Conceptual Modeling Applied to Genomics: Challenges Faced in Data Loading

Conceptual Modeling Applied to Genomics: Challenges Faced in Data Loading Matthijs van der Kroon september 2011 2 Contents 1 Introduction 7 1.1 Domain description .......................... 11 1.1.1 Genetics 101 ......................... 11 1.1.2 Genome structure ...................... 13 1.1.3 Genotype to phenotype ................... 15 1.1.4 Genetic variation ....................... 19 1.2 Problem statement .......................... 21 1.2.1 Data chaos .......................... 22 1.2.2 Conceptual chaos ....................... 24 1.3 Contribution ............................. 25 1.4 Methods and materials ........................ 26 2 State of the art 29 2.1 Conceptual modeling in Information Systems ........... 29 2.2 Conceptual modeling in bioinformatics ............... 31 2.3 Ontology based solutions ....................... 33 2.3.1 Universals versus particulars ................ 34 2.3.2 Continuants versus occurrents ................ 35 2.3.3 GO’s relations ........................ 35 2.3.4 Conceptual modeling ..................... 36 3 The Conceptual Schema of the Human Genome 39 3.1 The structural view ......................... 39 3.2 The variation view .......................... 41 3.3 The phenotype view ......................... 43 3.4 The transcription view ........................ 46 3.5 The genome view ........................... 49 3.6 The bibliography/databank view .................. 50 4 Solving the HGMD case 51 4.1 HGMD description .......................... 52 4.2 Encountered problems ........................ 52 4.2.1 Intrinsic data properties ................... 53 4.2.2 Data representation ..................... 54 3 4CONTENTS 4.3 Data loading problem solutions ................... 56 4.3.1 Intrinsic data properties ................... 57 4.3.2 Data representation ..................... 58 4.4 Lessons learned ............................ 61 5 Solving the dbSNP case 65 5.1 SNP definition ............................ 66 5.1.1 SNP characteristics ...................... 66 5.1.2 Uses of the SNP concept ................... 68 5.2 SNP incorporation in the CSHG .................. 69 5.3 dbSNP description .......................... 72 5.4 Encountered problems ........................ 72 5.4.1 Data quantity problems ................... 73 5.4.2 Problems related to semantics ................ 75 5.5 Lessons learned ............................ 78 6 Solving the BIC case 81 6.1 BIC description ............................ 82 6.2 Encountered problems ........................ 82 6.2.1 Aligning reference sequences ................. 84 6.3 Lessons learned ............................ 85 7 Genoma Data Loader 87 7.1 Structure ............................... 88 7.2 Coverage ................................ 91 8 Conclusions 93 9 Future work 97 Acknowledgements 101 Appendices 111 A Data format samples 113 A.1 HGMD precise mutations format sample ..............113 A.2 HGMD imprecise mutations format sample ............114 A.3 dbSNP XML format sample .....................114 A.4 BIC data format sample .......................116 B Genoma Data Loader coverage 117 B.1 Variations per gene, per data repository ..............117 B.2 Translated variations per gene and data repository ........118 B.3 Translated unique variations per gene and data repository . . . . 119 B.4 Translated unique variations versus total number of variations per gene and data repository ....................120 CONTENTS 5 C Algorithms 121 C.1 Calculating absolute position ....................121 C.2 Separating reference from variation .................121 C.3 Normalizing variations ........................123 6CONTENTS Chapter 1 Introduction Todays genomic domain evolves around insecurity: too many imprecise concepts, too much information to be properly managed. Considering that conceptualization is the most exclusive human characteristic, it makes full sense to try to conceptualize the principles that guide the essence of why humans are as we are. This question can of course be generalized to any species, but in this work we are especially interested in showing how conceptual modeling is strictly required to understand the ”execution model” that human beings ”implement”. The main issue is to defend the idea that only by having an in-depth knowledge of the Conceptual Model that is associated to the Human Genome, can this Human Genome properly be understood. This kind of Model-Driven perspective of the Human Genome opens challenging possibilities, by looking at the individuals as implementation of that Conceptual Model, where different values associated to different modeling primitives will explain the diversity among individuals and the potential, unexpected variations together with their unwanted effects in terms of illnesses and susceptibilities to certain types of drugs. Genomics, from an Information Systems point of view, is largely situated in the first phase of systems design, the analysis. Due to the youth of the genomics domain, many aspects of what is driving the mechanisms of life are still unknown. This work presents an inter-disciplinary approach, in which experience in Information Systems development is put to practice in genomics. Concretely speaking, by applying a conceptual modeling approach, fixing the present day knowledge about the human genome in a visual form. The genomic domain can be characterized by three distinct properties: large data quantities, high complexity and rapid evolution. The first property poses certain challenges on resources, like processing time and storage space. Processing the large linkage disequilibrium data files provided by the HapMap [1] resource (around ∼40 Gb compressed, ∼224 Gb uncompressed) for instance certainly counts as challenging, but is by no means impossible due to the regular structure. It is the high complexity, rapid evolution and the ambiguity that often results from those aspects that pose the real problems on the long run as they call for a stable and homogeneous structure, while at the same time allowing for efficient and easy 7 8CHAPTER 1. INTRODUCTION evolution of this same structure. No amount of processing power can compensate for a lack of structure, if what one is looking for simply has not been stored. It is true that genomics is often not an exact science due to the immense complexity of nature and its processes. It is true that basic concepts like genes, alleles and mutations are frequently variable in their definition. Their exact denotation often depends on both context and position in time. A gene for instance can be defined as a locatable region of genomic sequence, corresponding to a unit of inheritance, which is associated with regulatory regions, transcribed regions, and or other functional sequence regions. However, the existence of splicing, in which gene expression changes at runtime, complicates this view as is very well described by [2]: ”in eukaryotes, the gene is, in most cases, not yet present at DNA-level. Rather, it is assembled by RNA processing”. Pearson [3] and Gerstein et al. [2] provide recent insights on the evolution of the gene concept. But it is the central point of this work that for efficient research to take place, clear definitions of concepts and a common vocabulary are crucial. This is especially the case in research where worldwide various separate research groups are collaborating and exchanging information on a regular basis. Formally describing concepts, ruling out ambiguity and relating concepts to each other and their context is the main objective of model driven software development. Simply put, a conceptual model is a simplified representation of reality, devised for a certain purpose and seen from a certain point of view. The objective of a conceptual model is simulation of reality, it therefore needs to react to input in the same way as reality would. It is in this context, where a precise connection between the genomics domain and the Conceptual Modeling approach makes full sense. Describing a system by means of conceptual models means viewing the world as consisting of objects that belong to different classes, have distinct properties, and are related to each other in various ways. This way of viewing a system provides a powerful representation and reasoning tool. Modeling a domain in term of concepts thus roughly has either one of two applications, or both. Either the models serve by use as a reasoning tool to gain a deeper understanding of the domain at hand, usually as part of the design of an Information System, or they guide the creation of a system which is ultimately directed at controlling or modifying that same domain. In the first application, the conceptual model serves as a visual representation of the domain, linking concepts and their respective behaviors while the latter resembles the blue print used in traditional building. Not surprisingly, conceptual models and ontologies are closely intertwined. There is still a large debate about the use of ontologies in the bioinformatics domain. Section 2.3 will dig deeper into this subject, further refining the definitions of both conceptual models and ontologies and how they interconnect in the context of this work. In general, current efforts doing exactly what this work pretends to do, capturing the semantics of the Human genome in a formal manner, are almost all ontology based. The Gene Ontology [4] is considered 9 to be the de-facto standard here. It is however the contribution of this work to show how conceptual modeling techniques can improve on these efforts by providing a solid, well researched base which -due largely to its visual and comprehensive characterenables the engineer to capture these earlier mentioned semantics in close collaboration with domain experts. All this without losing the clear advantages and many promises that ontologies, as conceived by the bioinformatics domain, present while at the same time adding a few more. By allowing for a minor paradigm-shift, the human genome (or any genome) can be considered an Information System, a natural and highly complex form maybe but an Information System nonetheless. Stated in a very simplified manner, data that is stored in DNA, undergoes recombination, processing and ultimately translation to proteins. It is the combination of these proteins and the influence from external factors (the environment) that define the way an individual looks and behaves. These characteristics map neatly to the generic characteristics usually associated to Information Systems, only replacing manmade hardware in the form of chips and circuits, by biological molecules and physics. As Chikofsky and Cross [5] state, reverse engineering is defined as the process of analyzing a subject system to (i) identify the systems components and their inter-relationships and (ii) create representations of the system in another form or at a higher level of abstraction. Following this philosophy, just like software can be reverse-engineered in order to apply after-market changes or facilitate maintenance [6], it might very well be possible to reverse engineer life itself and create a higher level of abstraction representation in the form of a conceptual model. In this case, after-market changes and maintenance include treatments and/or prevention of previously untreatable disorders and disease. Basically, said in Information Systems jargon, debug life itself. The value of a conceptual model of the human genome is thus two-fold: first, it allows for a visual, and formal representation of the domain. By doing so, fixing a vocabulary and conceptual gamut from which scientists can draw in order to ensure communication takes place based on the same dictionary, using the same concepts. It deserves mentioning here, that the conceptual model as proposed in this work is thus expected to evolve along with the advancing understanding of the domain in time. This evolution capacity is an extra value in itself for a domain where knowledge is continuously being generated, and thus continuously subject to change. Only by having a well-defined, precise conceptual background the Conceptual Modelcan this new knowledge be properly incorporated, be understood and be adequately managed. Second, the conceptual model can be used to (semi-)automatically generate software. When seen as a Platform Independent Model (PIM) the Conceptual Schema of the Human Genome can be considered as the base from which various Platform Specific Models (PSM) can be deduced. Each of these PSMs can then be transformed into a functional implementation. The degree of functionality, and interference of the engineer depends on the level of expressiveness of the model. Current efforts, like OO-method [7], suggest that it is possible to generate fully functional Information Systems from nothing but models, in which low-level programming becomes redundant. It is well known that programming 16 CHAPTER 1. INTRODUCTION During transcription, a DNA sequence is read by RNA polymerase, which results in a complementary RNA copy in which all instances of thymine have been replaced with uracil. Various types of RNA exist, the most aberrant one being messenger RNA, or mRNA, in which the enclosed message potentially encodes for a protein. Other examples include ribosomal RNA, transfer RNA and micro RNA. Splicing is the process of modifying an RNA after the initial transcription process in which introns are removed and exons joined. Splicing rarely involves RNA that is not of the mRNA type, thus in the context of this work only this case will be considered. As explained earlier, in section 1.1.2, genes can be decomposed in coding sections, exons, and non-coding sections, introns. In order for the messenger RNA to produce a correct protein through translation, these sections need to be either eliminated or joined. Depending on various factors like in which tissue the particular cell is situated and cell age, the splicing process might also exclude specific exons, resulting in different mRNAs. The process of splicing allows for a ’run-time’ reordering of genetic information that enables for one gene to produce a variety of products. The earlier mentioned organism complexity issues can largely be explained by exactly this; organisms considered to be complex, like eukaryotes, splice many protein-coding mRNAs and some non-coding RNAs. Prokaryotes, on the other hand, splice rarely and mostly non-coding RNAs. Figure 1.3: The translation of mRNA and the synthesis of proteins by a ribosome. 1.1. DOMAIN DESCRIPTION 17 Translation Once the genetic message has been transcribed from DNA onto an RNA molecule, it is able to leave the cell nucleus and enter the cell’s cytoplasm. It is here that the ribosomes are located. These ribosomes act like ’protein-factories’ in which mRNA is translated, or synthesized, to proteins. The mRNA sequence binds to the ribosome and is interpreted in codons, which means nucleotide triplets. As the former holds true, the way of breaking the sequence in three letter codons -the reading framedefines how the sequence is read. There are three possible reading frames in an mRNA strand: each reading frame corresponding to starting at a different alignment. The reading frame that has a start codon, and a subsequent region usually containing a multiple of 3 nucleotides, but excluding the stop codons is called the open reading frame (ORF). Each possible combination of the mRNA bases -uracil (U), adenine (A), cytosine (C) and guanine (G)- represents a particular semantic value, as shown in figure 1.4, either relating the triplet to a specific amino acid or providing syntactical clues. The latter refers to start (AUG) and stop codons (UAA, UAG and UGA). Figure 1.4: Semantics of the triplet codes for each amino acid. Once the ribosome encounters a start codon, the translation is initiated and for each following codon the corresponding amino acid is connected to a growing chain of amino acids. This chain is what ultimately will lead to the finished protein product. Once the ribosome encounters one of the stop codons it initiates a process that induces the binding of a release factor protein which will release the created protein from the mRNA-ribosome assembly. As can be seen from the table, there exist more distinct codons (27) then amino acids (20). This allows for natural redundancy in which evolutionary speaking important 18 CHAPTER 1. INTRODUCTION amino acids appear multiple times. Metabolic pathways The resulting products from transcription and then translation then interact in a series of chemical reactions known a metabolic pathways. In each pathway, a principal chemical is modified by these series of chemical reactions. These reactions are catalyzed by specific proteins, also known as enzymes, and often require additional chemicals like minerals, vitamins and other cofactors. The involved compounds are referred to as metabolites. Pathways may include other pathways, and thus create an intricate network known as the metabolic network, see figure 1.5 for an overview of how the major metabolic pathways interact with each other. Figure 1.5: The major metabolic pathways. Metabolic pathways describe the process in which an initial molecule, the substrate, is transformed through a series of chemical reactions to form another product. Pathways that break down compounds are said to be catabolic, while anabolic pathways often create new biomolecules as final end-products. Although technically speaking all chemical reactions are reversible, the conditions in the cell are often such that it is thermodynamically more favorable to flow in only one direction of a reaction. 