:Analysis prov:used prov:Entity
sn:DataSource qb:Observation
qb:dataset prov:wasGeneratedBy prov:wasDerivedFrom
sn:ComputedObservation
clock" which allows one to understand the role of genes with respect to these
developmental stages which are analogous to hours in a clock. All stages except
Ectoderm were identi ed by van de Leemput et al. [
In order to link these data and statistical analyses with structured resources, we
begin by converting it from its original tabular form into an RDF graph structure
using the semantic techniques of Lebo et al. [
The code implementing the statistical analyses from Section 4 is modeled as prov:Entity and applications of analyses to observations are captured using prov:Activity. We make extensive use of STATO terminology (subclasses of obi:0200000 in Fig. 2) and include extensions to STATO where appropriate.10
Consider the task of normalization. The code to compute the z-score is modeled as a method in an R le that, given set of observations, computes the mean and standard deviation of the set, and returns the z-score for each observation. When this task is applied to the dataset of interest, the sn:Analysis used to model this activity is an instantiation of the transformation class stato:0000104 that uses the speci ed R code and input data. The output set of z-scores is derived from the input observations and modeled as sn:ComputedObservation.
The resulting RDF products are then made available as a dataset using a combination of the Vocabulary for Interlinked Datasets (VOID) and the W3Crecommended Data Catalog Vocabulary (DCAT) and Data Cube Vocabulary.
Databases
The DatabaseSource object is used by SemNExT to interface with relational
databases. We brie y describe each database incorporated by SemNExT.
Cortecon The Cortecon project from the Neural Stem Cell Institute provides
gene read data for a stem cell culture di erentiated into neurons during the rst
77 days of development [
Ensembl. Ensembl is a product of the EMBL - EBI and Wellcome Trust Sanger Institute to automatically annotate eukaryotic genomes.11 SemNExT uses its mappings from genes to transcripts to proteins to link di erent datasets. StringDB. StringDB is a protein-protein interaction (PPI) database containing both known and predicted interactions based on literature textmining.12 SemNExT uses StringDB as one of two background PPI knowledge bases. Uni ed Medical Language System (UMLS). UMLS provides a metathesaurus integrating healthcare and bioinformatics databases.13 SemNExT uses UMLS as a link set to bridge databases. 6.2
Linked Data
SemNExT integrates directly with a number of linked data resources via
the built-in SPARQLSource data source. The framework makes heavy use of
owl:sameAs links as well as inverse functional properties on identi ers to infer
sameness between entities in di erent datasets and drive the annotation process.
Bio2RDF. We make use of a number of linked life science resources through
the Bio2RDF project [
ReDrugS. The Repurposing Drugs with Semantics (ReDrugS) project14 is a rich
linked data resource that links Online Mendelian Inheritance in Man (OMIM),
IRefIndex, and DrugBank from Bio2RDF to identify likely paths for o -label
drug uses [
Uniprot. We use the SPARQL endpoint provided by the Uniprot Consortium15
to obtain curated Gene Ontology annotations and expression data [
Linking strategies Reuse of existing identi cation schemes for linking is a primary goal of SemNExT. Our URI schemes make extensive use of database identi ers for genes, e.g. Entrez IDs and Ensembl IDs. It is not always the case that a clear identi cation scheme is available for a particular class, however. SemNExT therefore uses three di erent strategies to perform text-based entity linking (e.g., by matching rdfs:labels), the most trivial being exact text and substring matching.
The matching algorithm extensively uses type hierarchies across data sources to limit the search space and return relevant matches. The SemNExT ontology is a hand curated ontology mapping that integrates relevant concepts from our input data sources to enable this work.
When multiple concepts in a dataset match an entity, the framework will exploit broader relationships in an attempt to nd a root, or archetype, node encompassing all of the matched nodes. This allows us to answer, for example, a single disease when many genetic variations may exist. If the user provides a speci c variation name, the more general concept will not be used to override it. 6.4
Enrichment analysis using linked data resources
After the data sources have been linked, we can use the structured resource
to perform a gene set enrichment analysis [
We also attempted to perform a similar analysis on the Tissue Speci city Annotations provided by Uniprot. Unfortunately, the annotations provided are textual rather than URI based, which leads to very poor results without additional processing in the form of natural language processing. 6.5
Linking Example: Septin-9 Consider the gene Septin-9 (SEPT9), which is an important gene in cell division and its mutations are implicated in some cancers. To build a network with this gene, SemNExT begins by searching all of its datasets for the identi er SEPT9 (e.g., dc:identi er ). It then queries for relations and attributes of SEPT9. For example, the system nds SEPT9's Entrez ID (10801) in Cortecon and adds that information to the gene's shared in-memory model. A UMLS database annotator then executes because it can map the Entrez ID to the UMLS concept identi er for SEPT9. Extracted information includes the fact that SEPT9 is part of the process \Cytokinesis" and that it has an Ensembl Gene ID (ENSG00000184640). 16 http://geneontology.org/ Ensembl's annotator triggers on the addition of the Ensembl Gene ID to the shared model, which allows the system to pick up the Gene Ontology annotations provided by Ensembl, for example \cytoskeleton." This procedure continues until no new information can be obtained from the background knowledge sources. 7
SemNExT provides a RESTful interface that produces JSON-LD descriptions of a genetic network that drives visualizations such as our Chord and Heat Map (CHeM) diagrams for understanding how mutations associated with disease may produce changes in cerebral cortex development implemented using D3.js.17 Figure 3 shows a CHeM diagram for Holoprosencephaly, a disease in which the brain fails to form two hemispheres. Genes are organized around the outside of the diagram according to the transcriptomic clock revealed by SVD. The stages of development found by FCM are shown by di erent colors in the inner band. 17 A demonstration is available at https://semnext.tw.rpi.edu/chords.html Chords between genes are PPI extracted from structured sources of proteincoding genes. The width of each gene varies with the number of connections. The diagram's outer ring shows the normalized RNA-Seq counts of each gene as a heatmap for each day of measurement, with day 0 closest to the center and day 77 in the outermost ring.
