A CitationRank algorithm inheriting Google technology designed to highlight genes responsible for serious adverse drug reaction
1 Bio-X Center, Key Laboratory for the Genetics of Developmental and Neuropsychiatric Disorders (Ministry of Education), Shanghai Jiao Tong University, Shanghai 200030, 2 Institutes of Biomedical Sciences, Fudan University, 138 Yixueyuan Road, Shanghai, 200032, 3 College of Life Sciences, Nanjing Agricultural University, 1 Weigang, Nanjing, 210095 and 4 Institute for Nutritional Sciences, Shanghai Institutes of Biological Sciences, Chinese Academy of Sciences, 320 Yueyang Road, Shanghai, 200031, PR China
* To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT 1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
Motivation: Serious adverse drug reaction (SADR) is an urgent, world-wide problem. In the absence of any well-organized gene-oriented SADR information pool, a database should be constructed. Since the importance of a gene to a particular SADR cannot simply be defined in terms of how frequently the two are cited together in the literature, an algorithm should be devised to sort genes according to their relevance to the SADR topics.
Results: The SADR-Gengle database, which is made up of gene–SADR relationships extracted from Pubmed, has been constructed, covering six major SADRs, namely cholestasis, deafness, muscle toxicity, QT prolongation, Stevens–Johnson syndrome and torsades de points. The CitationRank algorithm, which inherits the principle of the Google PageRank algorithm that a gene should be highly ranked when biologically related to other highly ranked genes, is devised. The algorithm performs robustly in recovering SADR-related genes in the presence of extraneous noise, and the use of the algorithm has been extended to sorting genes in our database. Users can browse genes in a Google-type system where genes are ordered according to their descending relevance to the SADR topic selected by the user. The database also provides users with visualized gene–gene knowledge chain networks, helping them to systematize their gene-oriented knowledge chain whilst navigating these networks.
Availability: The SADR-Gengle is freely available at http://Gengle.Bio-X.cn/SADR/.
Contact: helinhelin{at}gmail.com
Supplementary information: Supplementary data are available at Bioinformatics online.
|
1 INTRODUCTION |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
Serious adverse drug reaction (SADR) has always been a concern, particularly when Vioxx® events (Furberg, 2006) and Avandia® events (Nissen and Wolski, 2007) are reported. Better drug safety could be achieved if the genetic risk factors and the mechanisms of SADR (Wilke et al., 2007) could be identified. However, candidate gene selection is hampered by a lack of knowledge of the SADR mechanism (Need et al., 2005). Genes at the pharmacokinetic level, e.g. the drug metabolite enzymes or transporters, do not give a satisfactory explanation for type B SADRs, namely the dose-independent ADRs (Pirmohamed and Park, 2001). Genome-wide association studies are handicapped by the lack of case-control samples, and their results cannot be properly interpreted without knowledge of the molecular mechanism of SADR. Researchers therefore need a database pooling comprehensive information on SADR-associated genes. For example a researcher exploring the Stevens–Johnson syndrome (SJS) should be able to access the relevant genes and their relationships with SJS on a relevant semantic web. Such gene-oriented databases do exist, covering gene-disease (Allen et al., 2008; Bertram et al., 2007; Lin et al., 2006) and gene-drug (Altman, 2007) relationships. However, to our knowledge, no database has so far been set up to cover the gene–SADR relationship.