1.1. DOMAIN DESCRIPTION 19 1.1.4 Genetic variation It is clear that the biological mechanism of life is very complicated, and knows many failure points. A common cause for both evolutionary advantages and disadvantages is genetic variation. Genetic variations occurs through mutation, a change in the nucleotide sequence of a gene. A variation will result in phenotypic variation if the change in DNA sequence affects the order of amino acids resulting from translation, and the resulting difference in amino acid sequence influences the shape, and thus function of the protein. This phenotypic change can then have either one of three distinct effects on the capability of the individual to reproduce; (i) positive, when for instance the change makes the individual resistant to certain disease, (ii) negative, when the change increases susceptibility of the individual to disease and (iii) neutral, in which case the individual perceives no altered effect. It deserves mentioning here that geographic variations largely influences the effect of genetic variation on the individual. One can imagine the case of a genetic variation that impedes the proper expression of the EPAS1 gene, resulting in the translation of a less functional Endothelial PAS domain-containing protein 1 transcription factor. The function of this protein is to switch on genes that produce proteins associated to oxygen binding. Someone living at high altitude in the Himalayas will perceive this particular genetic variation as having a negative effect, clearly posing an evolutionary disadvantage over ’healthy’ individuals. But someone living at sea level, a situation in which this particular mutated EPAS1 gene is probably never expressed, will classify the variation as being neutral. Common ways to classify variations include (i) by effect on structure, (ii) by effect on function, (iii) by effect on fitness, (iv) by impact on protein sequence and (v) by inheritance ability. For the scope of this work we will confine the description to the first three classifications, namely (i) by effect on structure, (ii) by effect on function and (iii) by effect on fitness. Each of these will now be discussed in the following sections. By effect on structure When classifying genetic variation in terms of the effect on structure one emphasizes the alterations that occur in the genetic sequence of a given gene. Depending on where these variations occur, and whether they alter the function of the protein product will define the effects on health. •Small-scale mutations Point mutations, these variations substitute one nucleotide for another, thereby slightly changing the sequence of the gene, and therefore the resulting gene product. Common subclassifications for this type of variations include (i) silent mutations, (ii) missense mutations and (iii) nonsense mutations. Insertions, variations that insert one or more nucleotides at a given position in the genetic sequence are considered insertions. Insertions, like 20 CHAPTER 1. INTRODUCTION deletions, might trigger a frame-shift which is very likely to significantly alter the gene product. Deletions, remove one or more nucleotides from the genetic sequence. Very much like insertions these variations might affect the reading frame in such a way that the gene product is significantly altered. •Large-scale mutations Amplifications, duplicates certain areas of the chromosomal regions, altering the gene products significantly. Deletions, large-scale deletions differ from small-scale deletions such that in the former a small amount of nucleotides is involved while in the latter case multiple genes might simply disappear. Complex reorganizations, reorders pieces of DNA in such a way that gene products significantly change. Common subclassifications for this type of variations include (i) chromosomal translocations, (ii) interstitial deletions and (iii) chromosomal inversions. Loss of heterozygosity, in which case an entire allele is lost By effect on function Another way of looking at genetic variation, is classifying it according to how function is affected by the change. We distinguish five different types of variations in this particular case. •Loss-of-function variations, cause the gene product to have less or no function. In case of complete loss of function this type of variation is often referred to as an amorphic mutation. •Gain-of-function variations, modify the gene product in such a manner that the product acquires a functionality previously not present. •Dominant negative variations, the result of this type of variations is that the gene product acts in an exact opposite manner to the ’healthy’, or wild-type, variant. •Lethal variations, cause the death of the individuals that carry the variation •Back variation or reversion, this type of point variation cancels out variations restoring the original sequence and thereby the original phenotype. By effect on fitness Viewing genetic variations in terms of how they affect the fitness of the individual to cope with environmental factors and reproduce determines this classification. Seemingly, a variation has either one of two effects: harmful, or beneficial. In which case the former means a decrease in fitness of the individual, while the 1.2. PROBLEM STATEMENT 21 latter refers to an increase of fitness of the individual. In reality these discrete categories turn out to be an oversimplification. The earlier mentioned case of EPAS1 already illustrates this. Simply put, determining whether a variation has a harmful or beneficial effect is determined to large extend by the environment in which the particular individual resides. 1.2 Problem statement The sequencing of the Human Genome in the year 2000 by Craig Venter and Francis Collins [9] came with tremendous promises. These effects are most probably not yet apparent and science still struggles to process the huge accomplishment into usable artifacts. Scientists now broadly agree that reading the sequence of DNA is the easy part of genome analysis; figuring out what the sequence actually means is the real challenge. So following these insights a new scientific line of research was opened as the marriage of Information Technology (IT) and biology: bioinformatics. It is here that bioinformaticians try to combine rigorous yet computationally powerful IT with the ambiguous and fuzzy biology. And as Venter and Collins efforts started to bear fruits, more and more sequencing experiments have been performed world-wide generating large amounts of data. These data are stored in countless locations, according to countless formats and structures, leading to the following problem statement: how do we manage the data chaos? As technology advances the amount of data generated by sequencing experiments increases at an equally rapid pace. The task of biologists has traditionally been to analyze and interpret these data, effectively turning them into useful knowledge that eventually can be used in various applications. However, as data creation starts to greatly surpass the processing capability of a single human being, this analysis and interpretation activity becomes a daunting task. A possible solution to this data-overload would be the use of powerful computers, effectively pre-processing the data and separating the junk from useful data thereby allowing the expert to efficiently dedicate time to data worth looking at. For this to happen, the data needs both structuring into formats that allow for computational reasoning while at the same time computer programs that perform this act of reasoning need to be created. We thus identify the problem in current genomics as being two fold: on the one hand we have large amount of data. These data are stored globally in a very fragmented manner, by countless institutes each of whom often uses its own, ad-hoc, standards [14]. The data are considered to be highly complex and often inconsistent with each other due to the rapidly progressing comprehension of the domain. We have come to refer to this particular phenomenon as the ’data chaos’. On the other hand, apart from this obvious issue there is also a more hidden case, where conceptually speaking the domain is very much ambiguous and fuzzy. This ’conceptual chaos’ poses many barriers on promising reasoning technologies, that would otherwise enable exciting possibilities. These two pillars that define the main motivation behind this work will be discussed 22 CHAPTER 1. INTRODUCTION separately in the following sections. 1.2.1 Data chaos Genomic data is stored globally in a very fragmented manner. The lack of a consensus about these data, how to store and validate them, often results in inconsistencies and sometimes invalid information, as has been identified and confirmed by [15] and [16]. As an added problem, the fragmentation requires a high level of expertise for these data to be interpreted, which is the main task of a biologist. All in all, the core tasks of biologists become tedious, error prone and costly in terms of both time and money due to this lack of unity. In the context of this work we will now introduce some of the common repositories of genomic data, please notice that we follow a similar structure as above in section 1.1, starting with genomic structure subdomain, moving through the genotype to phenotype process and pathways towards genetic variation. The National Center for Biotechnology Information1(NCBI), established in 1988 as a national resource for molecular biology information, NCBI creates public databases, conducts research in computational biology, develops software tools for analyzing genome data. Since then they have continued to strike a compromise between the convenience and simplicity required for the everyday use of human gene nomenclature and the need for adequate definition of the concepts involved. In short, NCBI is a U.S. government-funded national resource for molecular biology information, providing access to a variety of data including genetic variations [17] and genetic sequences [18]. See [19] for a detailed overview of all services and tools provided by the NCBI. A data repository similar in mission to NCBI, is Ensembl2[20]. Ensembl is a joint project between EMBL-EBI and the Sanger Centre to develop a software system which produces and maintains automatic annotation on genomes. The goal of Ensembl was initially to automatically annotate the genome, integrate this annotation with other available biological data and make all this publicly available via the web. Since the website’s launch in July 2000, many more genomes have been added to Ensembl and the range of available data has also expanded to include comparative genomics, variation and regulatory data. The Online Mendelian Inheritance in Man3(OMIM) [21] [22] project, also a project under the NCBI umbrella, consist of a semi-structured collection of diseases and genetic mutations linked to these diseases. The full-text, referenced overviews in OMIM contain information on all known mendelian disorders and over 12,000 genes. OMIM focuses on the relationship between phenotype and genotype. It is updated daily, and the entries contain copious links to other genetics resources. It is available as a single file download but due to it’s low degree of structure very difficult to process automatically. It is for this latter reason that OMIM is not considered in this work. 1http://www.ncbi.nlm.nih.gov/ 2http://www.ensembl.org/ 3http://www.ncbi.nlm.nih.gov/omim 1.2. PROBLEM STATEMENT 23 Bioinformatics resources that provide information on pathway data include the REACTOME4[23] [24] and KEGG5[25] [26] [27] databases. REACTOME is an open-source, open access, manually curated and peer-reviewed pathway database. Pathway annotations are authored by expert biologists, in collaboration with Reactome editorial staff and cross-referenced to many bioinformatics databases. These include NCBI Entrez Gene, Ensembl and UniProt databases, the UCSC and HapMap Genome Browsers, the KEGG Compound, PubMed, and Gene Ontology. KEGG (Kyoto Encyclopedia of Genes and Genomes) is a collection of online databases dealing with genomes, enzymatic pathways, and biological chemicals. The PATHWAY6database records networks of molecular interactions in the cells, and variants of them specific to particular organisms. Apart from datasources that strive to provide structural information like genetic sequences of genes, chromosomes and entire genomes, there are also data repositories with the sole purpose of storing genetic variational data. The Human Gene Mutation Database7(HGMD) [28] provides a per-gene repository of mutations, or base changes that research has uncovered to be associated to disease. The full data set of HGMD can be acquired for a fee, while a less complete version is available for free in case of academic use. Given that it is curated by experts, reading scientific papers and extracting the mutational data from them, the source is considered to be highly reliable. At present day the HGMD is considered to be the primary source for obtaining genetic mutations among various genes, although alternatives exist. Many of these alternatives are non-profit based, meaning they provide open-access without limitations. Examples include the Catalogue of Somatic Mutations in Cancer8(COSMIC) [29], the HapMap9[1], the Breast Cancer Information Core (BIC) [30], dbSNP10 [17] and many Locus Specific Databases (LSDBs). An LSDB aims to store only data for a specific locus, or region of the genome. Often these type of data sources store genetic variations for one gene, or a limited amount of genes. An interesting effort to standardize these LSDBs, is the Leiden Open Variation Project (LOVD) [31].Table 1.1 shows an overview of available data sources that contain data on the BRCA1 gene. It is generally assumed that the data overlaps at least to some extend, but the exact amount is unclear. Adding to this already large list of data sources come the repositories that provide data on proteins. The mission of UniProt11 [32] is to provide the scientific community with a comprehensive, high-quality and freely accessible resource of protein sequence and functional information. InterPro12 [33] is an integrated database of predictive protein ”signatures” used for the classification and automatic annotation of proteins and genomes. InterPro classifies sequences 4http://www.reactome.org/ 5http://www.genome.jp/kegg/ 6http://www.genome.jp/kegg/pathway.html 7http://www.hgmd.cf.ac.uk/ 8http://www.sanger.ac.uk/genetics/CGP/cosmic/ 9http://hapmap.ncbi.nlm.nih.gov/ 10http://www.ncbi.nlm.nih.gov/projects/SNP/ 11http://www.uniprot.org/ 12http://www.ebi.ac.uk/interpro/ 24 CHAPTER 1. INTRODUCTION Table 1.1: Number of genes, and amount of mutations for the BRCA1 gene per genetic mutation source. Source Genes BRCA1 HGMD (academic) 3132 1085 HGMD (commercial) 4122 1339 COSMIC 18647 15 LOVD BRCA1/2 2 502 BIC 2 1446 at superfamily, family and subfamily levels, predicting the occurrence of functional domains, repeats and important sites. InterPro adds in-depth annotation, including GO terms, to the protein signatures. Clearly a large amount of databases exist, and every subdomain has at least two different repositories in use. The above only shows a fraction of what databases actually exists in the domain, which is estimated to be more then 1300. Some of those provide comprehensive ways to access their data, like Ensembl does with Biomart [34], while others provide no others means then either a manual copy-paste process or screen scraping based methods, like the HGMD. In general the data can be accessed through public FTP servers. The data are then provided in Tab Separated Files (TSV), Comma Separated Files (CSV) or XML. In other cases yet, the SQL-based database in use by the repository can be downloaded as a whole and installed as a local copy. It is clear that for each subdomain various data sources exist, apart from this each has its own manner of making the data accessible to third parties. 1.2.2 Conceptual chaos The conceptual chaos, although largely related to, is fundamentally different from the above stated data chaos. In short, having a fragmented data storage is one thing, not knowing the precise semantics of those data is of a different order. In rapid evolving domains like genomics it comes as no surprise that the state of knowledge keeps progressing, maybe even on a daily basis. Although in principle a benevolent process, this also means data have a risk of becoming outdated; essentially rendering them invalid. Apart from this fact, it is difficult to appropriately store data in such a manner that it allows for applications later on that were not envisaged at the time of storage. In biology, basic concepts like a gene, or allele still have not been defined to a level of formality usually seen -and often necessary for a proper applicationin Information Systems [2]. Fixing exactly this issue has been topic of research for over a decade, and the Gene Ontology13 [4] has commonly been recognized to be the de-facto standard as a result. Conceptual modeling based solutions include [35] [36], some first efforts, and [37], which builds upon the former. A detailed state of the art will be discussed in section 2.2. 13http://www.geneontology.org/ 1.3. CONTRIBUTION 25 Bringing order in the conceptual chaos has the potential benefit of allowing a computationally based data selection process through which relevant data can be sifted out and presented to a domain expert, often a biologist, for further interpretation. Clearly, the amount of data being generated currently -which in the future is only expected to increase [38]- is impossible to process by a single human being. It is therefore becoming increasingly difficult for biologists, whose core task is interpreting the various data, to assemble a ’complete’ picture and act on that. A possible solution exists in the use of computers to alleviate the load, and pre-select potentially relevant data from the data chaos. But in order to do this, a proper semantic understanding of the domain is required, exactly something that the conceptual schema of the human genome aims for. 1.3 Contribution First of all it is important to mention this work has been the result of a multidisciplinary effort, combining conceptual modeling techniques to the domain of bioinformatics, and genomics specifically. It can be considered a case study for the application of Information Systems methodologies in new areas, thereby defining a clear contribution to the IS community. At the same time, the contribution to bioinformatics is clear in that our approach shows clear evidence of improving current problems. Fully resolving the issues of chaos stated above represent an immense task, far outside the scope of this work. We do however aim to present a step in the direction we believe is right towards a bioinformatics utopia where both conceptually speaking, and in terms of data fragmentation the chaos is no long existing. Conceptual modeling as a methodology has been long recognized as a proven way to improve Information System quality, as will be discussed shortly in section 2.1. Bioinformatics in general is producing large amounts of complex data, that need to be interpreted by human experts. It is clear that to enable a proper management of this task, Information Systems can prove very useful. Seen from this perspective it makes perfect sense to use a well-defined conceptual schema as a basis. The Conceptual Schema of the Human Genome presented in this work can be used as a type of blueprint schema from which generic Genomic Information Systems (GIS) can be created, both manually and automatically using well-proven MDE methodologies. It is expected that this effort will allow for easier creation of more consistent and correct GIS, as has been shown to happen in other areas when applying MDE [39]. Another interesting of the CSHG introduces itself as the implicit specification of the underlying ontology. By viewing a domain as a set of concepts with properties, interconnected through various types of relations the engineer, along with the domain expert, creates a specification of a conceptualization. The added advantage of using a visual representation -as is used by the conceptual schemais that it is intuitive to both engineer and domain expert. This latter property enables the engineer to work much closer to the problem domain, effectively having a continuous and direct feedback loop 32 CHAPTER 2. STATE OF THE ART [57], [58], knowledge representation [59], [60], [61], qualitative modeling [62], [63], [64], language engineering [65], [66], database design [67], [68]. In these contexts, ontologies are usually understood as dictionaries, taxonomies, categorization schemata or modeling languages. Schulze-Kremer [69] confirms the need for a solution to the conceptual chaos by stating: ”to eliminate semantic confusion... or to provide an exact, semantic specification of the concepts”. He further proceeds to describe two , well-known at the time, biological ontologies: Cyc [70] and microKosmos [71]. The Gene Ontology (GO) as presented by Ashburner et al. [4] in 2000 is nowadays considered the de-facto standard as the biology ontology and can be considered an evolution of the Ontology for Molecular Biology as initially presented by Schulze-Kremer [69]. Schwarzer et al. [72] further elaborate on GO by describing the application of the Gene Ontology to the SNP concept. Coulet et al. [73] proceeds to provide an ontology-based converter that allows for solving the notational problems associated to heterogeneous SNP descriptions. Bard