Literature Based Discovery Support System and its Application to Disease Gene Identification
Dimitar Hristovski1, Borut Peterlin2, Sašo Džeroski3, Janez Stare1
1IBMI, Medical Faculty; Vrazov trg 2/2, 1105 Ljubljana, Slovenia
2Department of Human Genetics, Clinical Center Ljubljana; Zaloska, 1000 Ljubljana, Slovenia
3Institute Jozef Stefan; Jamova 39, 1000 Ljubljana, Slovenia
Abstract. We present an interactive discovery support system, which for a given starting concept of interest, discovers new, potentially meaningful relations with other concepts that have not been published in the medical literature before. The known relations between the medical concepts come from the Medline bibliographic database and the UMLS. We use association rules mining for discovering the relationship between medical concepts. Then we demonstrate a successful application of the system for predicting a gene candidate for a disease, the gene of which has recently been identified via the positional cloning approach. We conclude that the discovery support system we developed, is a useful tool complementary to the already existing bioinformatic tools in the field of human genetics.
With the rapidly growing body of scientific knowledge and with the ever more present over specialisation, it is likely that the scientific work of one research group might solve an important problem that arises in the work of another group. Yet, the two groups might not be aware of the work of the other one. However, most of the scientific knowledge is recorded at least in a secondary form in various databases such as the bibliographic database Medline for the field of medicine. Also more and more important for the current medical research are becoming various molecular sequences and genetic databases. In the present situation, these vast databases provide an opportunity for developing advanced methods and tools for computer supported knowledge discovery.
The main points addressed by this paper are: 1. Is it possible to discover new, potentially meaningful relations (knowledge) between medical concepts by searching and analysing the documents from a bibliographic database such as Medline?, 2. To what degree can be the discovery process automated? And 3. Can this process be used for the task of candidate gene discovery for a human disease? As an attempt to deal with these issues we developed an interactive discovery support system based on association rule mining of the Medline bibliographic database. Its intended use is as a generator for research ideas that should be then investigated by traditional medical methods.
The idea of using a bibliographic database for generating new medical discoveries that should be later verified by traditional follow-up studies was proposed by Swanson . He managed to make seven medical discoveries just by searching the Medline database with some smart strategies and by analyzing the bibliographic records. These discoveries were later confirmed and published in relevant medical journals. Swanson's discovery support process is based on the concepts of complementary literatures and noninteractive literatures. If one set of articles (XY) reports an interesting relation between concepts X and Y, and a different set of articles (YZ) reports a relation between Y and Z, but nothing has been published concerning a possible link between X and Z, then XY and YZ are called complementary literatures. Generally, XY and YZ are complementary if a potentially new relation can be inferred by considering them together that can not be inferred from either of them separately. For example, X might be a disease, Y a physiological function associated with X and Z a substance or drug which induces or regulates the physiological function Y. If the readers and authors of one literature are not acquainted with the other, and vice versa, as might often be the case with two different specialties, then the two literatures are noninteractive. By combining the concepts of complementary and noninteractive literatures, Swanson developed the concept of undiscovered public knowledge meaning that although the literatures XY and YZ represent publicly available knowledge, the potentially new relation between X and Z remains undiscovered and is a valuable source of new discoveries.
The first published example of a discovery Swanson made was about Raynaud's disease and fish oil . Articles on Raynaud's disease (X) and articles on eicosapentaenoic acid (Z) when considered together indicated that dietary fish oil rich in eicosapentaenoic acid might be beneficial for treating Raynaud patients. One Y concept was, for example, blood viscosity. The line of reasoning used was: dietary eicosapentaenoic acid (Z) can decrease blood viscosity (Y), which has been reported in patients with Raynaud's disease (X).
Another notable discovery made by Swanson was about the relation between migraine (X) and magnesium deficiency (Z) . It was discovered that magnesium deficiency was the cause of certain physiological effects (Y), which were associated with migraine.
In the beginning, Swanson performed the discovery process manually by searching the Medline database. Later he added software support for some of the stages of the process. His current system is called ARROWSMITH and is described in detail in . The Swanson's discovery methodology contains two steps which are usually done sequentially, but each one can be done independantly as well.
