Semantic Web technologies are core for the integration of disparate data resources. It can be used to exploit data from next generation sequencing (NGS) for therapeutic decisions regarding cancer. In this manuscript, we describe how di erent data resources, which inform on the expression of speci c genes in a tissue and its variants, can be brought together to indicate a risk for tissue-speci c cancer for NGS data. This approach can be used to judge patient genomic data against public reference data resources. The TCGA and COSMIC repositories are being processed to connect and query information concerning the expression of genes, copy number variants (CNV), and somatic mutations. We annotated sets of di erential expression data provided from the Illumina Body map 2.0 (HBM) concerning 16 di erent tissue types and identify genes with an RPKM (Reads Per Kilobase of transcript per Million mapped reads) value greater than 0.5 as measure indicating an associated risk for cancer. Thus, the di erential expressed genes from HBM can be associated with a tissue type and gene expressions in COSMIC and TCGA leading to a potential biomarker for that particular tissue speci c cancer. In the case of ovarian cancer, we retrieved the genomic positions (loci) and the associated genes of potential biomarker candidates, and suggest that this approach and platform can serve future studies well. Altogether, the presented linked annotation platform is the rst approach to represent the COSMIC data in an RDF format and to link the data with the TCGA datasets. The proposed approach enriches mutations by lling in missing links from COSMIC and TCGA datasets which in turn helped to map mutations with associated phenotypes.
Next Generation Sequencing (NGS) technologies open new diagnostic and
therapeutic ways for cancer research. However, the resulting high-throughput
sequencing data has to be processed in complex data analytics pipelines including
annotation services. Unfortunately, there is not yet a well-integrated platform
available for both clinical and translational [
In order to link and retrieve patterns of a gene and tissue speci c information from various cancer mutation (TCGA) and database with global mutation list and mutation type (COSMIC), we encountered the following three challenges: (i) to transform heterogeneous data repositories and their storage formats into standard RDF; (ii) to discover associations (aka. links) by nding speci c patterns (i.e. correlations) for a gene with regards to CNV, mutation and its gene expression data sets; and (iii) to query in a scalable way the large volume and frequently updating datasets covering 16 di erent tissue types and the gene expression data from di erent repositories.
The experiments conducted in this paper is aligned to the transcriptome study based on the Human Body Map 2.0 (HBM)3 from Illumina which covers the following tissues: adrenal, adipose, brain, breast, colon, heart, kidney, liver, lung, lymph, ovary, prostate, skeletal muscle, testes, thyroid, and white blood cells. The HBM provides gene speci c information across one or more tissue types and intends to support the identi cation of potential biomarker for targeted therapy. In this study our results not only depicts novel biological outcomes but also provides a linked annotation framework that assimilates clinical outcomes from related data repositories.
The rest of the paper is structured as follows: Section 2 motivates our working scenario exploring on the HBM use case and the annotation databases; Section 3 presents the methodology and architecture of the proposed functional annotation framework; Section 4 gives an evaluation of the functional annotation framework; Section 6 presents the related work in linking the TCGA repository and Section 7 draws the conclusion from our work.
In order to understand the outbreak of disease, in particular cancer, it is one approach to compare normal and diseased tissue samples to interpret the changes in the expression patterns of the genes with regards to the observed disease status. 1http://cancer.sanger.ac.uk/cosmic 2https://tcga-data.nci.nih.gov/tcga/ 3https://www.ebi.ac.uk/gxa/experiments/E-MTAB-513 In our case, HBM serves the purpose to identify similarities in gene expression patterns using the studies across di erent tissue types, where HBM discloses the similarities between human tissues on the molecular and genetic level. Due to overlaps between cancer behaviours, progression and mutated genes, we have annotated top 100 genes distilled by our ltering criteria with COSMIC to explore previously observed studies from TCGA database, e.g., somatic mutation, genomic loci and other mutations linked to these genes retrieved from healthy tissues.