The construction of a gene-oriented SADR database could be accomplished by text-mining the SADR literature. However, one obstacle is the entry recognition of gene names (Jensen et al., 2006). In addition, it is not easy to identify from a Pubmed entry whether the corresponding species is Homo sapiens. GeneRIF (Gobeill et al., 2008) of Entrez Gene includes a semi-automatic index of gene–Pubmed relationships examined by database curators. Furthermore, a gene2pubmed index was constructed from the GeneRIF and the GenBank. It possesses high precision but low sensitivity in retrieving SADR-related genes. The major problem is that there is no guarantee that gene A is less important than gene B to the SADR X simply because it is less frequently cited in X-related literature than B. Google's PageRank algorithm .(Brin and Page, 1998) presumes that a web page should be highly ranked if other highly ranked pages have hyperlinks to it. Morrison et al. (2005) extended this major principle, and successfully applied it to prioritize genes from a noisy microarray dataset. The algorithm could be extended to enrich and to sort SADR-related genes, so that a gene would be assigned a high relevance to X if it is co-cited (Jenssen et al., 2001) with other genes of high relevance to X, even if it is less frequently mentioned in X-related literature. The PageRank algorithm adopts the random walk hypothesis on a probability model. A recent study interpreting information flow within a network utilized an information-theoretic approach on the random walk model (Rosvall and Bergstrom, 2008), and was further applied in measuring the impact of journals (Bergstrom et al., 2008). Based on these algorithmic advances, we put forward an algorithm implementing the random walk model to evaluate genes' importance in the literature. The algorithm was set up upon a gene–gene knowledge chain network (GKCN) and its effectiveness has been tested on several noisy datasets. We have therefore utilized it to sort genes in our own database named SADR Gengle, with the objective of providing gene–SADR data on user-friendly interfaces.
|
2 METHODS |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
2.1 Construction of SADR-oriented bibliome and the genes' knowledge chain network
The six SADRs included in the database were those usually reported in the adverse event report system of US FDA. Users can receive up-dates on topics and genes by subscribing to the RSS feed. Six groups of Pubmed entries were retrieved using the following querying terms: cholestasis; deafness OR ‘hearing loss’; ‘long QT’ OR ‘QT prolongation’; rhabdomyolysis OR myalgia OR myopathy OR myositis; rash OR SJS OR toxic epidermal necrolysis; torsade de pointes. The records were downloaded through the eSearch and eFetch APIs, and were deposited into a rational database (MySQL 6.0). The reference impact factor for each Pubmed entry is the mean impact factor of the relevant journal in the science citation index from 2003 to 2007. The gene–Pubmed index called gene2pubmed was downloaded from the FTP site of Entrez Gene.
In the index, the corresponding species of genes other than Homo sapiens were excluded. Pubmed entries with a co-cited gene number greater than five were also eliminated, because empirically, they could not bear enough information to portray the clear gene–gene relationships. If two genes are co-cited in an entry, they tend to relate to each other biologically directly or indirectly (Hoffmann and Valencia, 2003, 2004; Jenssen et al., 2001). A connection between two genes is established if they are co-cited in at least two Pubmed entries. The GKCN of the six SADR topics was constructed based on the genes' co-citation, and were further transformed into XGMML (Shannon et al., 2003) so as to be visualized in the jSquid (Klammer et al., 2008) Applet.
2.2 Assignment of the core genes and
the extended genes
We were able to retrieve genes indexed in SADR
X-related Pubmed from the gene2pubmed index. For a gene i, the
ai, bi,
ci, di values,
representing the number of Pubmed entries containing gene i
(ai or bi) and not containing
gene i (ci or di)
under X or non-X, respectively (Supplementary
Table 1), were counted and the citation rate ratio (CRR) was
calculated as:
|
(1) |
A gene was assigned as a core gene if its CRR exceeded zero with a and b exceeding three. This threshold was set to insure the gene's knowledge was fully disclosed in X and non-X. Genes whose CRR was equal to zero but were connected to the core gene in the GKCN were assigned as extended genes. Genes whose CRR were greater than zero with either of their a or b less than three would be regarded as extended genes, with their CRR being set to zero. For X, the core and extended genes can be regarded as genes which were directly or indirectly associated with X, and the CRR of a gene to some extent reflected the importance of the gene in topic X.