and Rhee [74] offer a review of ontologies in biology. Some examples of relevant ontologies include the Cell Ontology (OBO)1,Galen2, a management architecture for clinical information that includes an ontology for human anatomy and MetaCyc3. It is interesting to note that in the 32 ”principal biological ontologies and other web sites” three entries can be directly linked to GO, and probably more when delving deeper into the separate entries. This confirms our earlier statement that the GO acts as a de-facto standard. Bard and Rhee proceed to clarify the use of ontologies in general, and applications for them in biology in particular as well as providing a future vision on the topic. Their main conclusion is that the key to the general use of ontologies will be access to the data in biological databases -the genetic variation data repositories discussed later in this work fall into this categorythat are annotated with the knowledge in these ontologies. An Information Systems approach to this specific biological problem space is not entirely new. Okayama [75] describes the conceptual schema of a DNA database using an extended entity-relationship model. Chen et al. [76] has indicated how an extended object data model can be used to capture the properties of scientific experiments, and [77] includes models for representing genomic sequence data. Paton et al. [35] advanced on this work by presenting a first effort in conceptually modeling the S. cerevisiae genome, which is a type of yeast, by proposing a collection of conceptual data models for genomic data. Among these conceptual models are a basic schema diagram for genomic data, a proteinprotein interaction model, a model for transcriptome4data and a schema for modeling alleles. Whereas [53] provides a broader view by presenting conceptual models for describing both genome sequences and related functional data sets, [78] further elaborated on the basic schema diagram for genomic data thereby 1obo.sourceforge.net/list.shtml 2www.opengalen.org 3http://metacyc.org 4The transcriptome is the set of all RNA molecules, including mRNA, rRNA, tRNA, and other non-coding RNA produced in one or a population of cells 2.3. ONTOLOGY BASED SOLUTIONS 33 narrowing the focus and specializing it for the human genome. Whereas Paton et al. provide a broader view by presenting conceptual models for describing both genome sequences and related functional data sets, [52] converged on the basic schema diagram for genomic data adapting it to the human genome and eventually produced a database, the human genome database (HGDB) corresponding to this model and following the standard rules of logical design. This database is now in the prototype phase and the first 2 genes, NF1 and BRCA1, have been partially loaded. Pastor et al. describes the evolution HGDB went through during the process of conceptually mapping HGDB and HGMD to each other in [53]. In [78] Pastor et al. describe the evolution of the model more in general and provide a descriptive overview of how the model came to be, and from where it evolved to what it is now. Banning ambiguity in the genomics domain has been subject of many earlier attempts, including ontologies and formal descriptions in natural language. Looking from the computer scientist perspective, ambiguity is usually considered an undesirable and often avoidable feature. Indeed, in computer design the behavior of the system is always intended to be known. In biology, and especially genomics, this is simply not the case. Complexity derived from the randomness which created the conditions allowing life to emerge is today obscuring the processes driving this very same system. 2.3 Ontology based solutions The sequencing of the Human Genome in the year 2000 by Craig Venter and Francis Collins [9] came with tremendous promises. These effects are most probably not yet apparent and science still struggles to process the huge accomplishment into knowledge artifacts. Scientists now broadly agree that reading the sequence of DNA was the relatively easy part of genome analysis; figuring out what the sequence actually means is the real challenge. Following these insights a new scientific line of research was opened as the marriage of IT and biology: bioinformatics. It is here that bioinformaticians try to combine rigorous yet computationally powerful IT with the ambiguous and fuzzy biology. And as Venter and Collins efforts start to bear fruits and technology rapidly advances, more and more sequencing experiments are being performed worldwide generating large amounts of data; leading to the following question: how do we manage the data chaos? Current solutions are often based on ontologies, most notably the Gene Ontology (GO). Literally translated from ancient Greek, ”oητoσ” means ”of that which is” and ”-λoγια” science, study. The science of Ontology (uppercase ”O”) is diverse and dates back to the early Greeks, where it referred to the analytic philosophy of determining what categories of being are fundamental, and in what sense items in those categories can be said to ”be”. In modern times, an ontology (lowercase ”o”) is considered many things. Gruber is credited for introducing the first compact, yet complete description: ”[an ontology is a] specification of a conceptualization” [55]. 34 CHAPTER 2. STATE OF THE ART Ontologies enable us to build large, maintainable knowledge bases that can codify what we know about specific areas of practice in precise, unambiguous terms. Adding to this, it allows us to reason over these structured knowledge bases, with the purpose of deducing new knowledge. In this short description we identify two different applications; knowledge management and knowledge deduction. An ontology can be of varying level of rigor, where a lower level, and as such more ambiguous ontology will be fine to deliver the promise of maintaining knowledge bases, but unable to allow for automated reasoning necessary to deduce new knowledge. A higher level of rigor will allow for both accurate knowledge management, and deduction of new knowledge at the cost of increased complexity. When the Gene Ontology was conceived in 2000 [4], it came with the promise of enabling a conceptual unification of biology by providing a dynamic, controlled vocabulary. Adding to this, it was hoped that the common vocabulary would result ”in the ability to query and retrieve gene and proteins based on their shared biology”, thus deducing new knowledge. Later research has shown that Gene Ontology in reality often lacks rigor to allow for this high level of ontology application. The early discussion started by [79] and [80], stating that ”It is unclear what kinds of reasoning are permissible on the basis of GOs hierarchies.” and ”No procedures are offered by which GO can be validated.’”. We now proceed to discuss the main points of their work. 2.3.1 Universals versus particulars In metaphysics, a universal is a meta-concept. It defines what particular things have in common, namely characteristics or qualities.The idea is that universals can be instantiated by particular things, or particulars (also called individuals, exemplars, instances, tokens). For example, the species E. coli, with the function to boost insulin production is a universal. Instantiated as the particular E. coli bacterium now existing in the Petri dish, its function is to boost insulin production in specific cells in your pancreas. Why is it important to have a distinction between particulars and universals? Consider the following case in which we have modeled a universal, say ”gene”, that corresponds to the biological concept by the same name and has a few properties: it is known for instance that a gene has a promotor and a terminator while being located on a specific chromosome. Once we now establish during a biological experiment that a certain stretch of DNA must be a gene (effectively instantiating the universal to an instance), we are now able to deduce from this knowledge that this particular stretch of DNA must have a promotor, terminator and that it is positioned on a chromosome. Imagine this same case but now we have not modeled the universal, instead storing a list of stretches of DNA (instances) that we consider to be genes. When we try to make the same deduction as earlier in case we find a new stretch of DNA that corresponds to the biological concept of ”gene”, we can not deduce what other properties it has. 2.3. ONTOLOGY BASED SOLUTIONS 35 2.3.2 Continuants versus occurrents Entities that continue to exist through time are often referred to as continuants. In the GO context, organisms, cells and chromosomes are all continuants, even while undergoing changes they do not cease to preserve their identity. Occurrents on the other hand, are never said to exist in full for a single instant of time. Rather they they unfold during successive phases, like for example a viral infection unfolds itself over time. (Biological) processes usually are characterized by passing through different states: where the nucleus is part of the cell, mitosis is a part of the cellular process. The continuant/occurrent opposition corresponds in the first place to the distinction between substances (objects, things) and processes. GOs cellular component ontology is in our terms an ontology of substance universals; its molecular function and biological process ontology are ontologies of function and process universals. But functions, too, are from the perspective of philosophical ontology continuants. For if an object has a given function which means a token function for a given interval of time, then this token function is present in full at every instant in this interval. It does not unfold itself in phases in the manner of an occurrent. If, however, the token function gets exercised, then the token process that results does indeed unfold itself in this manner. Each function thus gives rise, when it is exercised, to processes or activities of characteristic types. 2.3.3 GO’s relations The GO relation isa is referred to as meaning instance of, however in practice it is clearly used in such a way to indicate is a kind of or specialization between universals (e.g. ”p53 is a protein”). Adding to this, sometimes the isa relation is used to indicate part-of, as in the definition of vacuolar proton-transporting V-type ATPase, V0 domain (GO:0000220), which identifies the concept as isa vacuolar part, rather than as a component part thereof. The part-of relation as defined by GO indicates a transitive relation intended to conceptualize ”can be part of, not is always part of”. GO uses the partof relation for representation of parts of both substances and processes, and of functions/activities. The part-of relation is ambiguous in that it does not provide clear means of distinguishing between the following cases: •A part-of any B •A part-of some B •A part-of B, only when a condition holds However useful as a controlled vocabulary, Gene Ontology all to often fails to deliver on the promise of allowing for deduction of new knowledge due to a lack of necessary conceptual rigor. Ega˜na Aranguren et al. [81] captures the essence of this issue by stating that ”The computers’ understanding is determined by the semantics of the language”, thus if the semantics of the language are unclear, so 36 CHAPTER 2. STATE OF THE ART is the computers’ understanding. It is difficult for humans to adequately reason over situations they don’t understand, it is even more so for computers. 2.3.4 Conceptual modeling Conceptual modeling is the practice of creating models for the purpose of either designing an artifact, or achieving a higher understanding of a certain domain. Often these two overlap in a process referred to as Model Driven Architecture (MDA). In MDA the main objective consists of creating high quality software, made possible by extensive use of conceptual modeling techniques. The idea is to create models of the domain and the to-be created application, after which the software can automatically be generated from these models, in some cases without human interference [39]. Clearly, for this process to succeed the model must be formal and unambiguous, i.e. the semantics of the language must be clear. MDA is most often used in an Information Systems (IS) context, but is it so strange to view the biological mechanism of life as an IS? A very complex and organically based, but an IS nonetheless. Replacing human made, silicon based, chips with organic proteins and processor instructions with DNA transcription and translation to proteins. It is often through making analogies with systems we already comprehend, that we come to understand otherwise difficult to grasp mechanisms. Currently, a very attractive, challenging line of research is the personalized medicine context (have a look for instance at [82], where concrete applications of the IS-based working environment defended in this paper are presented). The first documented effort to combine the practice of conceptual modeling is [35] which proposes conceptual models for genomic data. Pastor et al. [83] and [37] then further elaborated on this initial work. It is important to understand that using either approach; conceptual models (CM) or ontologies is really not that different, as a matter of fact an ontology is always present when creating a conceptual model. It is defined implicitly by the model itself, while the other way around is not always true: not every ontology has a visual representation allowing it to be named a conceptual model. The universals versus particulars discussion is easily addressed by CMs: that which is reflected in the model must always be a universal, for instance a gene, having the following attributes: HUGO id, chromosome. While its particulars, for instance a gene that has the HUGO assigned id ”BRCA1”, and which is located on the human chromosome 17. These particulars are instances of the CMs concepts, and usually exists only at runtime. Runtime being interpreted broadly: both an executing object and a database tuple, stored persistently, are considered runtime. The GO expressiveness for conceptualizing relations among entities is not rich enough. By being unable to capture certain knowledge adequately, the uncontrolled use of non-standard operators becomes a tempting, ad-hoc, solution. Conceptual modeling offers a rich variety of relations: an aggregation is a weak whole-part relation where the members can exist without the containing or enclosing class, e.g. Staphylococcus epidermidis forms part of human skin 2.3. ONTOLOGY BASED SOLUTIONS 37 flora, but can survive without a human host. A composition is also a wholepart relation, but stronger than an aggregation such that the whole does not exist without its parts, e.g. a human individual can not exist without a brain. Further, inheritance allows in an intuitive way to identify hierarchies of concepts; a skin cell is a specified type of cell, thus inherits properties of its super type. The concept of a simple relation merely describes the abstract property of allowing communication between two entities: for instance the communication between neurons (this would be a reflective association). In case expressiveness is still not rich enough, the Object Constraint Language [84] allows for even finer specification by allowing the application of textual constraints to the models’ behavior. In general, ontologies and conceptual models are not all that different. We support the idea that in the creation of a conceptual model, an ontology is always present implicitly. The ontology as a formal representation of what is said to be in the modeled subdomain of reality can be inferred from the model itself. A main difference is the visual aspect of conceptual modeling, in which the domain expert can easily understand and therefore cooperate on the creation of the model. This latter implies that conceptual models can be created closer to the problem domain and therefore, we believe, lead to higher quality representations with a better fit with reality. 38 CHAPTER 2. STATE OF THE ART Chapter 3 The Conceptual Schema of the Human Genome Initially an ideal model of the Conceptual Schema of the Human Genome (CSHG) was created, essentially describing how the genomic concepts should be according to the latest of knowledge. However, as data was matched from various external sources to this ideal model, it soon became clear a dichotomy existed between the ideal model and the way data was represented in the real world. A second model was created, logically named the real model. This real model serves as a practical tool of resolving the encountered limitations, it therefore compromises on the aspect of understanding the domain. The intention of the real model is to adapt the modeling elements included in the ideal model to the way in which we found that data are stored and managed in practical settings. As there is always a conceptual mapping between concepts in the ideal model and how they appear in real models, we focus in this work on the ideal model. To simplify the presentation of the schema, 6 conceptual views are considered: the structural view, the variational view, the phenotype view, the transcription view, the genome view and the bibliography/databank view. The joins between these views are pointed out by shaded boxes. We will now proceed to further detail these views in the following sections. 3.1 The structural view A genome is defined as the entirety of an organism’s hereditary information, in this case a human being. By this we mean the full sequence of base pairs that make up the genetic sequence, consult section 1.1 for an introduction to genomics. The genetic sequence is rarely one consecutive DNA molecule, but rather a set of smaller DNA molecules, or chromosomes. These chromosomes then are further decomposed into units of heredity commonly known as genes. CSHG takes the concept of gene as a central point resulting in a gene centric conceptual schema. The structural view aims to capture exactly this by 39 40CHAPTER 3. THE CONCEPTUAL SCHEMA OF THE HUMAN GENOME conceptualizing common genetic concepts like Gene and Allele, figure 3.1. In terms of genetics these concepts are still surrounded with ambiguity, as shown by [2] and [3]. -id_symbol <<oid>> : string -id_HUGO : int -official_name : string -summary : string -chromosome : short -locus : string Gene -ord_num <<oid>> : short -start_position : long -end_position : long -strand : string Allele -Variant * -{id} 1 -sequence (derived) : string Allelic Variant -sequence : string Allelic Reference Type -regulator * -regulated * 1* -id_allele_db : string Allele Data Bank Identification * -{id} 1 -id_gene_db : string Gene Data Bank Identificacion 1..* -{id} 1 -Denoted * 1 -id_variation : int -description : string -id_variation_db : string Variation -Changes * 0..