Human Body Map (HBM) 2.0 from Illumina: HBM covers data from
transcriptome studies for 16 tissue types (see above). Samples for these 16
tissue types have been processed, aligned and nally expression level have been
determined [
The gene expression data extracted from HBM samples returns a very large
list of more than 52000 genes. For data processing reasons we chose to reduce
the list and therefore de ned the cut o for each RPKM value according to the
method suggested by Sandberg et. al[
Annotation databases (COSMIC & TCGA): The main focus of this work is the identi cation of patterns for cancer mutations (given by TCGA) and globally known mutations and their types (given by COSMIC) for selected di erentially expressed genes across di erent tissue types. Figure 1 shows the correspondences, i.e., the associations or connections that have been established between the TCGA and COSMIC databases for this purpose. For this task our primary concern have been the associations between the CNV, the known mutations and the gene expression data.
GENE_EXPRESSION CompositeElement REF Protein Expression
COMPLETE_MUTATION Hugo_Symbol Entrez_Gene_Id ChromosomeStart_Position UID Validation Status ……
CNV(COPYNUMBER) Chromosome Start End Segment_Mean
GENE_EXPRESSION ID_SAMPLE SAMPLE_NAME GENE_NAME REGULATION
COMPLETE_MUTATION Genename Samplename Tumourorigin TCGA_ID ID_tumour Samplesource …….
CNV(COPYNUMBER) ID_SAMPLE ID_tumour Primarysite Histology subtype MUT_TYPE Chromosome:G_Start. G_Stop
Fig. 1: Links between COSMIC and TCGA repositories
As part of this work, speci c basic curation for data re nements have been performed: we had to identify instances to link two databases or a couple of events within the databases (see g. 1). For example, MUTATION and GENE EXPRESSION data in COSMIC could be linked to GENE NAME but CNV had SAMPLE IDs as expected. Later we used SAMPLE IDs after rst iteration with GENE NAME. Also, chr:start end position and GENE NAME were used to link COSMIC and TCGA (see green arrows in Figure 1). The RDFized version (see section 3.1) has kept this redundancy problem to have FDR rate as low as possible.
The annotation architecture is summarized in Figure 2 showing all three major components. First, the RDFization component that generates Linked Data from the TCGA and COSMIC databases leading to several SPARQL endpoints for public use. Second, the linking component that searches and discovers correspondences between selected datasets (TCGA variants, COSMIC di expression, COSMIC Mutation, etc.). The links discovered by this component have an e ect on the e ciency in the source selection, on the query planning, and on the overall query execution in a decentralized setting. Third, the scalable query federation component: it a single-point-of-access through which distributed data sources can be queried in concerto.
Illumina Body Map 2.0 Differentially expressed genes (DEG)
SPARQL Query
SAFE Query Federation Engine
Results SumDmataaries
itrsvaaT_CnAG iiIfrssxS__eeCCpdnoOM ittSa_CCunoO M I M iitrrssxT__eeeCpnnpooAG s e ir o it s o p e R C I M S O C d n a A G C T d e k n i L
GENE_NAME Mutation_ID Sample_ID Tumor_ID Chr_start Chr_end TCGA_ID
GENE_EXPRESSION ion COMPLETE_MUTATION t CNV(COPYNUMBER) a z i FRD CCThhCrrG__Aset_naIdrDt
Variant_type Ref_Protien Expression TCGA
The scalable query federation is based on the SPARQL query federation
engine called SAFE [
RDFization COSMIC raw data les are delivered as tab separated text (tsv) and are being processed with the COSMIC RDFizer tool that generates the N3 triples for the SPARQL endpoint and statistical information related to the data. Only three types of data have been included, i.e., gene expression data, gene mutation and CNV data. Table 1 shows the overall statistics of the RDFization: row 1 represents for the COSMIC gene expression data the number of records (column 2), it's size (column 4), the corresponding triples generated (column 2) and again it's size. The other two rows represent the same type of data for the COSMIC gene mutation and CNV data. A total of 154 million records has been RDFized, producing approximately 1.2 billion triples. Row 5 represents the statistics for the RDF version of TCGA-OV (TCGA Ovarian), which forms a subset of the linked TCGA4 data. The RDF le for the COSMIC data can be made available inline with the COSMIC data policy. The main integration is based on owl:sameAs constructs as can be seen in the listing 1.1 where two COSMIC sample ids have been identi ed as being identical to two TCGA patient bar code ids. These links are at the core of facilitating data integration and the data analysis tasks.