2.3 The CitationRank
algorithm
Sorting SADR-related genes by their CRR value potentially
creates false negatives. For example, it would be wrong to assign
a low importance to a gene because currently its citation rate
is low, since our knowledge of the molecular mechanism of SADRs
is limited, especially when the gene is biologically linked to
other genes with high CRRs, it remains unclear, therefore, whether
this gene should be omitted. This problem is particularly acute in
SADR research, where knowledge of all SADR-related genes is scarce
anyway. Following the logic of the Google PageRank algorithm (Brin
and Page, 1998)
and the EigenFactor algorithm (Bergstrom et al., 2008),
we put forward an algorithm named CitationRank. PageRank is based on
the premise that the original rank of pagei can be
measured by the probability 1 –d, which is the probability of
an internet surfer randomly choosing it over all web pages. The
surfer can also arrive at pagei with probability
d from other web pages holding hyperlinks to it. In
CitationRank, the original rank of genei is defined as
the likelihood that a researcher would access it in the course
of looking at papers of a specific SADR topic whilst browsing
the literature. We use (1 – d)CRRi as a
measure of this possibility. The researcher can also ‘think of’
genei when looking at other genes which are
co-cited with genei in the literature. The links in
the algorithm can be defined as edges in GKCN. Thus in the following
iteration equation, the citation rank of
genei(cri) consists of two terms,
|
(2) |
denotes the citation rank (CR) of genei in the
nth iteration. The initial rank vector is taken as
R[0] = CRR/||CRR||1. W 
NxN is the
connectivity matrix of GKCN, with wij representing
the number of Pubmed entries where genei and
genej are co–cited. The linesj
equals 
wij. The CitaionRank uses the
parameter d in the range [0..1] to control the weighting of
CRR and the network connectivity in the rank calculations. It has
been proved that convergence of the iteration is guaranteed for all 0
< d < 1 (Morrison et al., 2005).
Hence the Jacobi iteration was performed to solve the vector
R.
2.4 Test for the robustness of
CitationRank algorithm against false negative noises
The aim of
applying the CitationRank is to solve the problem of the false
negative while using CRR to prioritize SADR-related genes. Here we
conducted the ‘leave one out cross validation’ (LOOCV) to test the
effectiveness of the algorithm when noises of the false negatives
were introduced. For SADR X, the CitationRank vector R for
Q X-related genes, including core and extended genes, was
first calculated. Genes were sorted by the ascending CR to generate
anorder vector O0. Then for M core genes, the
observation of the CRR of genei was set to zero in turn,
and the vector Ri was re-calculated based on
Q– 1 CRR observations. So an order vector
Oi was then constructed where the re-ordered
position for the omitted genei can be recovered as
Oii. Here the ‘boosting’ index B for
genei was computed as:
|
(3) |
2.5 Test for the robustness of the CR
value-based binary classification model
A gene can be classified as
a core or an extended gene when a threshold of a certain
classification variable of the gene is set. Different ROC curves can
be drawn using different classification variables, and the
effectiveness of the classification can be measured by the area under
curves (AUCs) of the ROC curves. The ‘1–specificity’ and
‘sensitivity’ of the x-axis and the y-axis are defined
as:
|
(4) |
|
(5) |
2.6 Enrichment of SADR associated
pathways
For SADR X, core and extended genes with high CR value
greater than 0.001 were included in the enrichment analysis of
KEGG Orthology (Kanehisa et al., 2008)
using KOBAS (Mao et al., 2005).
A threshold at 0.05 of the q value (Strimmer, 2008)
was used to choose the specifically enriched pathways. The enriched
terms and the contributing genes were deposited in our database.
|
3 RESULTS |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
3.1 Robustness of the CitationRank algorithm against false negative noises
We took the muscle toxicity related gene set as an example. We set the CRR of the core genes in turn to zero to intentionally generate false negatives. The CR value responded robustly to this anomaly, as most of the wrongly assigned core genes were recovered close to their original place in the ranked list (Fig. 1b-d). The performances of LOOCV were measured by C values. The best C was 0.63, which was achieved when the parameter d was set to 0.9 (Fig. 1d), namely 74 of 118 core genes being recalled in 90% of their original positions. However, a portion of genes failed to be recovered by CitationRank irrespective of what d values were set. Thiswas due to their lack of connection in the GKCN. These ‘orphan’ genes could not share the rank value spread from other genes. Similar trends of the algorithm could be observed in other SADR-related gene sets (Supplementary Figs 1 and 2).