* Figure 3.1: The structural view of the Conceptual Schema of the Human Genome In CSHG, a gene is considered to be a collection of identifying properties and 3.2. THE VARIATION VIEW 41 attributes. id symbol represents an alphanumeric code for the gene according to HGNC [85], it also functions as the primary key; id HUGO, a numeric code assigned to the gene by HGNC; official name, the full name of the gene; summary, a short description; chromosome, the chromosome on which the gene is located and locus, representing the location of the gene within the chromosome. Locus, also often referred to as cytoband, refers to stored information about the subregions of a chromosome that become visible with a microscope after staining during a specific cell cycle phase (Metaphase). An Allele is the ’version’ of a gene as encountered in organisms. By version we mean the exact sequence of base pairs that represent the unit of heredity we identify as gene. A particular gene might have one or many alleles associated. In some cases, different alleles lead to different phenotypes -of which some may have a negative or positive effect in terms of survivability of the individualand in other cases the different alleles don’t provoke differing phenotypes. In CSHG an Allele thus has a start position and an end position, relative to the beginning of the chromosome and a strand attribute to identify on which of the two DNA strands it is located, indicated by either ’plus’ or ’minus’. We identify two types of Alleles: Allelic Variants and Allelic Reference Type, each of which has a single attribute sequence. The idea here is that a reference allele is said to exist and represents a ’standard’ sequence of a particular gene, in the context of this work the so-called NCBI refSeq [18] is used. The Allelic Reference Type then can have various associated Variations, this latter concept will be discussed in detail in section 3.2. Each of these Variations triggers the possibility of a different Allelic Variant, which sequence is thus derived from the Allelic Reference Type and Variation. The Gene Data Bank Identification and Allele Data Bank Identification entities will be discussed into more detail in section 3.6. 3.2 The variation view Genetic variation is brought about by mutation, a change in the chemical structure of a gene. A variation can have either a negative, positive or neutral effect on phenotype. Chapter 5will go into more detail exactly about this latter effect and how the conceptualization was introduced in CSHG. Natural selection avoids the reproduction of individuals with genetic variation that poses evolutionary disadvantages. Genetic variation with negative effect is thus expected to occur infrequently. Section 1.1.4 describes genetic variations in more detail. 48CHAPTER 3. THE CONCEPTUAL SCHEMA OF THE HUMAN GENOME Transcription is the biological process in which the genetic sequence in DNA is transferred, or transcribed onto RNA, a molecule very similar to DNA. This allows the genetic sequence to leave the cell nucleus, in order to reach other biological systems that will eventually lead to the construction of complex proteins. A gene can be thought of a as a set of separate blocks of nucleotides. The coding blocks, or exons, eventually form the transcript, although which exons are included in which version of the transcript may vary. The process that regulates which exons are included, and discards all of the introns is called splicing, see [2] for more information. The Transcription view in [fig. 2], models the basic steps in protein synthesis. The Primary Transcript class represents the transcribed copy from DNA to RNA of the TranscribedSequence. In the biological process of transcription, the primary transcript is an RNA molecule, containing a literal copy of the DNA sequence of a gene, including all coding (exons) and non-coding (introns) fragments. Its sequence attribute is a derived attribute from the Segment class. The PrimaryTranscriptPath class models the different splicing factor-driven partitions of the Primary Transcript, its attribute ord num identifies a partition from the complete set of partitions of a Primary Transcript. In the ElementTranscript class, the ord num attribute identifies a specific fragment within the partition. The Exon and Intron classes specialize the type of partition fragments. The Spliced Transcript class represents different exon combinations (sequences) of a Primary Transcript, its ord num attribute identifies it among all the allele spliced transcripts. The result of these combinations will be the mRNA and others RNA types (specialized classes from Spliced Transcript). The mRNA contains a nucleotide sequence that could potentially encode a protein, this is known as ORF (Open Reading Frame). The id attribute of an ORF identifies it in the system, and the sequence attribute stores the codifying sequence. The Primary Polypeptide class describes the protein primary structure: the amino acid chain obtained as a result of the translation of an ORF. This amino acid chain undergoes chemical transformations and the final result is a functional protein, represented in the model as the Protein class. A protein can consist of one or more Primary Polypeptides. In the Protein, its name attribute represents the name of the resulting protein, and its sequence attribute the amino acid sequence. The ChromosomeSegment class represents the segments that comprise the chromosome. This class has a sequence attribute which stores the corresponding DNA sequence delimited by start position and end position attributes. A chromosome has two main types of segments: coding-related segments (GenicSegment) and non coding-related segments (NonGenicSegment). Two classes specialize NonGenicSegment: the IntergenicRegion class (the regions between genes) and ChromosomalElement class. The last one has three specializing classes that describe other elements of the chromosomes (Centromere,Telomere and ORI ) whose function is to keep the chromosome functional and are not involved in protein production. 3.5. THE GENOME VIEW 49 3.5 The genome view The Genome view, figure 3.5, models individual human genomes. This view is interesting for future applications, since massive parallel sequencing technologies will allow the complete sequencing of individual genomes at a very low price in the near future [87]. -id <<oid>> : short -description : string Genome -number <<oid>> : short -id_copy <<oid>> : short Chromosome -{id} 1 * -start_position <<oid>> : long -end_postition <<oid>> : long -sequence : string ChromosomeSegment -{id} 1 -Formed * Gene-Segment NonGenicSegment -Couple1 1 -Couple2 1 IntergenicRegion ChromosomaElement Centromere Telomere ORI -name <<oid> : string -description : string Research Centre -{id} 1 * -id_symbol <<oid>> : string -id_HUGO : int -official_name : string -summary : string -chromosome : short -locus : string Gene 1..1 -Corresponds 0..* -id_variation : int -description : string -id_variation_db : string Variation -SegmentChanges * * Figure 3.5: The genome view of the Conceptual Schema of the Human Genome 50CHAPTER 3. THE CONCEPTUAL SCHEMA OF THE HUMAN GENOME The class Research Centre represents the labs or research centers where an individuals’ human genome was sequenced. A genome (Genome) is considered a set of chromosomes (Chromosome). The number attribute identifies a chromosome in a genome. The couple relation on the Chromosome class represents the concept of homologue pairing, i.e. every human cell will carry two equivalent chromosomes -one from the father and one from the motherwith the same genes but different alleles for each gene. 3.6 The bibliography/databank view -id_symbol <<oid>> : string -id_HUGO : int -official_name : string -summary : string -chromosome : short -locus : string Gene -ord_num <<oid>> -sequence (derived) : string Spliced Transcript -ord_num <<oid>> : short -start_position : long -end_position : long -strand : string Allele -Variant * -{id} 1 -ord_num <<oid>> : short -start_position : long -end_position : long -sequence (derived) : string Segment -id_allele_db : string Allele Data Bank Identification *-{id} 1 -id_gene_db : string Gene Data Bank Identificacion 1..* -{id} 1 -id <<oid>> -title -authors -abstract -publication Bibliography Reference -Bibliography Name DB <<oid>> -URL Bibliography DB * -{id} 1* * ** * * * * -regulator * -regulated * -id <<oid>> : string -name : string -description : string Data Bank 1..* -{id} 1 1..* -{id} 1 -id_variation : int -description : string -id_variation_db : string Variation * * -Obtained1 * Figure 3.6: The bibliography/databank view of the Conceptual Schema of the Human Genome Chapter 4 Solving the HGMD case The first part of this work centers on emphasizing the importance of applying conceptual modeling techniques before constructing any Information System, and Genomic Information Systems in particular. The second part starts here and describes the results of loading the BRCA1 gene data from HGMD into the database that results from the Conceptual Schema of the Human Genome. The encountered problems serve as examples and a case proof of why an Information System without a conceptual modeling sound backbone results in myriad difficulties and undesired behavior, thereby enforcing our earlier mentioned statement. [83] reports a study of comparing the HGMD to the CSHG, in order to identify a conceptual mapping between the two. It is this mapping that is followed in this document, and the following section will report the encountered problems for actually loading the information from the HGMD into the HGDB for the BRCA1 gene. Roughly problems can be separated in two categories; intrinsic data properties and data representation, with the final intention of providing a concrete report on what type of problems the correct load of a conceptual model-based genome data base must face and solve. This is important because the main benefits of applying conceptual modeling principles to the understanding of the human genome are located in both having the right data structure and the right data contents. Identifying and classifying these data loading problems provide general solutions that can help in solving the same problems in other biological data loading contexts. Verifiable incorrect, inconsistent or incomplete data (tuples) are examples of these encountered mishaps with the actual data, or intrinsic data properties. Difficulties associated to physically extracting the data from the external source and ambiguously description of mutation properties are typical examples of data representation problems. Naturally, the division between the two categories is not strict and thus some overlap exists, it is however useful to keep in mind that intrinsic data property problems tend to be affecting the entire genetics domain, while the data representation difficulties are restricted to HGMD. 51 52 CHAPTER 4. SOLVING THE HGMD CASE 4.1 HGMD description The Human Gene Mutation Database (HGMD) represents an attempt to collate known (published) gene lesions responsible for human inherited disease. This database, whilst originally established for the study of mutational mechanisms in human genes [28], has now acquired a much broader utility in that it embodies an up-to-date and comprehensive reference source to the spectrum of inherited human gene lesions. Thus, HGMD provides information of practical diagnostic importance to (i) researchers and diagnosticians in human molecular genetics, (ii) physicians interested in a particular inherited condition in a given patient or family, and (iii) genetic counsellors. The HGMD can be accessed either publicly - having less entries but no fee is chargedor professionally -in which case a fee is charged for access to the full data set. The HGMD provides an overview of mutations causing or associated with human inherited disease. Adding to that, the repository includes disease-associated polymorphisms reported in literature. In general, these data consist of various types of genetic variation. In the current version of HGMD, both somatic and mitochondrial variations are not included. All genetic variations in HGMD have been found as a result of DNA analysis, excluding thus variations inferred from amino acids sequencing. The latter is to avoid ambiguity issues that still surround DNA sequence changes. Genetic variation in coding regions that does not alter the encoded amino acid are also not included. HGMD extracts its data from over 250 journals, which are not further specified. The papers are manually read and the data interpreted by experts, ensuring the data is in theory high quality, but vulnerable to human error. HGMD also includes some data from LSDB’s. At the time of writing (20-07-2011), the HGMD contained 113 247 genetic variations, of which 82 808 are publicly accessible. 4.2 Encountered problems HGMD distinguishes 10 mutation types: Missense/nonsense, Splicing, Regulatory, Small Deletions, Small Insertions, Small Indels, Gross Deletions, Gross Insertions, Complex Rearrangements and Repeat Variations. Roughly all the types can be mapped to the Variation and Precise concepts of the CSHG, except for the Gross Deletions, Gross Insertions, Complex Rearrangements and Repeat Variations. The latter are described in a very unstructured manner, almost natural language, and are thus considered impossible to process automatically. Appendix A.1 provides a screenshot of how precise mutations -the ones that have a precise description including position, change and phenotypeand appendix A.2 provides a screenshot of those earlier mentioned unstructured -or imprecisemutations. The CSHG facilitates these tuples as Imprecise, which stores a description of the mutation. 4.2. ENCOUNTERED PROBLEMS 53 4.2.1 Intrinsic data properties In some cases the HGMD mutational data simply lacks entries. For instance, the splice mutations overview provided by HGMD mentions 5 mutations in intron 22, while [88] states at least 2 other mutations; IVS22+67(T>C) and IVS22+8 (T>A). Three concrete examples of this problem were encountered, all three in splicing site mutations. However, this particular type of problem is very difficult to detect, since finding them involves rereading the articles HGMD provides which is hard to automate. Thus, although only three concrete occurrences of this problem have been encountered, it is likely more exist. Splicing site mutation CS961492 describes a C>T mutation, as a possible phenotype HGMD indicates Breast cancer. However, having the read the corresponding article [89], not once breast cancer is mentioned in combination with this mutation. The article does mention the mutation as being affiliated with men suffering from prostate cancer. Thus, deducing from the rather limited information made available by HGMD on this specific mutation, it is concluded HGMD made an error during data entry. Splicing site mutations CS063247 and CS011027 should be located near intron 4, according to the HGMD splicing site mutations overview. However according to the splice junctions overview HGMD provides, there exists no intron 4, nor an exon 4. Since indeed both of the papers [90] and [91] state the mentioned mutations near intron 4, it would be logical to presume the problem is on HGMDs side. So at first, this specific problem instance was considered to be either a major flaw in HGMD due to inconsistent reference sequences, or a result of human error during the HGMD loading procedure. However, deeper research revealed a more subtle situation. splicing site mutations are located in HGMD by using a splice junctions overview, which provides an overview of intronand exon borders in the gene. HGMD constructs this overview by using a NCBI reference sequence, in this case L78833. This reference provides a comment in natural language that explains the absence of an exon 4 as: ”Characterization of an aberrant BRCA1 cDNA clone in the original report [92] led to the misidentification of an inserted Alu element as exon 4. Not normally found in BRCA1 transcripts, insertion of this Alu would lead to introduction of a STOP codon. Hence, BRCA1 exons and introns are numbered 1a, 1b, 2, 3, 5, 6, etc.”. Splicing site mutation CS012667 indicates a G>A mutation in nucleotide +3 from the start of intron 2 . However neither the HGMD splice junction overview, nor the NCBI reference gene sequence indicates a G-nucleotide at this location. A very similar event happens with splicing site mutations CS001825 and CS991331. They both involve a mutation located 7 nucleotides upstream (+7) from the start of intron 22. However, the first mentions an A>G mutation, while the latter describes, for that exact same location, a T>C mutation. Since both the NCBI reference sequence and the HGMD splice junctions overview indicate an A nucleotide at the appropriate position, and the CS001825 reference article [93] indeed mentions an A>G mutation at the specific location, one could easily conclude the CS001825 A>G mutation in the correct one. However, the 54 CHAPTER 4. SOLVING THE HGMD CASE CS991331 reference article [88] does indeed point out a T>G mutation at the location, so the truth might be slightly more subtle. Since [88] involves an African-American population, while [93] entails a Chinese population, a possible reason for the irregularity might be general genetic differences between those ethnic groups. A phenomenon referred to as Single Nucleotide Polymorphisms, or SNP’s [94]. 5 Occurrences of this problem have been identified in splicing site mutations, 1 in Small Deletions and Small Insertions each, leading to a total of 7 occurrences. SNP’s are currently included in the CSHG as a separate concept, facilitating the placement of these type of mutations. Finding these type of genetic variations, mixed in with another type of variations does emphasize the need for an unambiguous data representation. 