ArticlePDF Available

Translating Clinical Findings into Knowledge in Drug Safety Evaluation - Drug Induced Liver Injury Prediction System (DILIps)

Authors:

Abstract and Figures

Drug-induced liver injury (DILI) is a significant concern in drug development due to the poor concordance between preclinical and clinical findings of liver toxicity. We hypothesized that the DILI types (hepatotoxic side effects) seen in the clinic can be translated into the development of predictive in silico models for use in the drug discovery phase. We identified 13 hepatotoxic side effects with high accuracy for classifying marketed drugs for their DILI potential. We then developed in silico predictive models for each of these 13 side effects, which were further combined to construct a DILI prediction system (DILIps). The DILIps yielded 60-70% prediction accuracy for three independent validation sets. To enhance the confidence for identification of drugs that cause severe DILI in humans, the "Rule of Three" was developed in DILIps by using a consensus strategy based on 13 models. This gave high positive predictive value (91%) when applied to an external dataset containing 206 drugs from three independent literature datasets. Using the DILIps, we screened all the drugs in DrugBank and investigated their DILI potential in terms of protein targets and therapeutic categories through network modeling. We demonstrated that two therapeutic categories, anti-infectives for systemic use and musculoskeletal system drugs, were enriched for DILI, which is consistent with current knowledge. We also identified protein targets and pathways that are related to drugs that cause DILI by using pathway analysis and co-occurrence text mining. While marketed drugs were the focus of this study, the DILIps has a potential as an evaluation tool to screen and prioritize new drug candidates or chemicals, such as environmental chemicals, to avoid those that might cause liver toxicity. We expect that the methodology can be also applied to other drug safety endpoints, such as renal or cardiovascular toxicity.
Content may be subject to copyright.
Translating Clinical Findings into Knowledge in Drug
Safety Evaluation - Drug Induced Liver Injury Prediction
System (DILIps)
Zhichao Liu
1
, Qiang Shi
1
, Don Ding
2
, Reagan Kelly
2
, Hong Fang
2
, Weida Tong
1
*
1Division of Systems Biology, National Center for Toxicological Research, United States Food and Drug Administration, Jefferson, Arizona, United States of America, 2ICF
International Company at FDA’s National Center for Toxicological Research, Jefferson, Arizona, United States of America
Abstract
Drug-induced liver injury (DILI) is a significant concern in drug development due to the poor concordance between
preclinical and clinical findings of liver toxicity. We hypothesized that the DILI types (hepatotoxic side effects) seen in the
clinic can be translated into the development of predictive in silico models for use in the drug discovery phase. We
identified 13 hepatotoxic side effects with high accuracy for classifying marketed drugs for their DILI potential. We then
developed in silico predictive models for each of these 13 side effects, which were further combined to construct a DILI
prediction system (DILIps). The DILIps yielded 60–70% prediction accuracy for three independent validation sets. To
enhance the confidence for identification of drugs that cause severe DILI in humans, the ‘‘Rule of Three’’ was developed in
DILIps by using a consensus strategy based on 13 models. This gave high positive predictive value (91%) when applied to an
external dataset containing 206 drugs from three independent literature datasets. Using the DILIps, we screened all the
drugs in DrugBank and investigated their DILI potential in terms of protein targets and therapeutic categories through
network modeling. We demonstrated that two therapeutic categories, anti-infectives for systemic use and musculoskeletal
system drugs, were enriched for DILI, which is consistent with current knowledge. We also identified protein targets and
pathways that are related to drugs that cause DILI by using pathway analysis and co-occurrence text mining. While
marketed drugs were the focus of this study, the DILIps has a potential as an evaluation tool to screen and prioritize new
drug candidates or chemicals, such as environmental chemicals, to avoid those that might cause liver toxicity. We expect
that the methodology can be also applied to other drug safety endpoints, such as renal or cardiovascular toxicity.
Citation: Liu Z, Shi Q, Ding D, Kelly R, Fang H, et al. (2011) Translating Clinical Findings into Knowledge in Drug Safety Evaluation - Drug Induced Liver Injury
Prediction System (DILIps). PLoS Comput Biol 7(12): e1002310. doi:10.1371/journal.pcbi.1002310
Editor: Greg Tucker-Kellogg, National University of Singapore, Singapore
Received June 29, 2011; Accepted November 1, 2011; Published December 15, 2011
This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for
any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Funding: The report study is a part of FDA’s Liver Toxicity Knowledge Base (LTKB) project that is supported by the FDA’s Critical Path program and Chief Scientist
Challenge Grant. This is an internal project supported by the FDA’s National Center for Toxicological Research. ZL is grateful to the National Center for
Toxicological Research (NCTR) of U.S. Food and Drug Administration (FDA) for postdoctoral support through the Oak Ridge Institute for Science and Education
(ORISE). The views presented in this article do not necessarily reflect those of the U.S. Food and Drug Administration. The funders had no role in study design, data
collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: weida.tong@fda.hhs.gov
Introduction
Drug-induced liver injury (DILI) poses a significant challenge to
medical and pharmaceutical professionals as well as regulatory
agencies. It is the leading cause of acute liver failure, which has a
high mortality rate (30%) as treatment is limited due to the
availability of livers for transplantation [1]. Although many
dangerous drugs are identified during animal testing thus
protecting humans from this damage, a consortium determined
that about half of the drugs that cause human hepatotoxicity
were not identified as having this potential in nonclinical
animal testing [2]. Many drugs have been withdrawn from the
market or have received restrictions and warnings due to DILI [3].