|
3.2 The robustness of the CR value based classification model
We further tested the power of mapping muscle toxicity related genes onto core and extended classes using CRR and CR as the classification variable. Given that the total number of the core genes was 118, parameter N was set at 30, 60 and 90 in three tests. According to the Equation (2), the CR vector was equivalent to the CRR vector when d equals zero. The lowest AUC was always observed when CRR was used as the classification variable (d = 0 in Fig. 2a) whichever N was applied. A higher d-value resulted in higher AUCs (Fig. 2). When the noise was not strong, namely N was set at 30 and 60 (Fig. 2a and b), the CR-based classifier did not perform significantly better than the CRR-based classifier, and the d parameter did not seem to be a key factor in determining the AUC. However, when N was set at 90 (Fig. 2c), the AUC decreased significantly when d was set at 0 and 0.5 compared to the corresponding results in Figure 2a and b, but the AUC remained unchanged in the classifier when d was set at 0.9. Even when these 90 core genes (76%) were wrongly assigned, an AUC of 0.81 (Fig. 2c) was still achieved with this classifier, indicating that considerable robustness could be derived from the latter term in Equation (2) of the CitationRank algorithm.
|
Similar trends existed in other SADRs (Supplementary Figs 3–4). We therefore set d at 0.9 to prioritize core genes and to highlight extended genes in the database. We also estimated empirically that about 50% of the true core genes were missed by the gene2pubmed index, thus we mainly studied the situation in Figure 2b where 60 of 118 (about 50%) core genes were wrongly assigned since this mimicked the real situation. A sensitivity of 0.80 and a specificity of 0.86 was reached around the inflexion of the ROC curve when the CR threshold was chosen at 0.074 (d = 0.9). Thus in the case of the muscle toxicity topic, extended genes whose CR was greater or less than 0. 074 could be regarded as highly relevant or not highly relevant to the topic.
3.3 Browsing the SADR-related genes
in SADR Gengle
As described above, the CitationRank algorithm can
solve the problem of the false negatives, indicating that the basic
premise of the algorithm is sound. This premise was then naturally
extended to the sorting of the core genes in our database. Genes
should be sorted and presented to the client by their decreasing
relevance to topic X, and such relevance was measured by the CR
values in our database. In Google, web pages relevant to the
user's key word are sorted mainly based on their PageRank (Brin
and Page, 1998).
In our system one can browse the SADR-related genes as if browsing
the search results from Google. The key data statistics of SADR
Gengle was listed in Table
1. Taking SJS for example, core genes are sorted by their
descending CR and are presented as gene cards with general
information presented to help the user to pursue their interests.
Within each card, the relevant extended genes, if any, in GKCN are
displayed, and the SJS-related literature containing this gene is
displayed. Here the impact factor of the gene denotes the mean
reference impact factor of the SJS-related Pubmed entries carrying
this gene. One can go to the detailed information page by
clicking on the gene name. In the detailed page of HLA-B under
SJS, all SJS-related papers describing HLA-B are listed. Gene
ontology terms, KEGG pathways and OMIM entries of HLA-B, if
any, are displayed. Detailed information of any ‘child’ genes
can be retrieved by navigating the ‘NETWORK NAVIGATOR’ on the
right of the page.
|
3.4 Case study of the CitationRank algorithm
To comprehend the parameter d in the Equation (2) and how could robustness be achieved against noises, we examined the CR values' change within a local network of ‘muscle toxicity’ related genes by assigning different d values (Fig. 3). All genes were extended genes except for SGCE and HLA-DRB1, whose CR values were equivalent to the CRR values when d was zero (Fig. 3a). The size of each node reflected their CR values. After setting d at 0.5, the CR value of SGCE spread to LAMA1, TOR1A and HLA-DRB1 (Fig. 3b). Such ‘spread effect’ became stronger when d was set at 0.9 (Fig. 3c). So even with the CR value being set to zero intentionally, HLA-DRB1 could still share the CR values from its neighbors such as TOR1A, hence had the chance to appear near its original position of the ranking list. On the other hand, if a gene, such as LAMA1, was absent in the gene2pubmed index, it could also be highlighted by the CitationRank algorithm.
|
3.5 Navigating the genes' knowledge chain network
The GKCN was used to define the linkages between all genes which were required by the CitationRank algorithm. In addition, the GKCN was also able to denote the biological relationship among genes (Hoffmann and Valencia, 2003; Hoffmann and Valencia, 2004; Jenssen et al., 2001), and could be applied intuitively in the SADR Gengle, enabling users to jump over the genes to systematize and create their own gene-oriented knowledge chain on SADR.