The goal of the first step is, for a given starting concept X, to find potentially new relations to concepts Z which are unknown at the begining. The user starts by searching Medline for all the articles about a starting concept of interest (X). This can be done using any search software through which Medline is accessible and is not part of ARROWSMITH. The articles found are then uploaded into ARROWSMITH. Then the titles of the articles are analyzed and a list of all words and phrases is made. This list is, of course, very large and a stop word list is used to reduce it. Another way to alter the list is by manual user editing. The remaining words and phrases are considered concepts (Y) that are somehow related to the starting concept (X). Now a set of search strategies is generated in order to search Medline for each of the Y concepts. The search results are uploaded into the system and again a list of words and phrases appearing in the titles is produced. This is the set of concepts Z related to Y. Those Z concepts for which there are articles in Medline containing both X and Z are eliminated. The remaining concepts in Z represent possible candidates for new relations between X, Y and Z, where Y is some intermediary concept linking X to Z.
The goal of the second step of the Swanson's discovery methodology is, for given concepts X and Z, to find intermediate concepts Y through which X and Z are related. It is possible that more than one Y leads from X to Z and ARROWSMITH orders the Z concepts by decreasing number of Y connections. So, in ARROWSMITH, the frequency of words or phrases in article titles is used as a measure of relation between medical concepts. The publicly available version of ARROWSMITH support only the second step of the discovery methodology.
Some of the Swanson's discoveries were repeated with different methods by Gordon and Lindsay  and by Weeber . Weeber also discovered several hypothetical new therepeutical applications of existing drugs. For more details about the Swanson's approach and its comparison with the others, see Weeber chapter in this volume.
Our system is based on Swanson's ideas, but there are however, several notable differences between our approach and theirs. Instead of using title words as a representation of the Medline documents' meaning, we use the MeSH descriptors. We use association rules as a measure of relationship between medical concepts while Swanson uses word frequencies. We have built a large association rule base by pre-calculating and storing the association rules in a database management system. This allows us to build a truly interactive discovery support system with fast response.
The Medline database is a product of the US National Library of Medicine (NML). Because of its coverage and free accessibility, Medline is the most important bibliographic database in the field of biomedicine. It contains bibliographic citations and author abstracts from over 3,900 biomedical journals. Each citation is associated with a set of MeSH (Medical Subject Headings) terms that describe the content of the item (Figure 1). Presently the database comprises over 9 million records dating back to 1966 . Medline is available for free searching on many websites of government and health agencies. One of the most popular is the NLM Web based product PubMed. There are also about 80 commercial Medline products.
Fig. 1. An example of a Medline record. Only the title, authors, a part of the abstract and the MeSH terms (descriptors) fields are shown.
In our system, we use Medline as the source of the known relations between biomedical concepts. We extract these relations and store them in a kind of a knowledge base. The discovery algorithm then operates on this knowledge base as described latter.
MeSH comprises NLM's controlled vocabulary and thesaurus used for indexing articles and for searching MeSH-indexed databases, including Medline. It contains the biomedical subject headings (descriptors), subheadings, and supplementary chemical terms. MeSH terminology provides a consistent way to retrieve information that may use different terminology for the same concepts. MeSH organizes its descriptors in a hierarchical structure that permits searching at various levels of specificity from narrower to broader. This structure also provides an effective way for searchers to browse MeSH in order to find appropriate descriptors. A retrieval query is formed using MeSH terms to find items on a desired topic. Similarly, indexers always use the most specific descriptors available to describe the subject content of an article. Problems with MesH indexing may arise when all important terminology in a field is not covered: when using descriptors, new concepts may be worded in a way that nonexpert users can not readily identify. Problems may also arise because of inconsistency of human indexing. The current MeSH includes more than 19,000 descriptors, 110,000 Supplementary Concept Records (formerly Supplementary Chemical Records), and over 300,000 synonyms and related terms .
In our system, MeSH represents the set of biomedical concepts we are dealing with. In the first phase we extract known relations between these concepts and in the second we try to discovery new relations between them.
Providing improved access to search terms as well as databases has led to the initiation of the Unified Medical Language System (UMLS) project that NLM began in 1986. The UMLS project was undertaken with the goal of providing a mechanism for linking diverse medical vocabularies as well as sources of information because the proliferation of disparate vocabularies, none of which was compatible with any other was recognized a significant impediment to the development of integrated applications. The project develops "Knowledge Sources" that can be used by a wide variety of applications programs to overcome retrieval problems caused by differences in terminology and the scattering of relevant information across many databases. There are now tree components of UMLS Knowledge Sources: the Metathesaurus, Specialist Lexicon and Semantic Network [6,7].