4.2.2 Data representation The HGMD public version, used for this research presents data through a website in HTML tables. This makes an automatic extraction procedure very difficult, but not impossible. Using a ’screen-scraping’ approach including HTML parser technology, the individual tuples can be isolated and extracted. This approach however has several drawbacks, among which its inflexibility of coping with changing environments, which especially in this rapid evolving domain is undesirable. It deserves mentioning here that the professional license, available for a fee, does allow for acquisition of the HGMD database as a .sql file, making this extraction process a lot easier. Also, the public version has a delayed update cycle, while the professional license includes the entire up-to-date dataset. Some data is provided in natural language. For instance the fact that the first two BRCA1 exons are alternative non-coding exons is only mentioned in the header of the Splice Junctions overview: ’The first 2 exons are alternative non-coding exons and The translation initiation codon is located within exon 2, this mainly affects locating splicing site mutations (1 instance). Adding to this, in Small Deletions (2 instances) and in Small Insertions (3 instances) some mutations are located through mouse-over tags, the information communicated by these tags is highly unstructured to a degree that we might call it natural language as well. Also, in the case of imprecise mutations (Gross Deletions, Gross Insertions, Complex Rearrangements and Repeat Variations), the greater part of the information presented by HGMD is in natural language, impeding an automated approach severely in the affected cases. In some cases, the HGMD database uses different ways of locating mutations, within the same type of mutations. For instance, Small Insertion mutations CI030168, CI962219 and CI022582 happen in non-coding areas of the gene, just like the Small Deletions mutations CD991644 and CD994433. Since HGMD generally uses a cDNA codon referenced way of locating these types of mutations, and given that non-coding sequences simply not exist in the cDNA, HGMD locates these earlier mentioned mutations in a different way. In the case of Small Insertions, HGMD provides a Splice Junction reference, very much like the method used to locate splicing site mutations. In this case the CI030168, CI962219 and CI022582 mutations are located at IVS20+21, IVS20+48 and 4.2. ENCOUNTERED PROBLEMS 55 IVS20+64 respectively. So IVS20 indicates the intron number, where +21 indicates the offset, however since no acceptor/ donor information is provided, it is unclear from which side of the intron the offset should be referenced. In the case of Small Deletion mutations CD991644 and CD994433 at first sight, no indication of how to locate them is provided. However, this information is provided through mouse-over tags in the Splice Junctions referenced form, described earlier. CD991644 is thus located by I7E8-24, aka IVS7 -15 del10. and CD994433 is located by I12+34 / polymorphism ?. This problem was thus encountered 3 times in Small Insertions and 2 times in Small Deletions, making a total of 5 occurrences. As said, HGMD refers to codons in many cases, which are sets of three nucleotides. Since HGDB will be using a nucleotide referenced position to locate mutations, a transformation of HGMD provided data is necessary. In theory, acquiring the correct nucleotide would be a matter of multiplying the codon number by three, reality is slightly more complicated as will be discussed in the next paragraph. HGMD uses this way of locating mutations in the case of Missense/nonsense (320 instances), most of the Small Deletion mutations (288 instances), most of the Small Insertion mutations (98 instances) and Small Indels (11 instances), leading to a total of 715 instances of this problem. Retrieving the corresponding nucleotide in DNA would simply be a question of multiplying the codon number by 3, if not for the existence of introns and exons. cDNA only comprises of the genes’ exons, thereby excluding the introns, contained in the DNA. Due to this fact and given that HGDB will be incorporating a DNA referenced scale, a linear transformation, by multiplying the codon number by 3, simply is not possible. HGMD splicing site mutations are located by referring them to so-called splice-junctions. These splice junctions indicate the borders between exon and introns. HGMD thus indicates an intron border, by giving an intron number and specifying which border by providing either a donor- (ds) or acceptor- (as) site of the intron. The donor site corresponds to the side closest to the 5’ end of the DNA strand, while the acceptor site corresponds to the side closest to the 3’ end of the DNA strand. Then an offset is given, to indicate the amount of nucleotides between the indicated splice junction and the actual mutation. In the so-called splicing mutations overview HGMD then provides a sample sequence for each intron/exon-junction contained in the gene. This method of locating mutations is used primarily in splicing site mutations (80 instances), but in some exceptional cases HGMD also uses this notation to provide locational data for other types of mutations. For instance, In Small Deletions (2 instances) and in Small Insertions (3 instances). In the HGMD data exists ambiguity; for instance, mutations may or may not result in a certain phenotype, this is indicated by a question mark following the supposed phenotype. However, no probability scores are stated and a mutation without a (noticeable) phenotype is considered to be a variation with neutral effect. Since variations and mutations are considered to be two different concepts in the HGDB datamodel, this poses problems with loading the database correctly. 94 instances of this problem have been identified: missense/nonsense 56 CHAPTER 4. SOLVING THE HGMD CASE mutations account for the most instances (73), splicing site mutations contains another 16, small deletion mutations 2 and small insertion mutations account for 3 instances. In short to summarize the encountered problems in the data loading process. Nine distinct problems have been identified, each of which has been discussed separately in the above section. The following sections will provide useful insights on how to resolve future instances of these. The problems as encountered, will now be summarized in the following list. Intrinsic data properties 1. Lacking data 2. Data entry errors 3. Reference to non-existing introns 4. General inconsistency Data representation 1. Data difficult to access 2. Use of natural language 3. Inconsistent way of positioning mutations 4. Inconsistent with NCBI refSeq 5. General ambiguity 4.3 Data loading problem solutions Inserting the 804 precise and 90 imprecise variations provided by the noncommercial version of HGMD manually seemed like a cumbersome and more importantly, error prone operation. For this reason a series of scripts was devised to automate the procedure, while at the same time providing useful experience and knowledge about how to further process the variational data automatically. A full technical report on this process, including the source code of these scripts can be found in [95]. The basic function of the software is to convert the HGMD provided mutational data into the format used by HGDB and detect inconsistencies in the data. The main tools are the cDNAsequences provided by HGMD [Genebank, U14680] [Genebank, L78833], the NCBI DNA reference sequence [NCBI NG 005905.1] and the NCBI Coding Sequences (CDS, located in NCBI RefSeq NG 005905.1). The HGMD cDNA sequence is the aggregation of all coding sequences (exons) of the BRCA1 gene, thus excluding introns. The NCBI DNA sequence then is the complete sequence of the BRCA1 4.3. DATA LOADING PROBLEM SOLUTIONS 57 gene, including introns. The CDS information specifies what parts of the NCBI DNA sequence are coding and which are not. The scripts extract, and in some cases, calculate the variables which are to be inserted into the HGDB Variation, Precise and Imprecise tables. At the same time detecting inconsistencies in the HGMD database that can not be resolved in an automated way, indicating a manual approach is needed in those cases. Since dealing with the above mentioned problems was necessary to devise those scripts, they are considered to be the crystallized solutions to the earlier mentioned problems. The main disadvantage of this screen-scraping approach, is high vulnerability to changes in HGMD data structure. This means the scripts will require regular updating. Also, by depending on a medium like HGMD, trust is invested in the integrity of the source. However, at this point the exact reliability of HGMD has not been identified and some of the encountered problems during this project clearly suggest reasons to doubt this reliability. 4.3.1 Intrinsic data properties HGMD lacks data entries for two reasons. HGMD might simply not be up-todate with the latest information provided by scientific research on the subject, or HGMD missed entries during the manual loading of the database. A partial solution to this problem involves using the professional version of HGMD, which includes more and more up-to-date entries. However, it is also a characteristic of the immatureness of the field that no full coverage exists, simply not enough research has been performed to identify all existing mutations. Therefore the database will inherently be incomplete, and thus resolving falls outside the scope of this paper. HGMD provides erroneous data for a variety of reasons, human error on either HGMD or the source paper side might be an issue. Inconsistencies between HGMD and source paper notation style might play a role. In either case, this is a very difficult problem category to detect since detection involves rereading the papers. A possible, but unsatisfactory to some degree, solution would be to perform a deep investigation on a limited amount of mutational data provided by HGMD, thereby uncovering error frequency. This error frequency, provided it is investigated according to scientific measures, can then be extrapolated to the rest of the database, thus providing a handle from which to calculate reliability of the data-set. Since inconsistencies can be detected, a manual check of the apparent inconsistent data is possible. Indeed, the scripts indicate only a few inconsistencies and so the shear amount of papers to be read manually can be reduced drastically, making the manual approach possible in these cases. In case the source is locating a mutation within a non existing intron, the solution might be to manually complete the splice junctions overview, combining the information from the NCBI reference sequence and coding sequences information (CDS), however contradiction exists about whether an intron 4 actually exists. The implications of this, as has been discussed earlier that reference papers are using different reference sequences, are far more serious and indicate a structural flaw 64 CHAPTER 4. SOLVING THE HGMD CASE Chapter 5 Solving the dbSNP case Explaining the various uses of the SNP concept, separating them into their appropriate conceptual context and presenting a conceptual modeling approach will allow solving this ambiguity. This is achieved by answering the main research question; What makes a genetic variation, classified as a SNP different from those genetic variations not classified as SNP?. The main contribution is to show how a model driven approach to the genomic domain can aid in the understanding and adequate representation of relevant concepts. Previous solutions applied to the problem space, including ontologies [72] and literal descriptions in natural language, often fail to comprehensively model the concept. As a result of the ambiguity intrinsically associated to natural language, defining the concept exhaustively this way has proven to be a daunting task. In addition to this, relating the concept to its context is equally as important as defining it intrinsically, and conventional methods often fail exactly at satisfying this requirement. The contribution of this work is twofold however, on the one hand fixing the semantics of the concept in a conceptual model while on the second hand enforcing the use of conceptual models in the bioinformatics domain. Bioinformatics has seen a rapid evolution in the past decade, starting with the sequencing of the human genome in 2000 by Venter and Collins [9], and the amount of data being generated is constantly growing. Managing these data has been done by ad-hoc solutions mostly until now. These solutions often do not cope well with changing environments –one that bioinformatics is without a doubt– scale poorly and do not support reusability well. Conceptual modeling, as the cornerstone of the Model Driven Architecture, has been proven to show significant improvements in engineering Information Systems [39]. Conventional solutions require a large amount of manual coding effort each time the domain changes in order to keep implementations in line. The conceptual modeling based solution allows for a process much closer to the problem space by allowing the engineer to create models that can be understood by the domain expert. From these models partial, or in some cases full, implementations can be generated following sets of rules and formal transformation processes, greatly reducing the effort required to maintain implementations consistent with a changing domain. 65 66 CHAPTER 5. SOLVING THE DBSNP CASE 5.1 SNP definition The dbSNP data repository [17] is an NCBI [97] driven effort to collect and integrate all available data on genetic polymorphisms. The name suggests emphasis on single nucleotide polymorphisms, but reality is slightly different. As is mentioned on the dbSNP website: ”This collection of polymorphisms includes single-base nucleotide substitutions (also known as single nucleotide polymorphisms or SNPs), small-scale multi-base deletions or insertions (also called deletion insertion polymorphisms or DIPs), and retroposable element insertions and microsatellite repeat variations (also called short tandem repeats or STRs)”. Clearly NCBI is stretching the definition of SNP to fit any type of polymorphism, thereby blurring the concept. Confusing as it may be, we believe dbSNP should actually be called dbVariation, as stated by NCBI itself in the Frequently Asked Questions section of their website. Yue and Moult [98] use a very similar definition of SNP, stating that a SNP can be either neutral or deleterious. This resource appears to be using the literal meaning of the label, simply indicating every polymorphism that influences a single nucleotide as being a SNP, regardless of its effect on phenotype. A different definition originates from the Department of Biological Sciences, Oakland University, Rochester, MI, USA [99]: ”Single nucleotide polymorphism (SNP) is the simplest form of DNA variation among individuals. These simple changes can be of transition or transversion type and they occur throughout the genome at a frequency of about one in 1,000 bp. They may be responsible for the diversity among individuals, genome evolution, the most common familial traits such as curly hair, interindividual differences in drug response, and complex and common diseases such as diabetes, obesity, hypertension, and psychiatric disorders. SNPs may change the encoded amino acids (non-synonymous) or can be silent (synonymous) or simply occur in the noncoding regions. They may influence promoter activity (gene expression), messenger RNA (mRNA) conformation (stability), and subcellular localization of mRNAs and/or proteins and hence may produce disease”. It resembles the conventional definition loosely, as it mentions a frequency of occurrence in an individual (about one in 1,000 bp) and pathogenicity (and complex and common diseases such as). Depending on interpretation, one could state that also the size dimension is mentioned implicitly, considering the definition handles single nucleotide polymorphism. 5.1.1 SNP characteristics During the investigation of various definitions of the SNP concept existing in todays genomic domain, the determinant characteristics, or dimensions, have been identified. They form the conceptual spectrum in which the SNP concept is positioned. Each concrete definition represents an instance of these abstract dimensions, thereby fixing a position within the gamut. It is expected that by incorporating these dimensions into a database, instead of fixing the concept on one definition, flexibility is acquired and data can be queried dynamically depending on the desired SNP interpretation. The commonly accepted defini- 5.1. SNP DEFINITION 67 tion for SNP is explicitly based on 2 dimensions: occurrence in the population and genotypic size. According to [100] a SNP is a single base change in a DNA sequence. It is considered that the least frequent allele should have a frequency of 1% or greater. The first dimension, occurrence in population, hides a third and fourth characteristic. As a direct result of natural selection, for a polymorphism to happen more than a given percentage in the population, it needs to have either a neutral or a positive phenotypic effect. A negative, or deleterious effect is impossible to happen in a substantial portion of the population as it eliminates itself. Therefore, SNPs considered according to the conventional definition have no direct relation to disease. For this very same reason, SNPs as defined in the conventional sense are relatively common among each individual. On average, ∼8.33 SNPs happen every 10kb [94]. Various literature sources use various definitions of the SNP concept. dbSNP forms one extremity of this spectrum by defining the concept very loosely. The conventional definition then describes the other side of the spectrum by being more rigorous. Following the various definitions leads to the identification of four dimensions on which SNPs are commonly defined: occurrence in population, occurrence in sequence, pathogenicity and size. Every dimension will now be clarified briefly. Occurrence in population This dimension entails the aspect of occurrence of a SNP in a given population. Examples of different populations include Asian, African or European. A higher population occurrence than a fixed low percentage, suggests a polymorphism with neutral effect on phenotype. This barrier is not defined rigorously, but commonly accepted percentages vary from 0.5% to 1%. Occurrence in sequence Due to natural selection, polymorphisms with deleterious or negative effects are usually filtered out of a population. This process conserves neutral polymorphisms without observable effect and polymorphisms with positive effect on phenotype. As a result of this mechanism, these type of polymorphisms stack up in the genome of an individual. According to conventional definition, in the human genome every ∼1200 base pairs a SNP is expected as stated by [94]. This dimension however, describes a characteristic of the SNP concept in general. Indeed, a specific SNP does not happen once every ∼1200 base pairs, rather a different SNP is expected to be occurring with this frequency. It is therefore a characteristic of a genome; stating the amount of SNPs happening in it, rather than being a defining criterium whether a specific polymorphism is considered to be a SNP or not. 68 CHAPTER 5. SOLVING THE DBSNP CASE Pathogenicity The association of a SNP to hereditary disease is captured in this dimension. The general belief is that SNPs are not causatively associated to disease. Rather, due to a phenomenon called Linkage Disequilibrium [101] and by using SNPs as markers, susceptibility to hereditary disease can be detected without having to sequence an entire individuals genome. In short, the SNP itself has no relation to a pathogenic effect. SNPs used as marker, however, predict a different mutation, happening at another position. This second, predicted, mutation usually does have a direct pathogenic effect. Size Clearly, a single nucleotide polymorphism should involve only one nucleotide, hence the name. However, considering the definition in which the SNP label is used to identify non deleterious polymorphisms this description might be stretched. Considering that DNA is read in triplets (combinations of three nucleotides), the polymorphisms that are not a multiple of three often result in frame-shift, changing the transcribed mRNA and thus the resulting protein. Frame-shift often leads to an entirely different protein, and is thus very likely to result in negative phenotype. A non-deleterious polymorphism, in theory, can thus be any variation in which the affected nucleotides are a multiple of three. However, every changed triplet, if non-synonymous, also changes an amino-acid eventually leading to an increased chance of significantly changing the resulting protein. Thus, although in theory a non-malicious polymorphisms might include more then one nucleotide, it is most likely to involve a single nucleotide change. 5.1.2 Uses of the SNP concept For a proper understanding of the genomic concept known as SNP, and to define it properly in terms of a conceptual model, it is necessary to know what purpose the term serves. Not only needs to be investigated what the domain exactly considers to be a SNP, but also how this concept is used in the domain. This exact understanding of the concept, will allow for an adequate application of the conceptual modeling approach. After applying the conceptual modeling perspective to the problem of properly specifying SNPs, the concepts behind them will less ambiguous and clearer. By more clear is meant that specific uses in particular contexts are labeled with an adequate term, while the SNP notion emerges clearly from the conceptual schema. Three distinct uses of the SNP concept have been identified. (i) dbSNP, (ii) distinguishing harmless polymorphisms from harmful variations and (iii) as a genetic marker. dbSNP uses the term to cover a wide variety of polymorphisms known to happen in the human genome. It thereby stretches the conventional definition of the term, and should thus actually be called dbVariation, as it mentions itself. Another application of the term is identified when assuming SNP to be defined according to measures on the dimensions shown in 5.1. 