DILI information and guidance for pharmaceutical industries has
been released by regulatory agencies such as the U.S. Food and
Drug Administration (FDA) (http://www.fda.gov/downloads/
Drugs/GuidanceComplianceRegulatoryInformation/Guidances/
UCM174090.pdf), European Medicines Agency (EMA) (www.
ema.europa.eu/pdfs/human/swp/15011506en.pdf) and Health
Canada (http://www.hc-sc.gc.ca/dhp-mps/alt_formats/pdf/con-
sultation/drug-medic/draft_ebauche_hepatotox_guide_ld-eng.pdf),
highlighting both the significance and difficulties in DILI
research. In the FDA, the Critical Path Initiative identified
DILI as a key area of focus in a concerted effort to broaden the
agency’s knowledge for better evaluation tools and safety
biomarkers (http://www.fda.gov/ScienceResearch/SpecialTopics/
RegulatoryScience/ucm228131.htm).
Determining the potential for a drug candidate to cause DILI in
humans is a challenge. First, the standard pre-clinical animal
studies do not effectively predict DILI events in humans. In one
notorious example, five subjects in a phase 2 clinical trial
experienced fatal hepatotoxicity induced by fialuridine, an
investigational nucleoside analogue that showed no liver damage
in animal studies [4]. Out of 221 pharmaceuticals, the overall
concordance of liver toxicity in humans and experimental animals
is as low as 55%, which is in sharp contrast with the concordance
of other target organs such as the hematological (91%),
gastrointestinal (85%), and cardiovascular (80%) systems [2].
PLoS Computational Biology | www.ploscompbiol.org 1 December 2011 | Volume 7 | Issue 12 | e1002310
Secondly, even well-controlled clinical trials fail to accurately
predict post-marketing DILI events. The main reason for this is
the statistical power of the trials – the risk of severe DILI of an
idiosyncratic nature is very low per exposed subject, while clinical
trials are usually carried out with only several thousand patients
[5], rendering them significantly underpowered to predict rare
DILI events.
To enhance the predictability of DILI, novel approaches have
been explored by many researchers. Notable examples include (a)
development of new DILI biomarkers [6], (b) introduction of high-
content screening [7], (c) adoption of more sensitive animal models
[8,9,10], and (d) utilization of toxicogenomics [11]. Most of these
investigations are focused on developing biomarkers using either
animal or in vitro models for predicting DILI in humans. This still
would involve synthesis of the drug and elaborate testing. An in
silico approach could inform chemists at the earliest point in the
drug discovery pipeline and enable them to select the best
chemical structures.
We hypothesized that there exists a distinct set of liver side
effects that can be used to characterize the DILI risk of drugs in
humans. We identified 13 types of hepatotoxicity (hepatotoxic side
effects or HepSEs) from the organ levels of hepatobiliary disorders in
the Medical Dictionary for Regulatory Activities (MedDRA)
ontology (http://www.meddramsso.com/). We found that these
13 HepSEs can discriminate DILI drugs from non-DILI drugs
with high accuracy (,83%). Since the side effects are clinical
observations obtained either from clinical trials or from post-
marketing surveillance with limited utility in drug discovery, we
developed quantitative structure-activity relationship (QSAR)
models for each of the HepSEs. We then constructed a DILI
prediction system (DILIps) based on the 13 HepSE models with an
improved prediction strategy using a ‘‘Rule of Three’’ (RO3)
criterion (incriminated by 3 or more HepSE models). The systems
were evaluated in several external test sets with performance
surpassing most in silico models in the field. We screened the entire
drug list using the DILIps and evaluated the RO3 drugs in terms
of therapeutic use and drug targets.
Results
Identification and assessment of hepatotoxic side effects
(HepSEs)
Figure 1 is an overview of the approach taken. First, the
identification and assessment of HepSEs were performed. We used
the SIDER database [12] to identify drugs and associated side
effects. Out of 1450 side effects in the database, we selected only
those that were caused by more than 20 drugs (an arbitrary cut-
off). This yielded 473 side effects. The distribution of 888 drugs
over 473 side effects and vice verse were depicted in Supplemen-
tary Figure S1, indicating that over 90% drugs were associated
with at least 10 side effects. These side effects were then directly
mapped onto low level terms of MedDRA. The terms were linked
to the system organ classes (SOC) level according to the
hierarchical structure of MedDRA (Supplementary Table S1) in
order to determine the terms’ attributes at the organ level. Finally,
we considered side effects defined by the MedDRA ontology as
related to the hepatobilliary disorders SOC term as HepSEs, and
identified 13 HepSEs: bilirubinemia, cholecystitis, cholelithiasis, cirrhosis,
elevated liver function tests, hepatic failure, hepatic necrosis, hepatitis,
hepatomegaly, jaundice, liver disease, liver fatty, and liver function tests
abnormal.