There are two ways to navigate the GKCN. Taking SJS for example again, one can jump from one gene to its ‘neighbors’ using network navigator or the visualized GKCN in the Applet. The navigation can begin from the gene list page in part I of Figure 4, where SJS-related genes are sorted by their CR values, and are presented in a Google pattern. Clicking on HLA-B will lead the user to the detailed information page (part II in Fig. 4), where SJS-related Pubmed entries carrying HLA-B will be displayed. Its neighbors in GKCN, the extended genes, are listed in the network navigator. Following the link of an extended gene (LTA) will lead the user to another navigator (part III in Fig. 4) where the ‘parent’ genes of LTA are displayed. Clicking on IL4R in the navigator, for example, will take the user to another detailed page, where some new gene–SJS knowledge is presented (part IV in Fig. 4). Users can jump between core and extended genes in such a way to retrieve a gene-oriented knowledge chain of SJS. Furthermore, one can also identify this knowledge chain in the visualized GKCN (part V in Fig. 4). Core genes are in grey dashed rectangles, whereas extended gene LTA is in blue dashed rectangle. One can also access the detailed page of a gene from the Applet through right clicking on a particular gene to retrieve a popup link. In this example, the user has constructed a knowledge chain of ‘SJS–HLA-B–LTA–IL4R–SJS’ extracted from GKCN. He could deduce that LTA relates to HLA-B and IL4R, two core genes that are involved in the pathogenesis of SJS (Hung et al., 2005; Ueta et al., 2007), implying a putative functional linkage of this extended gene to SJS through MHC I-mediated pathways (Chessman et al., 2008) or cytokine-mediated pathways (Ueta et al., 2007). On the other hand, the logical connectivity of these three genes might prompt the user to conceive the hypothesis that a crosstalk between a MHC I-related pathway and a cytokine-related pathway might be mediated by LTA, which might give a systematic explanation of the pathogenesis of SJS.
|
Another case of using GKCN to retrieve potential molecular knowledge about SADR again concerns TOR1A, a muscle disease associated gene (Zorzi et al., 2009), which co-cited with ‘muscle toxicity’ associated literature (Wong et al., 2005) but does not meet the criteria for becoming a core gene. However, it links to the core gene, SGCE, which plays an important role in myoclonus dystonia (Ettinger et al., 1997). By navigating network navigator or the GKCN, the user can easily identify TOR1A as a potential muscle toxicity related gene, for it not only achieved a CR of 0.11, the seventh highest CR among 1763 extended genes, but was also identifiable as a ‘child’ gene of SGCE in the visualized GKCN. The examples above indicate that the SADR Gengle could help to mine embedded knowledge, generating important hypotheses for prospective validation of the candidate genes for this poorly understood subject.
SADR Gengle focuses on providing the bibliomic information (Searls, 2001), the gene sorting algorithm and the methodology for user to navigate the literature. Users might also refer to protein–protein interaction (PPI) and gene co-expression data while reading the literature in the database. To facilitate the users, we provide hyperlinks for each gene to the STRING server (Jensen et al., 2009), which harbors comprehensive PPI and co-expression data.
3.6 Enrichment of SADR associated
pathways
To achieve a more precise enrichment of pathways that were
shared specifically by the genes under a certain SADR topic, we
included core genes and highly ranked extended genes for enrichment
analysis. Pathways enriched were usually consistent with existing
knowledge. In the case of the SJS topic, for example, ‘T cell
receptor-signaling pathway’ (q = 9.8E–20), ‘Antigen
processing and presentation’ (q = 1.4E–17) and ‘B cell
receptor signaling pathway’ (q = 1.0E–8) tended to occur
more frequently in SJS-related genes than in a random human gene
selection at the significance level of 0.01. The result corresponded
with the known SJS mechanism (Borchers et al., 2008),
indicating that the genes involved in the analysis, which was
highlighted by CitationRank, could to some extent represent the
molecule mechanism of this SADR. Furthermore, most of the results
could be replicated by another enrichment tool (DAVID) (Huang da
et al., 2007).