We chose to use JSON-LD as our primary serialization to exploit the D3 visualization toolkit while maintaining interoperability with semantic web capable agents. The framework publishes JSON-LD contexts for its network representation that makes use of best practice ontologies for linked statistics, including STAT-O, RDF Data Cube, PROV-O, and the Gene Ontology. 8
We evaluated our integration by determining the overlap of concepts across the data sources that SemNExT interfaces with. Table 1 presents a breakdown of the number of concepts identi ed in each dataset. Queries to match genes and proteins were limited to the species Homo Sapiens. Coverage was computed by pairwise selecting datasets containing a shared concept (disease, gene, protein, directed and undirected protein interaction) by using the techniques identi ed in Section 6 to infer owl:sameAs links. Intersections of all datasets sharing a concept were computed in a similar fashion, except that the entity being linked was required to exist in every dataset containing its concept to be counted.
We found that di erentiation of classes in di erent resources was a challenge. For example, Cortecon uses high level concepts such as Disease that in UMLS are decomposed further into concepts such as \Disease or Syndrome," \Congenital Abnormality," and \Pathologic Function." Where appropriate, we attempted to limit linking to relevant classes to improve search and integration responsiveness.
Lastly, edge extraction for the chord diagrams uses any available data source that asserts an edge between two proteins coded by genes. However, systems such as ReDrugS that we incorporate have more robust scoring mechanisms validated by biostatisticians that could further re ne appropriate edge choices. 9
The combination of statistics and semantics can provide robust error checking and synchronicity between members of a project. While migrating the CHeM visualizations from a Matlab version to the JSON-LD based API written in Python, a project developer noted discrepancies in the values of z-scores between the two implementations.
Listing 1.1. An excerpt of a provenance trace from a SemNExT analysis where z-score values were mismatched between Matlab and Python analyses.
Comparing the provenance of the two methods (Listing 1.1) made it clear that z-scores from the Matlab implementation were based on the corrected sample standard deviation18 whereas the Python implementation z-scores were based on an uncorrected sample standard deviation.19 This di erence in behavior was the result of Matlab's implementation defaulting to a normalization of N 1 whereas NumPy's normalized by a default of N . Due to the use of manual ontology mappings between data sources, our approach is limited in its direct applicability to other domains without a time investment by knowledge representation experts. However, the mapping we generated contained a limited number ( 100) of axioms so its construction was relatively short (on the order of hours). Larger, more complex domains may face scaling issues unless automated ontology mapping approaches are employed. 18 https://www.mathworks.com/help/matlab/ref/std.html 19 http://docs.scipy.org/doc/numpy/reference/generated/numpy.std.html
Using relational databases is another limitation. Annotators accessing databases must be hand-coded whereas SPARQL-based accesses make use of class hierarchies, class mappings, owl:sameAs assertions, and reuse of URIs. We leave exploration of tools to replace hand-coded SQL accesses (e.g., D2R) to future work and note that more resources are becoming available via SPARQL. 10
We modeled statistical analyses of genomic data using best-in-class ontologies. These statistical outputs were linked with additional structured resources using linked open data techniques. The resulting RDF graphs were then visualized using a combination of chord diagrams and heat maps. We evaluated our mapping techniques by comparing individual mappings inferred by our API to the total number of individuals across all datasets for a relevant selection of classes. Lastly, we identi ed some bene ts to being able to represent statistical and linked data transformations and their provenance in RDF.
We recommend a number of practices based on our experiences: (1) Statistical resources should be modeled with the RDF Data Cube and annotated with appropriate provenance and frameworks should take advantage of these structured resources when available; (2) JSON-LD provides a balance between the verbosity of existing semantic web serializations and the succinctness of formats such as CSV to incorporate semantics into web-based visualizations; and (3) Capturing of provenance of statistical operators, especially when such operations may have di ering default behavior between implementations, is important as it enables replicability and can aid in debugging analyses. 10.1
Future Work We are looking to expand the number of supported analyses and to apply the SemNExT framework to datasets outside of bioinformatics. SemNExT currently does not automatically reconcile concepts with adjective variations in labels that tend to violate traditional stemming rules, for example \Spinal cancer" in Cortecon compared with \Spine cancer" in UMLS. We intend to investigate resources such as Wordnet as an initial means of resolving these con icts and will explore other natural language processing techniques as necessary to improve the framework's ability to link resources without identi ers.
We are also interested in modeling the inputs and outputs of statistical analyses using languages such as OWL-S20. This will enable linking of relevant analyses at runtime rather than at code design time. 20 http://www.w3.org/Submission/OWL-S/ Mr. Patton was supported by an NSF Graduate Research Fellowship and RPI internal funds. Ms. Brown, Ms. H. De Los Santos and Dr. Bennett were supported by NSF Grant 1331023 and RPI internal funds.