Listing 1.1: COSMIC and TCGA Linking Example <Link 1> <Source>COSMIC</Source> <Target>TCGA OV</Target> <l i n k > http : / / cosmic . s e l s . i n s i g h t . org / schema / ID Sample /TCGA 13 0920 owl : sameAs http : / / t c g a . d e r i . i e /TCGA 13 0920 </l i n k > </Link 1> <Link 2> <Source>COSMIC</Source> <Target>TCGA OV</Target> <l i n k > http : / / cosmic . s e l s . i n s i g h t . org / schema / ID Sample /TCGA 24 1850 owl : sameAs http : / / t c g a . d e r i . i e /TCGA 24 1850 </l i n k > </Link 2> 3.3
Scalable query federation
SAFE has been developed for accessing sensitive clinical data in data cubes at
di erent locations [
The listing 1.2 shows a sample SPARQL query, which federates across COSMIC and TCGA data asking for genomic loci of a mutated gene by chromosome start points which then returns the disease metastasis information along with the mutation type. Answering such a query requires the integration of COSMIC with TCGA and merging results from both TCGA and COSMIC, and thus has to make use of query federation. The results for the rst four triples in the given query (i.e. cosmic-s:ID Sample, cosmic-s:Gene Name, cosmic-s:Chrom start) are fetched from COSMIC and the results for the next three triples (i.e. tcga:tcga id, tcga:start) are fetched from TCGA. To produce the required information, both results are merged on the basis of the last triple which integrates COSMIC with TCGA. Sample results for this query can be seen in Figure 5.
Listing 1.2: Federated SPARQL Query PREFIX cosmic s : <http : / / cosmic . s e l s . i n s i g h t . org / schema/> PREFIX t c g a : <http : / / t c g a . d e r i . i e /> PREFIX owl : <http : / /www. w3 . org /2002/07/ owl#> SELECT WHERE f ? c o s m i c r e s u l t a cosmic s : r e s u l t ; cosmic s : ID Sample ? i d s a m p l e ; cosmic s : Gene Name ? gene ; cosmic s : c h r o m s t a r t ? c h r o m s c o s m i c . ? t c g a r e s u l t a t c g a : r e s u l t ;
t c g a : Id ? t c g a i d ; t c g a : S t a r t ? t c g a c h r o m s t a r t . ? i d s a m p l e owl : sameAs ? t c g a i d . g 4
We analysed all genes have an RPKM > 0.5 [
Initially, we have sampled 99 genes that are highly expressed in all 16 tissues as shown in Figure 3 to retrieve their CNV, mutation and gene expression annotations from cBIO portal (for TCGA) and CNV annotator (for COSMIC) to determine current state of the art and provide a baseline comparison for the proposed linked annotation solution. The results for TCGA 4 clearly indicate an elevated distribution of these genes in uterine and ovarian cancer with a large number of mutations and CNVs. As an outcome, ovarian cancer has been selected as a good candidate for further investigation due to its elevated ampli cation rate and its multiple repetition in di erent experiments. Further studies have been retrieved that were conducted to understand the somatic relevance and the loci(genomic position) of these genes, and further detailed information for mutations could be retrieved as well. This study demonstrated a focus to genes such as ACTC1, B2M, CRP, FABP3, FABP4, FGA, FGB, GC, MYH7, RPRH2, SLC26A3, TG, TXNIP which form most relevant driver genes transforming healthy human tissues into ovarian cancer. The same 99 genes have been queried against 99 genes from HBM 2.0 to get the results for the somatic mutations in cancer. This repeat annotation will not only provide detailed statistics reported in COSMIC but also a validation for out earlier experiments. Table 2 clearly indicates that locations chr1,chr4,chr14 and genes CRP,FGA,FABP3,MYH7 could be potential genes with high relevance for the development of ovarian cancer.