The Metathesaurus provides a uniform, integrated distribution format from about 60 biomedical vocabularies and classifications used in patient records, administrative health data, bibliographic and full-text databases and expert systems.
The Lexicon contains syntactic information for many terms, component words, and English words, including verbs, that do not appear in the Metathesaurus.
The Semantic Network contains information about the types or categories (e.g., "Disease or Syndrome," "Virus") to which all concepts have been assigned and the permissible relationships among these types (e.g., "Virus" causes "Disease or Syndrome"). The Semantic Network, through its 132 semantic types, provides a consistent categorization of all concepts represented in the UMLS Metathesaurus.
UMLS research has made progress on some of the many research issues associated with interpretation of user queries, mapping between the language of different information sources, and medical natural indexing and retrieval techniques. Much of serious investigation and prototype system development involving links between automated patient data and knowledge-based information has been performed using UMLS components . NLM's own applications include Internet Grateful Med and PubMed.
In our system, we use from UMLS information regarding the semantic types of the biomedical concepts and their coocurence in Medline records.
Association rules  were originally developed with the purpose of market-basket analysis, where it is of interest to find patterns of the form X -> Y, with the intuitive meaning "baskets that contain X tend to contain Y". A basket corresponds to a single visit of a customer to a store and is called a transaction, while individual products in the basket are called items. The approach is general enough to apply to bibliographic databases, where transactions are documents and items are words or descriptors used for indexing the documents. Association rules here have the form Word1 -> Word2 or Descriptor1 -> Descriptor2 (e.g. Disease X (Multiple sclerosis) -> Symptom Y (Optic neuritis)). Another example could be: Disease (Multiple sclerosis) -> Treatment (Interferon-beta).
Goal and Basic Premises
The system we developed is an interactive discovery support system for the field of medicine and is supposed to be used as a generator of new, potentially meaningful relations between a starting known concept of interest and other concepts.
The Medline database is used heavily by biomedical researchers. Traditionally it is used to check what is new in the literature on a particular topic of interest or to check if a medical discovery has already been published. In the latter case the researcher has already made the discovery or at least has a general discovery idea and just wants to check if some one else has published that discovery before. Also the various information retrieval systems used for searching Medline are geared towards the task of searching for documents about a topic well known in advance. In contrast to the traditional use of Medline, where it is used to check if a discovery is new or not, our system actively helps in the discovery process by generating potentially new discoveries and research ideas by analyzing the Medline database.
When building our discovery support system we started with the same basic ideas as Swanson. However, the methods we use for discovering new relations between concepts are different and the levels of automation and interactivity of the discovery process are much higher in our approach.
We used the major MeSH descriptors assigned to a Medline record as a representation of the contents of the article the record is about. Some of the MeSH descriptors are designated as major (followed by an asterisk in the Medline record). Major descriptors are those that are the main topic of the article. See Figure 1 for an example of a Medline record and MeSH descriptors.
We used association rules  between pairs of medical concepts as a method to determine which concepts are related to a given starting concept. In our system an association rule of the form
means that in confidence percent of articles containing X, Y is present and that there are support number such articles. In other words, we take concept co-occurrence as an indication of a relation between concepts. If X is a disease, for example, then some possible relations might be: has-symptom, is-caused-by, is-treated-with-drug and so on. We do not try to find out the kind of relation currently. This can not be done directly by using the MeSH descriptors assigned to an article because there is no explicit information about the relation between the descriptors stored in the Medline record.
We calculated all the associations between the major MeSH descriptors. We did this regardless of the confidence and support values and for two Medline time segments: 1990-1995 and 1996-1999. The calculated associations are stored in a database management system: there are currently more than 11.000.000 associations in the rule base. The calculation of the association rules was much simplified by the use of the data contained in the UMLS, especially the co-occurrence files. Actually, for these calculations it was not necessary to access the full Medline records at all.
The large association rule base is a foundation upon which the algorithm for discovering new relations between concepts proceeds as described in Table 1. The main idea is to first find all the concepts Y related to the starting concept X (e.g. if X is a disease then Y can be pathological functions, symptoms, ...). Then all the concepts Z related to Y are found (e.g. if Y is a pathological function, Z can be a chemical regulating that function). As a last step we check if X and Z appear together in the medical literature. If they do not appear together then we have discovered a potentially new relation between X and Z. The user of the system should then evaluate the proposed (X, Z) pairs and select among them those that deserve further investigation.