When the enrichment results of six SADRs were summarized (see SADR
Gengle online), most of them were found to share several common
pathways, such as ‘Gap junction’, ‘Toll-like receptor signaling
pathway’ and ‘Adherens junction’. Although these SADRs are
triggered by different drugs and occurred in different tissue, the
sharing of the same pathways implied that these common biological
procedures might be responsible for the pathogenesis of the SADR and
thus were worthy of experimental validation.
|
4 CONCLUSIONS |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
- The CitationRank algorithm has the potential for sorting genes
by their relevance to a topic, and is effective in uncovering
false negatives of the gene set relevant to a particular topic;
- At a time when molecular mechanisms are poorly understood
and gene-oriented knowledge is not well organized, SADR
Gengle enables users to broaden and systematized their
gene-oriented knowledge on SADR;
- The SADR Gengle is an instance of the Gengle knowledge
depositing model, which inherits some of the ‘genes’
from Google, and is designed to organize structured
gene-topic knowledge. The model could be instantiated
onto topics other than disease, such as longevity, pain
and reproduction etc., about which knowledge has begun
to accumulate but is poorly organized. Such information
could be quickly organized on a gene-oriented basis using
the methodology described in this study.
|
ACKNOWLEDGEMENTS |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
We appreciate Jian Chen, Shuiqing Huang, Zichun Hua, Bin Wang, Gongli Xia, Xiangzhe Zhang and Zhenhua Xia for helpful discussions. We are grateful to the developers of the Pubmed and Entrez Gene. We thank the suggestions from the anonymous reviewers in improving the manuscript.
Funding: China Postdoctoral Science Foundation (PSF) [20070420660 to L.Y.], Shanghai PSF [61444 to L.Y.], National Natural Science Foundation, 863, 973 projects of China [07DZ22917, 2006AA02A407, 2006CB910600, 2006BAI05A05, 2007CB947300] and the Shanghai Leading Academic Discipline Project [B205].
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Limsoon Wong
Received on March 12, 2009; revised on
June 9, 2009; accepted on June 9, 2009
|
REFERENCES |
|---|
| TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 CONCLUSIONSACKNOWLEDGEMENTSREFERENCES |
|---|
Allen NC, et al. Systematic meta-analyses and field synopsis of genetic association studies in schizophrenia: the SzGene database. Nat. Genet. (2008) 40:827–834.[CrossRef][Web of Science][Medline]
Altman RB. PharmGKB: a logical home for knowledge relating genotype to drug response phenotype. Nat. Genet. (2007) 39:426.[CrossRef][Web of Science][Medline]
Bergstrom CT, et al. The Eigenfactor metrics. J. Neurosci. (2008)
28:11433–11434.
Bertram L, et al. Systematic meta-analyses of Alzheimer disease genetic association studies: the AlzGene database. Nat. Genet. (2007) 39:17–23.[CrossRef][Web of Science][Medline]
Borchers AT, et al. Stevens-Johnson syndrome and toxic epidermal necrolysis. Autoimmun. Rev. (2008) 7:598–605.[CrossRef][Web of Science][Medline]
Brin S, Page L. The anatomy of a large-scale hypertextual Web search engine. Comput. Networks ISDN Syst. (1998) 30:107–117.[CrossRef]
Chessman D, et al. Human leukocyte antigen class I-restricted activation of CD8+ T cells provides the immunogenetic basis of a systemic drug hypersensitivity. Immunity (2008) 28:822–832.[CrossRef][Web of Science][Medline]
Ettinger AJ, et al. epsilon-Sarcoglycan, a broadly expressed homologue of
the gene mutated in limb-girdle muscular dystrophy 2D. J. Biol. Chem.
(1997) 272:32534–32538.