80% cy 70% eun 60% freq itona 50% r e ltA 40% 30% 20% 10% 0% Cancertype Mutationdata + + + + + + + + + + + + + + + + - + + + + - - + + + + + + + + + + + + + + + + + + + +
CNVdata + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + UterineCS(OTCvaGrAia)Mne(TlaCnGomA)a(TBClaGddABe)lard(TdCerG(ATLC)unGgALaupdnuegbn)aod(eTnCoG(ATC)LGuAngpusbq)u(TCLGivAOe)vra(rTiaCnGS(ATtoC)mGaAchpu(TbC)GSAtompLuaubcn)hg(sTqCuG(HATeC)aGdHA&epanudebc&)kn(eTcCkG(ATC)GABpreuabs)tS(TaCrcGomAU)ate(rTinCeG(ATC)GAUtpeuribn)eP(aTnCcGreAa)s(CTeCrGviAcEa)slo(TpChaGgAu)s(TCDGLBABC)re(TaCstG(ATC)GPAropsutabt)eC(oTloCrGPecrAota)slta(TteCC(GToAClo)GreAct2a0l1(T5C)GAccpRuCbC)(TGCliGomA)a(TCGGBAM)(TCGACAC)(TCpRGCAC)GB(TMC(GTAC)GAc2h0R1C3cC)cR(TCCCG(ATC)GAPpCuPbG)(TcChRGCATC)hy(rToCidG(ATC)GTAhpyurobi)dA(TMCLG(ATC)GApAuMb)GLB(TMC(GTAC)GA2008)
Mutation
Deletion
Amplification Multiple alterations
Fig. 4: TCGA query output from cBIO Portal[
We now query the same 99 genes from before using the linked annotation mechanism. The result snippet for ovarian cancer is depicted in Figure 5 with detailed functional annotations together with the TCGA ids, again for ovarian cancer.
cs:result cs:Sample_Name c:TCGA-13-0920-01; cs:Gene_Name c:MYH7; ts:result ts:bcr_patient_barcode t:TCGA-13-0920; cs:Regulation "over". ts:beta_value "0.0419.."; cs:chr_no c:1; ts:chromosome t:14; cs:chrom_start_m c:23418303; ts:chromosome t:1; cs:chrom_stop_m c:23418303; ts:start t:1288070; cs:chr_no_m c:14; ts:stop t:1293914; cs:mut_type "GAIN"; ts:scaled_estimate "773.555". cs:Primary_Site c:ovary; cs:Primary_Histology c:carcinoma;
Fig. 5: Linked annotations for MYH7 - COSMIC
Figure 5 represents the COSMIC and TCGA annotations, respectively.
MYH7 corresponds to chr-1 which is evident from previous annotations and
replicated again in our study along with its TCGA ID:TCGA-13-0920-01. Its
mutation type is primarily the GAIN type of a mutation for chr1 and chr14 which
is a dominant mutation with all its regulation of over, under and normally
expressed. Translational researchers may want to repeat and re-validate the study
for Pubmed ID:1398522 additionally with beta value (measure of methylation)
of 0.041999536 and scaled estimation (Tumour purity) of 773.555 also supports
this gene from the epigenetic point of view. Further multiple genomic locations
will help clinical practitioners to track CNV for the targeted study and which
ultimately leads the direction towards a better prognosis.
Kandoth et al. [
Likewise a reduced version of the COSMIC database has been RDFized to
explore on the mechanism of TP53 [
In this paper we have presented a linked data infrastructure for functional annotation which enables querying di erent types of mutations and genomic alterations to contribute to molecular and clinical insights of cancer by de ning most relevant variants and their prioritization. This knowledge could be highly advantageous for a targeted therapy and personalized medicine based on gene expression data. The presented experiments are based on TCGA, COSMIC and HBM 2.0 datasets and have been used to identify sets of genes with relevance for ovarian cancer and with comprehensive set of mutations. Similar studies have to be performed for other cancer types. We have covered CNV, gene expression and mutation data from COSMIC and TCGA (only for ovarian cancer). We have processed 1.2 billion COSMIC triples and 100 million TCGA triples which in turn generated 27 GB of data. In future, this work will be expanded to cover level 1, 2 and 3 along with other datasets from COSMIC to provide in-depth biological insight for each queried gene.
This publication has emanated from research supported by the research grant from Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289.