Table 1. The algorithm for discovering new relations between medical concepts.
Because in Medline each X concept can be associated with many Y concepts, each of which can be associated to many Z concepts, the possible number of X -> Z combinations can be extremely large. In order to deal with this combinatorial problem, the algorithm incorporates filtering (limiting) and ordering capabilities. By filtering, the number of X -> Y or Y -> Z associations is limited and also the number of accidental associations is minimized.
The default filtering that can not be relaxed is that only the associations between major MeSH headings are considered by the system. The other filtering possibilities are optional and can be interactively enforced by the user of the system.
The related concepts could be limited by the semantic type to which they belong. Each MeSH descriptor belongs to one or more semantic types. For example, if the starting concept X is a disease (semantic type disease or syndrome) then the user can request that Y concepts are of semantic type pathologic function and that Z concepts are of semantic type pharmacologic substance. As a consequence, the system will only consider chains of associations of the form: disease or syndrome -> pathologic function -> pharmacologic substance. The information about the semantic types to which a concept belongs is drawn from the Semantic Network component of the UMLS. The last possibility for limiting the number of related concepts is by setting thresholds on the support and confidence measures of the association rules in steps 2. and 3. of the algorithm. In fact, all of the filtering options can be interactively set alone or several of them in combination.
Because the usefulness of the system relies to a large degree on human judgement, special attention is paid to the order in which the candidate related concepts are presented to the user of the system. Thus, the goal of the ordering is to present best candidates first to make human review as easy as possible. Currently the default ordering is by the decreasing association rule confidence, but it is also possible to order by support or semantic type.
Figure 2. shows the user interface of our discovery support system. Apart from it, the user also needs a standard web browser to access Medline through the Entrez system, which is publicly available. Through the Entrez system it is also possible to search the GenBank, SwissProt, OMIM and some other databases. Now we will describe the elements of the user interface, its use in the discovery process and its integration with the Entrez system.
^ The user interface of the interactive discovery support system.
Searching for a Starting Concept X
The user starts a discovery session by searching for a starting concept X, which is usually from his own research area. The query is performed with query by form method using the same screen form for browsing data and for searching. The user can specify a full or partial name in the Str field. When specifying a partial name, the % (percent) sign is a wild-card character and is a replacement for any character string, including an empty one. After entering the search criteria, the user presses the Query button to retrieve the results. The found starting concept is shown in the Starting Concept frame and the semantic types it belongs to are shown in the right frame.
When specifying a partial name, it is possible that more then one concept matches the search criteria. In that case, the system shows the first matching concept. However, by using the navigation buttons (<<, <, >,>>), it is possible to show the other concepts as well.
The fields above the concept name show the frequency of occurrence of the concept in Medline when used as a major MeSH descriptor for two time intervals (1990-1995) and (1996-1999).
The concepts Y related to the starting concept are found by pressing the Find Related button which is under the Starting Concept frame. Before finding related concepts, the user can specify limits such as the semantic type of the related concepts or the minimal confidence and frequency (support) of the association rules. This is done in the upper Limits frame. Also the order in which the related concepts are presented can be specified in the upper Order by frame. It can be descending or ascending by confidence or frequency (support). When the user wants to try a new limiting and ordering combination, he/she has to press the Find Related button again.
The related concepts are presented in the ^ frame. The Sab column shows the source of relationship (MBD - Medline 1990-1995, and MED - Medline 1996-1999). Apart from the related concept names, the frequency of co-occurrence in the particular segment is shown as well as the confidence of the association rule between the starting concept and the related concept. The user can browse through the list of related concepts and select those that need further investigation. This is done by checking the check box to the left of the related concept name.
Finding the Related Concepts Z
Now the user can press the Find Related button which is under the ^ frame. This action will find the Z concepts related to the Y concepts found in the previous step and show them in the RELATED CONCEPTS2 frame. As described earlier, the user can specify limits and the order of the related concepts.