Furberg CD. Adverse cardiovascular effects of rofecoxib. N Engl J Med (2006) 355:204–205. 204; author reply.[Web of Science][Medline]
Gobeill J, et al. Gene Ontology density estimation and discourse analysis for automatic GeneRiF extraction. BMC Bioinformatics (2008) 9(Suppl. 3):S9.
Hoffmann R, Valencia A. Life cycles of successful genes. Trends Genet. (2003) 19:79–81.[CrossRef][Web of Science][Medline]
Hoffmann R, Valencia A. A gene network for navigating the literature. Nat. Genet. (2004) 36:664.[CrossRef][Web of Science][Medline]
Huang da W, et al. DAVID Bioinformatics Resources: expanded annotation
database and novel algorithms to better extract biology from large gene lists.
Nucleic Acids Res. (2007) 35:W169–W175.
Hung SI, et al. HLA-B*5801 allele as a genetic marker for severe cutaneous
adverse reactions caused by allopurinol. Proc. Natl Acad. Sci. USA
(2005) 102:4134–4139.
Jensen LJ, et al. STRING 8 – a global view on proteins and their functional
interactions in 630 organisms. Nucleic Acids Res. (2009)
37:D412–D416.
Jensen LJ, et al. Literature mining for the biologist: from information retrieval to biological discovery. Nat. Rev. Genet. (2006) 7:119–129.[CrossRef][Web of Science][Medline]
Jenssen TK, et al. A literature network of human genes for high-throughput analysis of gene expression. Nat. Genet. (2001) 28:21–28.[CrossRef][Web of Science][Medline]
Kanehisa M, et al. KEGG for linking genomes to life and the environment.
Nucleic Acids Res. (2008) 36:D480–D484.
Klammer M, et al. jSquid: a Java applet for graphical on-line network
exploration. Bioinformatics (2008) 24:1467–1468.
Lin BK, et al. Tracking the epidemiology of human genes in the literature:
the HuGE Published Literature database. Am. J. Epidemiol. (2006)
164:1–4.
Mao X, et al. Automated genome annotation and pathway identification using
the KEGG Orthology (KO) as a controlled vocabulary. Bioinformatics
(2005) 21:3787–3793.
Morrison JL, et al. GeneRank: using search engine technology for the analysis of microarray experiments. BMC Bioinformatics (2005) 6:233.[CrossRef][Medline]
Need AC, et al. Priorities and standards in pharmacogenetic research. Nat. Genet. (2005) 37:671–681.[CrossRef][Web of Science][Medline]
Nissen SE, Wolski K. Effect of rosiglitazone on the risk of myocardial
infarction and death from cardiovascular causes. N. Engl. J. Med.
(2007) 356:2457–2471.
Pirmohamed M, Park BK. Genetic susceptibility to adverse drug reactions. Trends Pharmacol. Sci. (2001) 22:298–305.[CrossRef][Medline]
Rosvall M, Bergstrom CT. Maps of random walks on complex networks reveal
community structure. Proc. Natl Acad. Sci. USA (2008)
105:1118–1123.
Searls DB. Mining the bibliome. Pharmacogenomics J. (2001) 1:88–89.[Medline]
Shannon P, et al. Cytoscape: a software environment for integrated models
of biomolecular interaction networks. Genome Res. (2003)
13:2498–2504.
Strimmer K. A unified approach to false discovery rate estimation. BMC Bioinformatics (2008) 9:303.[CrossRef][Medline]
Ueta M, et al. Association of IL4R polymorphisms with Stevens-Johnson syndrome. J. Allergy Clin. Immunol. (2007) 120:1457–1459.[CrossRef][Web of Science][Medline]
Wilke RA, et al. Identifying genetic risk factors for serious adverse drug reactions: current progress and challenges. Nat. Rev. Drug Discov. (2007) 6:904–916.[CrossRef][Web of Science][Medline]
Wong VC, et al. Stiff child syndrome with mutation of DYT1 gene.
Neurology (2005) 65:1465–1466.
Zorzi G, et al. Early onset primary dystonia. Eur J Paediatr Neurol. (2009).