5.2. SNP INCORPORATION IN THE CSHG 69 When sequencing a human genome, various polymorphisms are usually encountered. Every ∼1200 base pairs, a mismatch happens between the sequence used as reference and the sample at hand. Most of these polymorphisms happen without negative effect on the phenotype, as has been explained earlier. Considering that DNA is read in triplets, or combinations of three nucleotides, the polymorphisms that are not a multiple of three often result in frame-shift, rendering the transcribed mRNA sequence senseless. This taken into account, the most probable size in which a polymorphism is expected to not alter the genetic sequence is one nucleotide, or a multiple of three. In the absence of negative effect, they are likely to happen in more than 0.5% of the population. Hidden among these relatively innocent polymorphism, lurks the one that is associated to the development of disease. More than one polymorphism with negative effect on phenotype, happening in any individual is very unlikely. Discriminating between polymorphisms with negative effects, and polymorphisms without is thus very important. The SNP concept, shaped in the way as is done in 5.1 serves this purpose elegantly. Table 5.1: The SNP concept defined to discriminate polymorphisms with neutral effect on phenotype from polymorphisms with negative effect on phenotype. Property Value Occurrence in population >0.5 % Pathogenicity None Size 1 nucleotide SNPs are also commonly used as genetic markers. Due to relative high pricing of DNA sequencing, ways of using the available sequencing capacity efficiently had to be come up with. One approach to this cost-saving strategy, is making use of linkage disequilibrium. LD describes the nonrandom correlation that exists between a marker and the presence of polymorphisms at a locus [101]. Following LD it is possible to asses the probability of a polymorphism happening in a subjects genetic code, by knowing the state of a marker position. Vignal [100] describes the use of SNPs as molecular markers in animal genetics. The use of SNPs as genetic markers can thus be considered an imperfect solution to high costs associated to DNA sequencing. 5.2 SNP incorporation in the CSHG Incorporating SNPs into the Conceptual Schema of the Human Genome is essential. Figure 5.1 demonstrates a representation of the extension. It provides means of discriminating various types of polymorphisms. The most important being the separation between harmful and harmless genotypic variations. In the schema, harmful variations are considered variations with mutant effect, while SNPs are considered to be variations of the type neutral polymorphism. It is for this reason that data on SNPs need to be stored in such a way, that it facilitates the earlier mentioned dimensions on which the concept is defined. 70 CHAPTER 5. SOLVING THE DBSNP CASE This will allow for dynamic creation of queries, extracting the correct tuples depending on which definition at runtime is chosen. To solve the ambiguity associated to the SNP concept, an often considered sufficient solution involves providing a concise, literal description of the term. Due to the various uses of the concept, a static definition is believed to be insufficient. Also, this approach fails to relate the concept to its context, an aspect considered very important in the complex reality of genomics. Rather, coping with the variability of the term seems appropriate and is facilitated by the extension to the conceptual model of the human genome. In practice, this means that the earlier identified dimensions of the definition needs to be represented in the schema, allowing for dynamic capture of the concept. An instance of the concept can then be instantiated with different values, thereby allowing for the different uses and definitions that have been identified. Thus, by separating the instance of the concept from the structural definition, flexibility is obtained. First of all, the SNP concept is considered to be a specification of the Indel concept, which in turn is a specification of the Variation concept. It deserves mentioning this implementation does not rule out SNPs of size greater then 1, leaving room for flexibility on the size dimension. The occurrence in population (populationOccurrence) is regarded as a SNP specific attribute. Occurrences in populations might differ among various populations, it is therefore crucial to also store the investigated population that resulted in the discovery of the SNP at hand. Other polymorphisms, either mutant ones or with unknown effect, are expected to have different effects among populations as well. It is for this reason, the Population concept is related to both the SNP and Variation concepts. The pathogenicity dimension relates the neutral polymorphism concept in the conceptual model to the SNP concept, securing that a SNP is never associated to disease. According to the schema, a polymorphism associated to negative phenotypic effects is considered to be mutant and is thus captured already in the form of a Variation, the SNP super-type. 5.2. SNP INCORPORATION IN THE CSHG 71 -ord_num <<oid>> -sequence (derived) : string Spliced Transcript -description : string Imprecise -position : int Precise -Changed0..* 1..* -sequence (derived) : string Allelic Variant -sequence : string Allelic Reference Type RegulatorSequence -ord_num <<oid>> : short -start_position : long -end_position : long -sequence (derived) : string Segment GenicChromosomicNeutral PolimorphismMutant -identificacion? -description : string Chromosomic Mutation 1 -Influences 1..* Location Description -Denoted * 1 -id <<oid>> -title -authors -abstract -publication Bibliography Reference Unknown consequence OthersSplicing MissenseRegulatory Effect 1 1 1 1 -sequence : string -repetition : int Insertion -bases : int Deletion -ins_sequence : string -ins_repetition : int -del_bases : int Indel -bases : int Inversion -id <<oid>> : string -name : string -description : string Data Bank -id_variation : int -description : string -id_variation_db : string Variation -Changes * 0..* * * -obtained from 1 * -id_syndrome : int -name : string -description : string Syndrome -id_value : int -description : string Value * * ** -populationOccurrence SNP Population * * -is predicted by * -predicts * -occurs in * -is known to have * Figure 5.1: Conceptual model of the human genome including the SNP concept. 72 CHAPTER 5. SOLVING THE DBSNP CASE As one might notice, the dimension occurrence in sequence is not incorporated in the schema. The occurrence in sequence represents not a characteristic of a single SNP, rather a characteristic of a specific genome. Indeed, a specific SNP does not occur every 1200 base pairs, but rather every 1200 a SNP is found. It is therefore not associated to the SNP concept and thus left out of the schema. Also, it is considered unnecessary to be stored, for the reason that it can be derived for every genome from the database itself, simply combining the data on the total amount of nucleotides in an individuals genome, and the amount of found SNPs. To incorporate the marker functionality of the SNP concept, a new relation would need to be made between SNP and Variation. This relation would consist of a probability score, indicating the probability to which the SNP at hand predicts the variation. This use of the concept is considered to be a result of the immatureness of the domain. Using SNPs as markers merely satisfies a costefficiency strategy, and is expected to be rendered useless in the near future by advances in DNA sequencing technology. Although a correlation exists between neutral SNP occurrence and malicious polymorphisms, there appears to be no causal relation. It is for this reason, that has been decided not to incorporate this use of the SNP concept in the conceptual model of the human genome. 5.3 dbSNP description The Single Nucleotide Polymorphism database (dbSNP) [17] is a public-domain archive for a broad collection of simple genetic polymorphisms. This collection of polymorphisms includes single-base nucleotide substitutions (also known as single nucleotide polymorphisms or SNPs), small-scale multi-base deletions or insertions (also called deletion insertion polymorphisms or DIPs), and retroposable element insertions and microsatellite repeat variations (also called short tandem repeats or STR’s). Each dbSNP entry includes the position, the sequence context of the polymorphism (i.e., the surrounding sequence), the occurrence frequency of the polymorphism (by population or individual), and the experimental method(s), protocols, and conditions used to assay the variation. As established earlier, the dbSNP database contains much more then simply SNP’s, making the extraction process described below significantly more difficult. dbSNP stores data for 100 species, with a total of 87 536 603 entries (Rs#’s), of which 29 813 700 have been validated. Currently (20-07-2011, dbSNP build 132) the dbSNP repository contains 30 442 771 entries -of which 19 727 605 have been validatedfor the Human genome. 5.4 Encountered problems The problems encountered when loading dbSNP into HGDB can roughly be divided in two categories: data quantity and semantic issues. The first refers to the intrinsic difficulties of processing the large files, 22Gb compressed using the 5.4. ENCOUNTERED PROBLEMS 73 Gzip algorithm and approximately 120Gb when decompressed into XML files, dbSNP uses on a regular workstation with 8Gb RAM. The latter refers to the problems encountered associated to conceptual ambiguity or collision with the conceptual schema. Both will now be discussed separately. 5.4.1 Data quantity problems The dbSNP data repository provides access to its files through an open FTP server1, where the files are organized per organism and file type. The files can be downloaded as a SQL database dump, XML or ASN. The latter being a more space efficient representation of XML. For the purpose of this work the XML files are most appropriate. The already present solution for transforming data from outside sources to the CSHG format, which will be discussed in detail in chapter 7, is a Java application which makes it very suitable for integrating already existing XML parser libraries. The XML files are organized as one per chromosome, including the Mitochondrial chromosome, and a few summary files. All together they account for 22Gb in a compressed state, and approximately 120Gb when uncompressed. The algorithm used to compress them is Gzip. Processing these file sizes pose obvious performance issues when done on a regular workstation PC. The relevant specifications of the PC from which was worked are described in section 1.4. The file size for the individual files ranges from 56Mb compressed (591Mb uncompressed) in the case of the relatively small chromosome Y, to 1.7Gb compressed (18Gb uncompressed) for the largest chromosome on the human genome, chromosome 1. Thus simply opening one of these with a regular text editor - clearly a logical step in analysis of the filesposes serious difficulties, often ’freezing’ the system. Fortunately, Ubuntu provides powerful tools to handle this type of files directly from the command line. Each file has a header section with general data about dbSNP version, organism, taxonomy id etcetera. The absolute majority of the file is taken up by the dbSNP entries, each entered as a so-called Rs-item, identified by the <Rs>opening tag and </Rs>closing tag. In between, all data for that that entry is stored, ranging from the various submitted samples that lead to the establishment of that particular Rs entry to the validated details like position, and observed alleles. A sample section of a typical Rs entry can be consulted in appendix A.3. For more details on the exact structure of the XML file, please consult the XSD file as provided by dbSNP on their FTP server2. In general the elements that are of interest to the HGDB are contained within these earlier mentioned Rs entries. The Sequence element stores data about the observed alleles in the sequence (the Observed element), and provides flanking sequences -both upstream Seq5 and downstream Seq3-. Within each Rs entry there exist multiple Ss entries, that represent submitted samples which 1ftp://ftp.ncbi.nih.gov/snp/ 2ftp://ftp.ncbi.nih.gov/snp/specs/docsum_3.3.xsd 80 CHAPTER 5. SOLVING THE DBSNP CASE the philosophy of this project to only store data that is valid and consistent, ultimately leading to the reliable data repository that is so much needed in the domain right now. As a suggestion for improvement to the dbSNP database -beside the obvious name change from dbSNP to dbVariation to account for the conceptual ambiguity earlier mentioned-, the authors of this work suggest the addition of an extra element in the dbSNP XML files. This element should identify the type of each SNP; either being an insertion, deletion, indel or perhaps a more elaborate classification. An adequate, and equally obvious name for this element would be Type, and could be located within the Component element already present. This latter allows for expressiveness to cope with the same SNP having perhaps different observed behavior in different reference sequences. Chapter 6 Solving the BIC case The Breast Cancer Information Core (BIC)1[30] is an open access, on-line mutation database for breast cancer susceptibility genes. In addition to creating a catalogue of all mutations and polymorphisms in breast cancer susceptibility genes, a principle aim of the BIC is to facilitate the detection and characterization of these genes by providing technical support in the form of mutation detection protocols, primer sequences, and reagent access. Breast Cancer is the most common malignancy among women worldwide. Of these cancers, approximately 5-10% is attributable to inheritance of a mutation in one or more highly penetrant autosomal dominant susceptibility genes. Both BRCA1 and BRCA2 have been identified as exactly this type of genes. It was evident that scientific and clinical interest in BRCA1 and BRCA2 would lead to an effort to screen women at high risk of developing breast cancer for mutations in these genes. Building a knowledge base about cancer susceptibility genes is one of the steps towards realizing this goal. The BIC represents an effort aimed at exactly this. As of April 2000 the repository contained 3416 entries describing genetic variations in BRCA1, and 2292 entries for BRCA2. Unfortunately the BIC data repository seems to have been inactive the last decade or so. Since the publication by C. Szabo in 2000 [30] it seems no new publications have been made. The data remains available but it is unclear when it was updated, or whether new data keeps being added. However, browsing through the data reveals that the latest date in which entries have been added is the 25th of November, 2004. The data set in itself represents a collection of curated tuples and is thus regarded by the community as reliable. An interesting, but as of yet unanswered question is whether the data set has been super seeded by other data sources like the Human Gene Mutation Database, or various LSDB’s. The data set as is now has been found to accumulate 12 025 entries for the BRCA1 gene, and 11 331 for the BRCA2 gene. Of these entries, 1446 and 1692 have been recognized as valid and unique for BRCA1 and BRCA2 respectively. 1http://research.nhgri.nih.gov/bic/ 81 82 CHAPTER 6. SOLVING THE BIC CASE 6.1 BIC description The BIC dataset can be viewed from the database website through a fee less membership account, the overall layout of the BIC web site is shown in figure 6.1. Browsing the data can be done per gene, and to even finer granularity per exon. The BRCA1 gene is divided into 25 exons, in which exon 11 is subdivided into 11A until 11D, and no exon 1 and 4 exist (the numbering starts with exon 2). The absence of an exon 4 has been discussed in detail in chapter 4, section 4.3.1. The data can be downloaded as a Tab Separated File (TSV) either per gene or per chromosome. The data files consist of 32 columns and store a variety of data about each variation. The relevant ones are as follows: (i) Accession Number, identifying the variation internally, (ii) NT, locating the variation nucleotide precise within the cDNA, (ii) Base Change, describing which change was observed. Other columns of interest include Reference, which in some cases provides a literature reference for the variation, and Clinically Important which tells us whether the observed change either led to a phenotype change or not. The latter is strongly related to the SNP concept earlier discussed thoroughly in chapter 5. for a full overview of the BIC data format, consult appendix A.4. Positioning in BIC is done in two ways, the most common one is done by aligning the variation with a transcript while in some cases the variation is located against genomic DNA. All variations have been aligned with the same -although different per gene, obviouslytranscript reference sequences. The reference sequences used by BIC are as follows: BRCA1 GenBank U146802 and BRCA2 GenBank U437463. Some variations have also been positioned according to their genomic DNA location, but this only occurs for the BRCA1 gene, and BRCA1 GenBank L788334is used. 6.2 Encountered problems The BIC data set turns out to be quite comprehensive and accessible when compared to the previous two cases, HGMD and dbSNP. The data are readily available, although one needs to create a membership account free of charge, and can be downloaded as a single file. Interpreting the data proved to be no problem because of the [30] publication which comprehensively clarifies how the data is structured, as well as assigning semantical meaning to each data property. The Breast Cancer Information Core website is quite sparse in providing information on how to interpret the data, and it does not provide a list of publications associated to the BIC. Searching the well known channels for publications led to the retrieval of the above document, but it is unclear whether this is the last sign of life of the BIC or whether more recent publications exist. A more interesting issue encountered in the BIC repository, and one that is prone to reoccur, is the positioning problem. As stated above, BIC variations are 2http://www.ncbi.nlm.nih.gov/nuccore/U14680 3http://www.ncbi.nlm.nih.gov/nuccore/U43746 4http://www.ncbi.nlm.nih.gov/nuccore/L78833 6.2. ENCOUNTERED PROBLEMS 83 Figure 6.1: The Breast Cancer Information Core (BIC) website features. Overview and map of the BIC website. Details of the links contained in the Mutation Database section are shown. 