The frame RELATED CONCEPTS2 contains an important additional field designated as "Discovery?". This value of this field is YES if a relation (association) between the starting concept X and the current concept Z does not exist in the appropriate Medline segment and NO if such a relation exists. In other words, this field shows if the relation between the starting concept X and the current concept Z is considered a potential discovery by the system. Of course, the judgement of a human expert (hopefully the user of the system) is needed to verify how plausible such a potential discovery is. The user can browse through the list of these potential discoveries or she can select different set of Y concepts and try again to find some potential discoveries. It is possible to limit the Z concepts to only those considered potential discoveries by checking the check-box Discoveries only in the lower Limits frame.
When one starting concept has been dealt with, another one can be searched for and the whole procedure can be repeated. The user can interactively guide the discovery process by selecting promising concepts and by setting various limits.
To make the evaluation of the proposed potential discoveries easier to the user, the system offer the possibility to search and display the Medline records related to the concept currently under investigation. The user should read these records, decide which relations deserve further attention and guide the discovery process accordingly.
The search and display of Medline records is run by pressing either the Show Medline docs (X and Y) or the Show Medline docs (Y and Z) buttons. The first one searches for Medline records containing both the starting concept X and the current concept Y. Similarly, the second button searches for records containing both the current Y and Z. When the user presses one of these buttons, the system does the discovery system does the following: 1. prepares an appropriate search request, 2. starts a standard web browser, 3. connects to the search system Entrez, and 4. runs the prepared search request. The user can then browse through the resulting records and change the search request if necessary. With some basic knowledge of the Entrez system, it is also possible to display the related proteins and nucleotide sequences.
The end user program is developed using Oracle Forms 4.5. The association rule base and the other necessary tables needed in the discovery process are stored in an Oracle relational database management system on a UNIX server. The end user program communicates with the database server in a client/server manner over a TCP/IP network, which can also be the Internet. To make the system widely available to end users we plan to develop a Web version that would require only a Web browser on the user side.
The ultimate goal of the system is to assist in making medical discoveries that can be published in relevant medical journals. So far, however, we have only managed to perform a preliminary evaluation.
Medical Meaning of Related Concepts
The main goal of the evaluation done by the medical doctor was to check if the Y concepts, which the system finds to be related to the starting concept X, have medical sense. In other words, we wanted to check if the association rules are successful in extracting known relations between biomedical concepts from the Medline database. This is very important in our approach because by combining known relations our system proposes potentialy new relations.
The medical doctor selected as a starting concept multiple sclerosis (MS) which could be defined as a demyelinating disease of the central nervous system of putative autoimmune origin. Then he used the system to find the related concepts Y. Below is a list of the first 20 concepts related (associated) with MS ordered by decreasing support with a short description about the nature of the association:
8.-20. Immunosupresives (treatment), IgG (diagnostics), encephalomyelitis (symptoms), cognition disorders (symptoms), VEP (diagnostics), citokines (pathogenesis), TNF (pathogenesis), spinal cord (anatomical structure – organ involved), methylprednison (treatment), receptors-Tcells( pathogenesis), myelin protein (pathogenesis), psychological adjustment (treatment), demyelinating disorders (disease categories).
Among the analysed 20 concepts there are 6 related to pathogenesis, 4 to treatment, 3 to diagnostic methods, 3 to symptoms, 2 to target organs-anatomical structures and 2 are related to general disease categories.
It could be concluded that the concepts found as related by the system are associated with the current main focus of medical endeavors in the field of MS which is still oriented to treatment and therefore to better understanding of pathogenesis.
For the Z concepts related to the Y concepts the evaluation becomes much more difficult and involves more or less speculations. Nevertheless two things deserve some commentary:
Table 2 - The results of the prediction of new relationships between medical concepts in the newer Medline segment (1996-1999) based on the older segment (1990-1995) using the system. The column names ending with 1 are for the AVGS constraint and those ending with 2 for the 2*AVGS constraint. The columns have the following meaning: n - all the relationships that can be predicted; k - new relationships in the newer segment that were not present in the older segment; m - predicted relationships based on the older segment; l - successfully predicted relationships; p - probability of achieving l or more successfully predicted relations by chance; r - the number of successfully predicted relations by chance alone.
The ultimate proof of the system would be to (help) discover medical discoveries that could be published in relevant medical journals. However, we have managed to do a statistical evaluation so far. The goal of this evaluation was to see how many of the potential discoveries made by the system at some point of time become realised at a later time. For us, a potential discovery is a relationship between two concepts proposed by our system, but not present in Medline at some point of time. We consider the potential discovery realised if the two concepts later appear together in a document in the Medline database. In other words, the goal of the evaluation was to see how good our system was in predicting what discoveries would be made in the future.
We approached this goal by first dividing the Medline database and the corresponding association rules into two segments according to the publication date of the documents stored: the older segment is from 1990 to 1995 and the newer segment is from 1996 to 1999. We then analysed ten diseases, which are listed in Table 2.
Here we will give a discussion of the analysis of Multiple sclerosis (MS). MS appears in 2582 documents in the older segment. It is related to 1610 distinct concepts. When analysing the old segment, the system proposed 15617 concepts as potential discoveries. MS is related to 662 new concepts in the new segment that it was not related to in the old segment. Our system successfully predicted 95.5% (632 out of 662) realised discoveries in the new segment. However, only 4% (632 out of 15617) of the proposed potential discoveries got realised. It should be stressed that MS was not related to 15965 out of 17575 distinct concepts appearing in the older segment. The system proposed 97.8% of the concepts MS was not yet related to as potential discoveries. The conclusion is that without using limits on the strength of relationship the system is very successful at predicting future discoveries, but proposes far too many potential discoveries. Then we repeated the evaluation with two values for thresholds on the support level of the association rules. In one case the threshold was set to the average support of the associations between one concept and the others (AVGS) and in the other case it was set to 2*AVGS. Only associations with support greater or equal to the threshold were taken into account. The number of proposed potential discoveries dropped from 15617 without thresholds to 6848 for AVGS and to 3151 for the 2*AVGS threshold. The percent of successfully predicted realised discoveries dropped from 95.5% (632 of 662) without thresholds to 78.7% (521 of 662) for AVGS and to 55.2% (366 of 662) for 2*AVGS. However, the ratio of realised to proposed potential discoveries improved from 4% (632 out of 15617) without thresholds to 7.6% (521 of 6848) for AVGS and to 11.6% (366 of 3151) for 2*AVGS. We conclude from this that with the use of proper thresholds the usability of the system is much better because a smaller number of better potential discoveries are generated.
Results of the statistical evaluation for the ten selected diseases are in Table 2. The values obtained by our system were tested against the null hypotheses of random hits. Or to put it another way, we wanted to check whether the number of correct predictions obtained by our system could have occured by chance alone. This is done in the following way.
Let n be the number of all possible relationships that the system could predict based on the older Medline segment. Of these k actually appear in the new segment (successful predictions), and n-k do not (nonsuccessful predictions). Let m be the number of actual predictions made by the system, of which l are successful, and m-l nonsuccessful. Probability of such an event is
This distribution is known as the hypergeometric distribution. The question now is if l successful predictions represent a statistically significant results. In other words, we need a probability of obtaining l or more successful predictions if we were predicting completely at random. This probability is given by
Table 2 shows p values for given n, m, k and l for AVGS and 2*AVGS contraints respectively. Zeros in the p value columns actually mean that the probability is less than 10-16. If the predictions were random, we would expect of them to be successful.
Fig. 3. Searching for the gene candidate for the ^ disease. The system has discovered a potentially new relationship between Incontinentia Pigmenti and the NF-kappa B transcription factor through the intermediate concept Interleukin-6.
Disease Candidate Gene (re)Identification
In the preceeding section we showed that our system predicts new relations between medical concepts with statistical significance better than some predicting by chance. This time we wanted to evaluate the system for the task of candidate gene discovery for a disease. For this purpose we selected the incontinenta pigmenti disease.
Familial incontinentia pigmenti (IP; MIM 308310) is a genodermatosis that segregates as an X-linked dominant disorder and is usually lethal prenatally in males. In affected females it causes highly variable abnormalities of the skin, hair, nails, teeth, eyes and central nervous system. The gene for IP has been maped in the terminal part of the long arm of the X chromosome (Xq28) .
The pathogenesis of the disease is not known yet, however immune and haematopoetic system seem to play an important role [10, 11, 12].
Recently the gene NEMO (NF-kappaB essential modulator/IKK) mutated in patients with IP has been identified via positional cloning approach . NEMO is required for activation of the NF-kappaB transcription factor, which is involved in numerous immune, inflammatory and apoptotic mechanisms.
The NEMO gene has been cloned in the 1998 and localised in the Xq28 region of the human chromosome X in 1999 .