84 CHAPTER 6. SOLVING THE BIC CASE positioned along well established reference sequences. These sequences however, are quite old and for this reason outdated. The BRCA1 GenBank U14680 came to be as a result of combining the efforts by Miki [92] and Futreal [102] in 1994, the latest update as recorded by the NCBI dates back to the 10th of June 2002. The BRCA2 GenBank U43746 as first presented by Teng [103] in 1996 has not seen any modifications since then. It is not uncommon for genomic data repositories to refer to outdated reference sequences; sometimes the effort of realigning ’old’ data to new reference sequences is to costly or simply impossible. The challenge in the case of BIC consists of ’saving’ these data by realigning them with the currently used reference sequence, see section 1.4 for more details on which reference sequences have been used in the creation of HGDB. This work reports on the process of manually aligning the sequences and deducing the relevant corrections values from it. Due to time constraints no automated solution can be presented but creating this would offer interesting future research lines. In the following section the procedure that was followed to retrieve the BIC datasets’ positions will be elaborated on. 6.2.1 Aligning reference sequences Aligning genetic sequences has been subject of research for a long time. The de-facto standard has become an algorithm known as BLAST as presented in 1990 by [104], which stands for Basic Local Alignment Search Tool. Nowadays, various versions of BLAST exist, some refined for a specific purpose while other provide performance enhancements. Another popular algorithm is the BLAT [105], the BLAST-Like Alignment Tool. All of these tools provide an algorithm capable of comparing primary biological sequence information, such as the amino acid sequences of different proteins or the nucleotides of DNA sequences. A BLAST search allows a researcher to compare a sequence against a library or database of sequences , and identify library sequences that resemble the query sequence above a certain threshold. Because of the relatively uncomplicated nature of the required alignment, the conventional BLAST algorithm as provided by the NCBI was deemed sufficient. The NCBI implementation of BLAST5. For the purpose of this work not every observed change between the sequences is relevant. Structural differences that do not reflect a differing number of nucleotides between sequences can safely be ignored. Indeed, these latter differences do not change the positioning reference and are thus in the context of this particular application irrelevant. The differences that do represent a change between the number of nucleotides of both sequences, and thereby warp the positions accordingly need to be noted and taken into account when correcting for them. An example of the BLAST result file is shown in figure 6.2. The result of analyzing the BLAST result file lead to the identification of a series of structural differences, both in the BRCA1 and BRCA2 genes. Currently performed manually, there is clear advantage in devising an automated approach. These structural changes are then taken into 5http://blast.ncbi.nlm.nih.gov/Blast.cgi 6.3. LESSONS LEARNED 85 account when creating a new, ’virtual’, set of joins that represent the exon borders. These exon borders are required to calculate the absolute positions -genomic; referenced on a scale where both exons and introns are taken into accountof variations that are positioned in a relative style. The latter refers to positioning in which introns have been left out of the sequence, for a more detailed explanation refer to section 4.3.2 and appendix C.1. Figure 6.2: BLAST result file showing the alignment between GenBank U14680 and refSeq NG 005905. The green box identifies variations between the two sequences that did not affect structurally (length-wise) while the red box shows a structural difference between the two. In the latter case the query sequence, U14680, has a Cytosine base at a position where NG 009505 appears to have no nucleotide at all. 6.3 Lessons learned The BIC data set represents a clear added value to the HGDB. Although very outdated in some aspects, the data was once well curated and of high quality. The character of the data, being directly related to common malignancies among women, makes it of clear interest both in both clinical and research aspects. It would clearly be a waste to simply discard it, and this work has presented an effort to recovering it. The method used to recollect the BIC data entries, by aligning the outdated reference sequences to the current ones, is expected to be of future use in other sources as well. As the genomics domain rapidly advances, data sets become increasingly fast outdated. Sequences once considered the latest and a reference, can quickly be discarded from this status once progressing understanding proves them erroneous. As presented an obvious solution is to calculate the differences between those old sequences and their new counterparts, allowing the calculation of correction values. These corrections values can then be put to use in establishing the positions of the BIC variations. The results of resolving the problems encountered in BIC are displayed in figure 5.2. It shows an overview of the total number of variations per gene -of the 15 genes considered in this workand the number of entries finally loaded in the HGDB. Appendix B.4 provides an overview of how the amount of loaded variations relates to the total number of variations present in the repository for a given gene. In the case of BIC, the percentages are relatively low with an average of 13.48 % and a standard deviation of 1.45 %. This is due to the high amount of double entries as discussed earlier. Indeed of the 12 025 and 11 331 86 CHAPTER 6. SOLVING THE BIC CASE variations for the BRCA1 and BRCA2 genes respectively only 1446 and 1692 are considered unique and valid. Figure 6.3: Total number of variations per gene, and the number of loaded variations per gene for the BIC data repository. Chapter 7 Genoma Data Loader The analysis, interpretation and loading of the above mentioned databases lead to the establishment of a variety of lessons learned about these data, and their typical problems. Considering the vast amount of genomic data, a manual approach to loading them into HGDB would be rather cumbersome and probably impossible within a reasonable time span. The solution was found in creating a custom Extraction Transformation and Load (ETL) framework, built in the Java programming language. ETL applications have been used in many different domains to integrate data from distinct data sources and integrating them in a single repository, however all current solutions had to be discarded for two main reasons. No ETL solution was found to provide the level of flexibility required to cope with the wide variety of data sources (HTML, raw text, XML) as well as present enough capability to incorporate custom algorithms to change the data. These latter algorithms, although not always very complicated, ofter require the access to data already present in the database, or ’external’ files. These external files include cDNA sequences, that will after use not be loaded in the database. All in all, these latter requirements minimized the pool of possible ETL solutions to virtually none, however the applications that would maybe have offered the required functionality proved to exceed the budget in licensing fees. Once opted for the ’from scratch’ approach to build the ETL tool, the Java programming environment seemed the most appropriate tool to do it. Bioinformatics has traditionally favored Java for its multi-platform capacities and large amount of available libraries. For the purposes of this work, performance is subordinate to flexibility and availability. Indeed, it doesn’t matter that much whether the database loading takes 1 hour or a whole day, as the actual process of introducing new databases and maintaining the already considered up to date is much more time consuming. The wide arrange of libraries available for Java made it a clear alternative, as this would minimize the need for reinventing the wheel in many cases. At the same time, as the project was currently set up to use Windows, Linux and Mac OS X firmly intertwined, it was reasoned that a future solution should in theory be capable of running on all these platforms, 87 88 CHAPTER 7. GENOMA DATA LOADER as well as being developed on the various platforms. The constant evolution of the Conceptual Schema, and thus the resulting database ´and the ETL solution, meant that for a useful solution the application needed to cope well with change. It was decided that a plugin-based solution would be most adequate, in which a framework would take care of the regular execution flow of the application, while plugin based execution blocks would allow for a precise definition of how each external data repository should be treated separately to transform its data to the appropriate format. Looking at the future, in which not only the input side of the application would change but also the storage side, and thus a ”plugout” concept was introduced. Currently the only plugout that has been implemented is the Oracle database output interface, which allows for loading of the transformed tuples to the Oracle database. It is not difficult to imagine that various other plugouts might be created, like one that makes a flat file dump or an XML output. The use of plugins and plugouts enables the application to quickly evolve along with changes to existing data sources, as the required modifications to programming code are confined to as few classes as possible. At the same time, adding new data sources becomes a quick and easy process in which previous lessons can be taken advantage from as much as possible and requires a minimum amount of code creation; in the best case scenario only two classes need to be created, with relatively little programming complexity, to load a new data repository. Saying that the Genoma Data Loader has known issues, yet to be resolved, can be considered an understatement. There are many known, and probably even more still unknown. This, however, should not be considered the result of a lack interest or effort as invested by both the author of this work, and many others. The GDL represents a step in the ambitious endeavor at unifying genomic data. In a domain where data is considered the main currency -characterized by complexity, large quantities and still surrounded by so much that is left unknownthis unification is exactly what is craved for. In this section we will introduce the successes of the GDL in section 7.2.. 7.1 Structure Roughly the Genoma Data Loader can be decomposed in three layers: extraction, transformation and loading. Figure 7.1 provides a conceptual schema of the main structural components of the Genoma Data Loader framework. This corresponds to the well established concept of so called ETL (Extract, Transform, Load) applications. The idea of the Genoma Data Loader is to provide a framework environment that handles the execution flow while at the same time providing great flexibility and efficiency in adding additional functionality. The latter is achieved by the use of a plugin based approach, in which so called AGetter plugins are used to obtain raw data from external sources. The ATransformers plugins are used to transform the raw data into their destination format, this includes semantic as well as syntactic changes. The ALoader plugins then define classes that perform the loading procedure of the prepared 7.1. STRUCTURE 89 data into the Human Genome Database. The high level plugin is defined by the APlugin abstract class. The Tool object is used as a wrapper around the actual plugins, instantiated by the Main class as it is executed and the command line arguments are parsed. It deserves mentioning here that the framework can be considered a pipeline where the Getters generate output -in the form of ’raw’ data fileswhich then serve as input for the Transformers, whose output than -in the form of organized and ’clean’ data filesserves as input for the Loaders. AGetter ATransformer ALoader FtpGetterHgmdGetter TSVMutationTransformer APlugin XMLMutationTransformer AMutationTransformer MutationLoader Tool -is wrapped by0..1 -wraps1 AObjectCreator ARawMutation AXmlCreatorATsvCreator genoma.data .loader * 1 * -instantiates * APlugOut SqlOutManager FileOutManager 1 1 Figure 7.1: A conceptual schema that represents the structure of the Genoma Data Loader program. Currently two Getters have been implemented: HgmdGetter and FtpGetter. 96 CHAPTER 8. CONCLUSIONS can be seen as the crystallization of a certain body of knowledge. Exactly as is being attempted by Gene Ontology -creating a tool for the unification of biologyconceptual schemas might very well fulfill the promise. As has been argued in this work and previously by Smith and Kumar [79] [80], among others, the current de-facto standard leaves a lot to wish for. We firmly believe that exactly conceptual modeling techniques can bring that what is needed to make the promise of the Gene Ontology come true: by offering a visual tool, along with improved and formally described expressiveness a powerful reasoning tool can be acquired. One of our goals in this work was to start a discussion on the application of conceptual modeling techniques in genomics. Capturing biological understanding is a huge challenge we face, and the question is not whether we want to do it, but how do we do it. Specifications of conceptualizations are everywhere around us, and all of these correspond to the broad concept of an ontology. It is when we try and structure knowledge into a computationally sound structure and consider this to be the ontology, where not everything can stand that qualification anymore. Rigorous methods are needed, for a semantically sound ontology to emerge. The practice of conceptual modeling has been long around in Information System development, and has proven very suitable to capture domain knowledge in a controlled manner, both syntactically and semantically sound; ultimately even leading to the automated generation and validation of computer software. If we take the liberty of considering the biological system of life as analogous to an Information System it is not easy to miss the grand opportunities such a perspective provides. The analogy here is tempting and exciting; if we can create higher quality Information Systems by fully understanding their respective problem domains, perhaps we can improve life itself too by applying that very same method. After all, the mechanisms that underly the processes that drive life are not that different from our artificial, man-made Information Systems. Indeed, it requires only a small amount of imagination to see how information contained on long-term storage in DNA transfers to short-term storage RNA and is eventually ’executed’ in an intricate chemical processing unit to form an organism. Chapter 9 Future work One of the first future research lines that comes to mind is evolution of conceptual data modeling to cope with immense data quantities. It is clear that in bioinformatics very large amounts of data are being managed, and that this is very likely to increase exponentially the coming decades as technology further develops. The current MDA techniques allow for the automated creation of conventional relational databases like MySQL, Oracle and PostgreSQL. The rules for transforming platform independent models (conceptual schemas) to implementations that follow this relational paradigm have been well established. These technologies, although very mature and performing exceptional in some applications, date back to the ’70s. The current information age has changed a lot since then; data quantities have increased, hardware capabilities have changed, as well as user requirements, imposing entirely new challenges on storage technology. Current popular solutions include the so-called NoSQL-based solutions [108], in which conventional relational algebra is replaced with different paradigms that allow for far larger data quantities, faster access times while in some cases trading in some consistency in real time applications [109]. Currently NoSQL solutions are put to practice by large multinationals like Facebook [110], Google [111], Amazon and Twitter. Although often referred to as NoSQL, we prefer the term non-relational databases as it covers the semantics better. It would be a very interesting line of research to see if and how conceptual modeling can be applied to the application to these new technologies. In particular, discovering the set of rules and transformations required to translate a conventional conceptual schema (platform independent model) to a non-relational database, thereby combining the reduced development effort of model driven engineering with the ability to cope with today’s high data quantity requirements. Next, an interesting thought is the classification of different classes of loading problems. The Genoma Data Loader program is able to identify certain types of problems, like inconsistencies among data sources and format errors. The latter referring to for instance, typing errors in which a letter is introduced in the ’location’ attribute of a mutation. Classifying these problems into separate categories and determining how runtime encountered problems can be assigned 97 98 CHAPTER 9. FUTURE WORK to them is expected to spark well directed research lines. By applying statistical methods it is then possible to determine ’out of the ordinary’ data sets within data sources, and among them. To illustrate we will take the example of the HGMD, here we might assume that the average format error -for instance a letter in an integer columnaffects around 1 in a 100 mutations, per gene. We now know, through applying statistics, if a certain average format error for a newly loaded gene is significantly higher then the average overall that something might be off; either the Genoma Data Loader code does not cope well with this particular gene, or this is a new kind of formatting not yet implemented. Through this work we have used a gene-centric approach. The conceptual schema takes the concept of a gene as a central hub, an connects additional properties to this main artifact. Reality might be slightly different as will be illustrated with an example. Currently all variations in the Human Genome Database are associated to genes and their respective result on phenotype. The idea here is that if a gene is the atomic element of heredity, only those variations that affect a gene will have an effect on phenotype. This however turns out not be the case, as is now commonly believed among geneticists that also variations in non-coding parts of the genome can affect phenotype. This latter renders, at least to some extend, the currently chosen perspective invalid. A better approach, and one that is being investigated right now, is to use the chromosome centric approach. This allows for connecting variations to chromosomes, rather then genes. This is in turn makes full sense, as we take into account that the earlier gene centric solution was based on the shortcomings of technology at that time, being unable to sequence entire genomes at reasonable prices. Now that we possess over the short term prospect of acquiring cheap full genome sequencing, these DNA stretches between the genes will be analyzed in much more detail. The current evolution of the conceptual model, which took place in parallel with this work, is already exploring these ideas. One last interesting thought deals with data redundancy. It is currently unknown to what degree common genomic data sources like the BIC, LOVD and HGMD overlap. They each store very similar data, but in varying data formats that make a comparison non-trivial. The result of this work, basically the integration of these data in a common format, has made this comparison a very feasible research line. It would be particularly interesting to see how the entries in the relatively old and abandoned BIC made it, or not, to the other data sources. The difficulty will be to determine what are the basic properties of a variation that determine its identity. The authors of this work suggest that the actual change and position are all that is needed. Some discussion can exist about whether the phenotype should also be added. After all, it is still very unknown if one variation can have just one, or various phenotypes. And if a variation leads to various phenotypes, it must mean that other genetic changes also affect the process; shouldn’t they be part of the concept of Variation too, if that is what we want to indicate the genotypic blocks that lead to different phenotypes? Indeed, following this line of thought the Variation concept should not be restricted to a single change when it is connected to the Phenotype concept, but rather a new concept in between. This new concept could be 99 called Haplotype, following the genetic concept of variations being transmitted together. 