More precisely, our aim regarding IP was to find out whether NEMO could have been identified by our discovery support system using the data that was available before NEMO was officially identified as the gene for IP.
As the starting concept (X) the name of the disease has been entered – Incontinentia pigmenti (Figure 3). Regarding the fact that the immune system seems to play an important role in the pathogenesis of IP, related concepts of the 1st order (Y) were limited by the semantic type immune factor. Two related concepts have been found: Interleukin-6 (IL-6) and Thromboplastin.
IL-6 is a cytokine and has an important role in the development of the inflammatory response, the differentiation and activation of cells of the hematopoietic lineage, and the regulation of nerve cell and bone cell functions. Increased serum levels of IL-6 has been namely found in an IP patient who in addition to IP also demonstrated symptoms of Behcet disease .
In the next step, we searched for concepts related to IL-6 (concepts Z). The column “Discovery?” shows if there might be a potentially new relation between the starting concept X and the current concept Z. The value in this field is YES if there are no Medline documents containing both the X and Z concepts. Among the hits with higher co-occurrence frequency we identified three hits pathogenetically potentially related to IP: NF-kappa B, apoptosis and vascular endothelium.
NF-kappa B is a transcription factor involved in metabolic pathways related to immune system, inflammation and apoptosis, and as such very interesting regarding IP.
In the next step we used the fact that the gene for IP is located at Xq28, which is known from 1994 . When some of the buttons labeled ^ is pressed, an external Web browser is started and a search request is executed on the PubMed database to show the concepts that are currently under consideration. Figure 4. shows the results of the search request for documents containing the transcription factor NF-kappa B and chromosome location Xq28. The first article  reports about the discovery of the NEMO gene as responsible for IP. However, this article was published in 2000. The second article  is very important in our procedure because it states that the NEMO gene, which is necessary for the activation of the transcription factor NF-kappa B, is localised in the Xq28 region. This article does not make any reference to IP and was published in 1999. From this we can conclude that by using our system together with some Medline searching, it was possible to predict the NEMO gene as responsible for IP in 1999.
Fig. 4. The Medline articles containing both the transcription factor NF-kappa B and the chromosomal region Xq28.
The paper presented an interactive discovery support system for the field of medicine. For a given starting medical concept it discovers new, potentially meaningful relations with other concepts that have not been published in the medical literature before. The proposed relations should be evaluated and verified by a qualified medical professional.
As a measure of the relation between concepts we use association rules calculated from the Medline bibliographic database. We were in a dilemma whether to use the X -> Y or Y -> X direction of the association rule when finding concepts related to X. Because we have only binary associations the direction comes into play only when limiting and ordering the related concepts. We selected the X -> Y direction with the intuitive idea that the Y concepts that appear most often in documents regarding X have strongest relation to X. However, there are cases where some Y might appear not very frequently in X documents, but X might appear in almost all of Y documents. In this case we have a strong association in the Y -> X direction that should be considered also. We plan to further investigate this issue. One possibility might be to develop a heuristic approach in which some times is used a X -> Y association and some times a Y -> X one. Another possibility is to use some kind of a composite measure that takes into account both directions of the associations.
We do not use Medline directly, but rather we use UMLS. It simplifies the calculations considerably, however it introduces considerable limitations into our system as well: we can calculate only binary association rules, the association rules are only between major MeSH headings and we are limited to only two time intervals. We plan to analyse the full Medline database in the future if we get access to it. Currently in MeSH only a few dozen genes and other sequences are present. We plan to calculate direct association rules between the MeSH headings and a considerable number of molecular biology sequences and so increasing the functionality of the system significantly. And the last in out plan list, but not the least is the development of a Web version of the system that will increase the number of users. More users means better chance for real medical discoveries.
As part of the evaluation of the system, a medical doctor confirmed that most of the relations between concepts found by association rules are meaningful. We demonstrated a successful application of the system for predicting the gene candidate for the incontinentia pigmenti disease, by using the information known prior to the discovery of the gene. In the statistical evaluation, the system proved to be successful at predicting future discoveries. However, this came at the expense of generating a large number of potential discoveries that have to be judged and verified by the user of the system. The statistical evaluation also showed that properly set thresholds are crucial for successful use of the system. Thus, we plan to work on setting good default values for the thresholds that can be changed by the user if necessary.
We believe that literature based discovery support systems, such as ours, will help researchers make some important biomedical discoveries in the future.