100 CHAPTER 9. FUTURE WORK Acknowledgements The contents of this work are a result of the two and half years I have had the pleasure to work as a junior researcher at the Universidad Polit´ecnica de Valencia, Spain. Working together with so many wonderful people, and being able to reside in such a great environment -both from an academic ´and a general point of viewis what sparked my interest for the genomics domain, where so much is still left to be done. Naturally, a work of this size comes to be as a result of the collaboration of many and I take mere credits, if any, for assembling those lessons learned from so many into a single body of work. In the following section I will proceed to thank and credit those of you who have influenced me for the better, and stood by me when times were rough. Naturally, although all merits of this work ought to be shared with those mentioned here, I take sole responsibility for any errors you as a reader might discover. Where would we be, without our parents? Simply nowhere. This holds true from both an existential point of view for we would not exist, and a moral perspective because without them, we would not know how to act. Existential when we talk about genetics and the messages conveyed by it -some good and some maybe not so goodand moral when we refer to who we are as a human being and how we treat others. Parenthood only deals with the first, you see, while good parenthood is associated with both. In the end, we need our parents to shape us while later we need them for guidance and support, as much as we need them for the petty act of existence. It is for this reason, that although perhaps not much of the work presented here has been influenced directly by my parents, the mere existence is fully indebted to them. For this, and everything else I thank both of them. I, on the other hand, feel that this work represents my humblest effort of contributing to science in such a manner that maybe one day problems we face today, like cancers, may be overcome. Further, I want to pay explicit gratitude to my mentor and good friend ´ Oscar Pastor for creating so many opportunities and accepting me as one of his students. It has been an extremely educational and fun experience to work so closely with one of the founders of the Model Driven Architecture paradigm. Although -or perhaps becausewe disagreed on many occasions it has been trough the resulting discussions that new insights were born, many of which made it to this work. And isn’t that what science is all about? Special thanks also go out to Ana Levin, who I think of as the biology conscience of my work, as well as a very close friend. If it weren’t for her, I would probably not have been 101 102 CHAPTER 9. FUTURE WORK where I am now. Apart from monitoring my scientific development, correcting and teaching me when necessary she was also there to support me whenever necessary. My thanks go out to the entire ProS institute, with all it’s great members for providing such a pleasant environment to work in and welcoming a stranger with so much warmth. Special mentioning deserve Ainoha Martin Mayordomo for her contribution to the analysis of the BIC data repository, Veronica Burriel Coll for sharing her results of analyzing the LOVD data repositories and Arthur Baars for enlightening me with his great wealth of knowledge on Information Systems design and development. I also want to thank the full Genoma team, apart from the members already mentioned, for sharing so much valuable knowledge in a variety of domains. Bibliography [1] The International HapMap Consortium. “The International HapMap Project”. In: Nature 426.6968 (2003), pp. 789–796. [2] M.B. Gerstein et al. “What is a gene, post-ENCODE?” In: Genome Research 17.6 (2007), pp. 669–681. [3] H. Pearson. “Genetics: What is a gene?” In: Nature 441.7092 (2006), pp. 298–402. [4] M. Ashburner, C.A. Ball, and J.A. Blake. “Gene Ontology: Tool for the Unification of Biology”. In: Nature genetics 25.1 (2000), pp. 25–30. [5] E.J. Chikofsky and J.I. Cross. “Reverse Engineering and Design Recovery: a Taxonomy”. In: IEEE software 7.1 (1990), pp. 13–18. [6] G. Canfora and M. Di Penta. “Frontiers of reverse engineering: a conceptual model”. In: Proceedings of Frontiers of Software Maintenance (FoSM 2008). 2008, pp. 38–47. [7] O. Pastor et al. “The OO-method Approach for Information Systems Modeling: From Object-Oriented Conceptual Modeling to Automated Programming”. In: Information Systems 26.7 (2001), pp. 507–534. [8] B. Alberts et al. Essential Cell Biology. Ed. by E. Zayatz and E. Lawrence. 2nd. Garland Science USA, 2003. [9] J.C. Venter et al. “The sequence of the human genome”. In: Science 291 (2001), pp. 1304–1351. [10] E.S. Lander and et al. “Initial sequencing and analysis of the human genome”. In: Nature 409 (2001), pp. 860–921. [11] J.S. Mattick. “Challenging the dogma: the hidden layer of non-proteincoding RNAs in complex organisms”. In: Bioessays 25.10 (2003), pp. 930– 939. issn: 1521-1878. [12] J.S. Mattick. “RNA regulation: a new genetics?” In: Nature Reviews Genetics 5.4 (2004), pp. 316–323. issn: 1471-0056. [13] S.E. Ahnert, T. Fink, and A. Zinovyev. “How much non-coding DNA do eukaryotes require?” In: Journal of theoretical biology 252.4 (2008), pp. 587–592. issn: 0022-5193. 103 104 BIBLIOGRAPHY [14] E. Baralis and A. Fiori. “Exploring Heterogeneous Biological Data Sources”. In: 19th International Conference on Database and Expert Systems Applications. 2008, pp. 647–651. [15] H. Macauley, N. Wang, and N. Goodman. “A model system for studying the integration of molecular biology databases”. In: Bioinformatics 14 (1998), pp. 575–585. [16] M. van der Kroon et al. “Mutational data loading routines for human genome databases: the BRCA1 case”. In: Journal of Computing Science and Engineering 4.4 (2010), pp. 291–312. [17] S.T Sherry et al. “dbSNP: the NCBI database of genetic variation”. In: Nucleic acids research 29.1 (2001), p. 308. [18] K.D. Pruit and D.R. Maglott. “RefSeq and LocusLink: NCBI genecentered resources”. In: Nucleic Acids Research 29.1 (2001), pp. 137– 140. [19] T. Tatusova. “Methods in Molecular Biology”. In: ed. by C. Oliviero and F. Eisenhaber. Vol. 609. Biomedical and Life Sciences. Springer-Verlag, 2010. Chap. 2 - Genomic databases and resources at the National Center for Biotechnology Information, pp. 17–44. [20] T. Hubbard et al. “The Ensembl genome database project”. In: Nucleic Acids Research 30.1 (2002), pp. 38–41. [21] A. Hamosh et al. “Online Mendelian Inheritance in Man (OMIM)”. In: Human Mutation 15 (2000), pp. 57–61. [22] J. Amberger et al. “McKusick’s Online Mendelian Inheritance in Man (OMIM (R))”. In: Nucleic Acids Research (2008). [23] G. Joshi-Tope et al. “Reactome: a knowledgebase of biological pathways”. In: Nucleic Acids Research 33.issue suppl. 1 (2005), pp. D428–D432. [24] Lisa Matthews et al. “Reactome knowledgebase of human biological pathways and processes”. In: Nucleic Acids Research 37.suppl 1 (2009), pp. D619–D622. [25] M. Kanehisa and S. Goto. “KEGG: Kyoto Encyclopedia of Genes and Genomes”. In: Nucleic Acids Research 28.1 (2000), pp. 27–30. [26] M. Kanehisa et al. “From genomics to chemical genomics: new developments in KEGG”. In: Nucleic Acids Research 34.issue suppl. 1 (2006), pp. D354–D357. [27] M. Kanehisa et al. “KEGG for representation and analysis of molecular networks involving diseases and drugs”. In: Nucleic Acids Research 38.issue suppl. 1 (2010), pp. D355–D360. [28] P.D. Stenson et al. “The Human Gene Mutation Database (HGMD R ): 2003 update”. In: Human Mutation 21.6 (2003), pp. 577–581. [29] S. Bamford et al. “The COSMIC (Catalogue of Somatic Mutations in Cancer) database and website”. In: British journal of cancer 91.2 (2004), pp. 355–358. BIBLIOGRAPHY 105 [30] C. Szabo et al. “The Breast Cancer Information Core: Database Design, Structure, and Scope”. In: Human Mutation 16 (2000), pp. 123–131. [31] I.F.A.C. Fokkema, J.T. den Dunnen, and P.E.M. Taschner. “LOVD: Easy creation of a locus-specific sequence variation database using an ”LSDBin-a-box” approach”. In: Human mutation 26.2 (2005), pp. 63–68. [32] Amos Bairoch et al. “The Universal Protein Resource (UniProt)”. In: Nucleic Acids Research 33.suppl 1 (2005), pp. D154–D159. [33] Sarah Hunter et al. “InterPro: the integrative protein signature database”. In: Nucleic Acids Research 37.suppl 1 (2009), pp. D211–D215. [34] Damian Smedley et al. “BioMart - biological queries made easy”. In: BMC Genomics 10.1 (2009), p. 22. [35] N.W. Paton et al. “Conceptual modeling of genomic information”. In: Bioinformatics 16.6 (2000), pp. 548–557. [36] E.W. Bornberg-Bauer and N.W. Paton. “Conceptual Data Modeling for Bioinformatics”. In: Briefings in bioinformatics 3.2 (2002), pp. 166–188. [37] O. Pastor et al. “A Conceptual Modeling Approach to Improve Human Genome Understanding”. In: ed. by D.W. Embley and B. Thalheim. 1st. Handbook of Conceptual Modeling. Springer-Verlag, 2011. Chap. 16, p. 25. [38] F. Christen. “Global Sequencing: A Review of Current Molecular Data and New Methods Available to Assess Microbial Diversity”. In: Microbes and Environments 23.4 (2008), pp. 253–268. [39] O. Pastor and J.C. Molina. Model-driven architecture in practice: a software production environment based on conceptual modeling. SpringerVerlag. Berlin-Heidelberg, 2007. [40] D. Batra, J.A. Hoffler, and Bostrom R.P. “Comparing Representation with Relational and EER models”. In: Communications of the ACM 33.2 (1990), pp. 126–139. [41] D. Batra and G.M. Marakas. “Conceptual Data Modeling in Theory and Practice”. In: European Journal of Information Systems 4.3 (1995), pp. 185–194. [42] J. Larkin and H. Simon. “When a Diagram is (Sometimes) Worth Ten Thousand Words”. In: Cognitive Science 11.1 (1987), pp. 65–100. [43] B. Hailpern and P. Tarr. “Model Driven Development: The Good, the Bad, and the Ugly”. In: IBM Systems Journal 45.3 (2006), pp. 451–462. [44] H. Topi and V. Ramesh. “Human Factors Research on Data Modeling: A Review of Prior Research, An Extended Framework and Future Research Directions”. In: Journal of Database Management 13.2 (2002), pp. 3–19. [45] I. Davies et al. “How Do Practitioners Use Conceptual Modeling in Practice”. In: Data and Knowledge Engineering 58 (2006), pp. 358–380. Appendix A Data format samples A.1 HGMD precise mutations format sample This screenshot provides a typical example of how HGMD precise mutations are presented to the user. In this particular case we see the first 11 missense/nonsense mutations of the BRCA1 gene. The other precise mutations -small deletions, small indels and small insertionshave a very similar layout. The presentation is consistent among genes. 113 114 APPENDIX A. DATA FORMAT SAMPLES A.2 HGMD imprecise mutations format sample This screenshot provides a typical example of how HGMD presents imprecise mutations to the user. In this particular case the first few entries of the gross deletions overview of the BRCA1 gene are shown. The other imprecise mutations -gross insertions, gross indels, regulatory and complex rearrangementshave a similar layout. A.3 dbSNP XML format sample A sample of the XML file used by dbSNP to transfer data. What is shown is a single Rs entry, which represents a single Single Nucleotide Polymorphism. The data of this particular entry has been manipulated in order to fit on the page, this includes semantical changes. The entry as shown on this page therefore does not represent a correct dbSNP entry. <Rs rsId="3894" snpClass="snp" snpType="notwithdrawn" molType="genomic" genotype="true" bitField="050000000001010500020100" taxId="9606"> <Het type="est" value="0" stdError="0"/> <Validation byCluster="true" byOtherPop="true"> <otherPopBatchId>7179</otherPopBatchId> </Validation> <Create build="36" date="2000-09-19 17:02"/> <Update build="132" date="2011-01-11 12:25"/> <Sequence exemplarSs="3931" ancestralAllele="C"> A.3. DBSNP XML FORMAT SAMPLE 115 <Seq5>TAATCAGTCTCCTCCCAGCAAGTGATAT</Seq5> <Observed>C/T</Observed> <Seq3>AATTAGGAAGAGCTGGTACCTAAAAT</Seq3> </Sequence> <Ss ssId="3931" handle="OEFNER" batchId="489" locSnpId="M3" subSnpClass="snp" orient="forward" strand="bottom" molType="genomic" buildId="36" methodClass="DHPLC" validated="by-cluster"> <Sequence> <Seq5>TAATCAGTCTCCTCCCAGCAAGTGATATG</Seq5> <Observed>C/T</Observed> <Seq3>AATTAGGAAGAGCTGGTACCTAAAATGATTTT</Seq3> </Sequence> </Ss> <Assembly dbSnpBuild="132" genomeBuild="37_1" groupLabel="GRCh37" current="true" reference="true"> <Component componentType="contig" ctgId="224514809" accession="NT_011875.12" name="NT_011875.12" chromosome="Y" start="13798578" end="23901427" orientation="fwd" gi="224514809" groupTerm="GRCh37" contigLabel="GRCh37"> <MapLoc asnFrom="5297784" asnTo="5297784" locType="exact" alnQuality="1" orient="reverse" physMapInt="19096362" leftFlankNeighborPos="179" rightFlankNeighborPos="181" leftContigNeighborPos="5297783" rightContigNeighborPos="5297785" numberOfMismatches="1" numberOfDeletions="0" numberOfInsertions="0"/> </Component> <SnpStat mapWeight="unique-in-contig" chromCount="1" placedContigCount="1" unplacedContigCount="0" seqlocCount="1" hapCount="0"/> </Assembly> <RsLinkout resourceId="4" linkValue="60936"/> <MergeHistory rsId="56546490" buildId="130"/> <hgvs>NT_011875.12:g.5297785G>A</hgvs> </Rs> 116 APPENDIX A. DATA FORMAT SAMPLES A.4 BIC data format sample Accession Number Unique identifier generated at the time mutation is entered is entered into database Exon BRCA1 or BRCA2 exon in which mutation has been identified cDNA nucleotide Nucleotide # in the transcript at which mutation occurs; Reference sequences: BRCA1 GenBank U14680; BRCA2 GenBank U43746 gDNA nucleotide Nucleotide # in genomic DNA at which mutation occurs; Reference sequences: BRCA1 GenBank L78833 Codon Triplet codon # (ATG is +1) in which mutation occurs Base change Description of nucleotide difference compared to reference sequence Amino acid change Description of resulting change in the encoded amino acid sequence Mutation designation Designation of described mutation according to nomenclature guidelines Mutation type Frameshift, nonsense, missense, splice, in frame deletion Mutation effect Frameshift, nonsense, missense, splice, unclassified variant, polymorphism Depositor Contributor of mutation report Patient sample source Hospital or research institute from which patient sample originated Patient ID number Patient identifier designation in publications, or used by the hospital or research institute Number of times reported Number of family members carrying mutation Germline or Somatic mutation Detection method used SSCP, direct sequencing, DGGE etc. Proband tumor type Breast, ovarian, other Number of chromosomes screened Number of normal control chromosomes screened Frequency of polymorphism (A/C/T/G) Information on the frequency of variants in normal control chromosomes, if known Literature reference Literature citation in which mutation was first reported Contact person Email address of individual to whom inquiries should be addressed Notes Additional information describing the mutation, family history, patient series, age of onset, etc. Creation date Date on which mutation is entered into the public database Ethnicity Information on patient ethnicity, if known Nationality Information on patient nationality, if known Appendix B Genoma Data Loader coverage B.1 Variations per gene, per data repository Gene HGMD BIC dbSNP LOVD TOTAL BRCA1 1035 12 025 3091 0 16 151 BRCA2 776 11 331 2712 0 14 819 CDH23 86 0 4954 1119 6159 CLRN1 15 0 547 276 838 COL1A1 443 0 748 0 1191 COL1A2 288 0 787 0 1075 DFNB31 4 0 1289 132 1425 FBN1 848 0 2320 0 3168 GPR98 6 0 5409 41 5456 MYO7A 190 0 1205 2136 3531 NF1 1018 0 2558 0 3576 PCDH15 23 0 11 830 504 12 357 USH1C 16 0 647 223 886 USH1G 6 0 121 50 177 USH2A 119 0 7925 2212 10 256 TOTAL 4873 23 356 46 143 6693 81 065 117 118 APPENDIX B. GENOMA DATA LOADER COVERAGE B.2 Translated variations per gene and data repository Gene HGMD BIC dbSNP LOVD TOTAL BRCA1 1035 11 507 1662 0 14 204 BRCA2 776 11 169 998 0 12 943 CDH23 86 0 4754 1108 5948 CLRN1 15 0 565 276 856 COL1A1 443 0 809 0 1252 COL1A2 288 0 843 0 1131 DFNB31 4 0 1222 132 1358 FBN1 848 0 2399 0 3247 GPR98 6 0 5254 41 5301 MYO7A 190 0 1094 2090 3374 NF1 1018 0 2499 0 3517 PCDH15 23 0 12 186 489 12 698 USH1C 16 0 661 200 877 USH1G 6 0 111 50 167 USH2A 119 0 7965 2209 10 293 TOTAL 4873 22 676 43 022 6595 77 166 B.3. TRANSLATED UNIQUE VARIATIONS PER GENE AND DATA REPOSITORY119 B.3 Translated unique variations per gene and data repository Gene HGMD BIC dbSNP LOVD TOTAL BRCA1 947 1446 1140 0 3533 BRCA2 733 1692 684 0 3109 CDH23 74 0 3273 241 3588 CLRN1 15 0 370 16 401 COL1A1 358 0 575 0 933 COL1A2 244 0 595 0 839 DFNB31 4 0 827 54 885 FBN1 721 0 1638 0 2359 GPR98 6 0 3570 13 3589 MYO7A 170 0 734 299 1203 NF1 792 0 1747 0 2539 PCDH15 23 0 8208 91 8322 USH1C 12 0 455 54 521 USH1G 6 0 74 7 87 USH2A 107 0 5445 471 6023 TOTAL 4212 3138 29 335 1246 37 931 120 APPENDIX B. GENOMA DATA LOADER COVERAGE B.4 Translated unique variations versus total number of variations per gene and data repository Gene HGMD BIC dbSNP LOVD AVG S.DEV. BRCA1 91.50 % 12.02 % 36.88 % 46.80 % 33.19 % BRCA2 94.46 % 14.93 % 25.22 % 44.87 % 35.31 % CDH23 86.05 % 66.07 % 21.54 % 57.88 % 26.96 % CLRN1 100.00 % 67.64 % 5.80 % 57.81 % 39.08 % COL1A1 80.81 % 76.87 % 78.84 % 1.97 % COL1A2 84.72 % 75.60 % 80.16 % 4.56 % DFNB31 100.00 % 64.16 % 40.91 % 68.36 % 24.31 % FBN1 85.02 % 70.60 % 77.81 % 7.21 % GPR98 100.00 % 66.00 % 31.71 % 65.90 % 27.88 % MYO7A 89.47 % 60.91 % 14.00 % 54.79 % 31.11 % NF1 77.80 % 68.30 % 73.05 % 4.75 % PCDH15 100.00 % 69.38 % 18.06 % 62.48 % 33.81 % USH1C 75.00 % 70.32 % 24.22 % 56.51 % 22.92 % USH1G 100.00 % 61.16 % 14.00 % 58.39 % 35.16 % USH2A 89.92 % 68.71 % 21.29 % 59.97 28.69 % AVG 90.32 % 13.48 % 31.32 % 21.28 % 47.07 % S.DEV. 8.36 % 1.45 % 13.47 % 9.77 % 10.61 % 12.31 % Appendix C Algorithms C.1 Calculating absolute position public s tat ic int getAbsolutePos(List<Exon>exons , int cDNAPos) { int p o s i t i o n = 0 ; int l en gt h = 0 ; for (Exon e : exons ) { int exonLength = ( e . end exon −e . s t a r t e x o n ) + 1 ; i f ( exonLength + length <cDNAPos) { len gth += exonLength ; }else { p o s i t i o n = ( e . s t a r t e x o n + (cDNAPos − length −1) ) ; return p o s i t i o n ; } } return −1; } C.2 Separating reference from variation /∗∗ ∗in some cas es dbSNP ” h id es ” m u l t ip l e v a r i a t i o n s ∗in one <RS>entry ; t h i s can be i d e n t i f i e d by lo o k in g ∗at the ’ ’ observed ’ ’ f i e l d , which pr ovides observed ∗a l l e l e s s ep ar a te d by b a c k s l a s h e s (/ ) . Each o bse rv ed ∗a l l e l e t h a t doesn ’ t co rresp ond to t he n u c l e o t i d e a t ∗t h a t p o s i t i o n in r e fS eq means a unique v a r i a t i o n ∗according to CSHG. ∗ ∗@return ∗I f no observed a l l e l e corresponds with refSeq , ∗0 i s returned ∗/ private Li st <DbSnpMutation>parseObservedAlleles() { Li st <DbSnpMutation>variations = new ArrayList< DbSnpMutation>() ; 121