We evaluated these 13 HepSEs for their ability to differentiate
drugs that do and do not cause DILI using the Liver Toxicity
Knowledge Base Benchmark Dataset (LTKB-BD) [13] and
PfizerData [14]. For both datasets, we used only the drugs that
they had in common with SIDER. There are several differences
between two datasets to call a drug as DILI or non-DILI (see
Materials and methods), including (1) LTKB-BD is based on the FDA-
approved drug labeling while PfizerData is according to the case
reports; (2) two datasets apply different criteria for DILI
assessment; and (3) drugs are grouped differently between two
datasets. To obtain an objective evaluation for 13 HepSEs, we
took the following actions to select DILI positive and negative
drugs from two datasets: (a) in LTKB-BD [13], Most-DILI-
Concern drugs were classified as positive while No-DILI-Concern
drugs were classified as negative; and (b) in PfizerData [14], drugs
with evidence in human toxicity were considered DILI positive
while drugs with no evidence in any species were considered DILI
negative. Defining a drug as causing DILI if it was positive in any
of the 13 HepSEs, this approach yielded 91% and 74% accuracy
for LTKB-BD and PfizerData, respectively.
It is important to note that the 26 MedDRA SOCs are not all
strictly related to human organs in a conventional sense. For
example, ‘‘investigations’’ and ‘‘general disorders and administra-
tion site conditions’’ are not organs (the complete list of MedDRA
SOC is available in Supplementary Table S1). Some side effects
with DILI indication are resided in a SOC other than the
hepatobilliary disorders. For example, the SOC of ‘‘investigations’’
include the ‘‘elevate liver enzyme’’ and ‘‘alkaline phosphatase
increased’’, both are conventional DILI indicators. Moreover,
some side effects in the SOC of ‘‘general disorders and
administration site conditions’’ could also be the manifestations
of DILI. Thus, we conducted a permutation test with the purposes
of confirming that the 13 HepSEs do in fact have significant
performance over the chance to distinguish DILI drugs from non-
DILI drugs. We randomly selected 3, 5,…, 21 side effects from the
473 side effect pool with each selection repeated 20,000 times. As
shown in Figure 2, the classification accuracy of the 13 HepSEs,
indicated by the red dot, was considerably higher than the average
accuracy for each of the sets of randomly selected side effects,
demonstrating that the observed classification accuracy for the 13
HepSEs was not due to chance.
Author Summary
Translational research involves utilization of clinical data to
address challenges in drug discovery and development.
The rationale behind this study is that the side effects
observed in clinical trial and post-marketing surveillance
can be translated into a screening system for use in drug
discovery. As a proof-of-concept study, we developed an
in silico system based on 13 hepatotoxic side effects to
predict drug-induced liver injury (DILI), which is one of the
most frequent causes of drug failure in clinical trial and
withdrawal from post-marketing application, and also one
of the most difficult clinical endpoints to predict from
preclinical studies. We first identified 13 types of liver
injury which yielded high prediction accuracy to distin-
guish drugs known to cause DILI from these don’t. To
effectively apply these 13 hepatotoxic side effects to the
drug discovery process for DILI, we developed in silico
models for each of these side effects solely based on
chemical structure data. Finally, we constructed a DILI
prediction system (DILIps) by combining these 13 in silico
models in a consensus fashion, which yielded .91%
positive predictive value for DILI in humans. The DILIps
methodology can be extended in applications for address-
ing other drug safety issues, such as renal and cardiovas-
cular toxicity.
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 2 December 2011 | Volume 7 | Issue 12 | e1002310
Figure 1. Overview of the workflow of DILIps development and its evaluation.
doi:10.1371/journal.pcbi.1002310.g001
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 3 December 2011 | Volume 7 | Issue 12 | e1002310
Figure 2. Box plot of classification accuracy with the number of selected side effects using a permutation test. (a) The test consisted of
128 drugs with a ratio of 69 DILI positives versus 59 DILI negatives in the LTKB-BD, and (b) 258 drugs with a ratio of 168 DILI positive drugs and 90 DILI
negative drugs in PfizerData. Given a randomly selected number of side effects, a drug showing positive in any of the side effects was considered as a
DILI positive drug. The process was repeated 20,000 times for each of the selected number of side effects. The red dot denotes the data based on the
13 HepSEs selected from the MedDRA hepatobilliary disorders category.
doi:10.1371/journal.pcbi.1002310.g002
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 4 December 2011 | Volume 7 | Issue 12 | e1002310
DILI prediction system (DILIps)
As illustrated in step 2 (Figure 1), QSAR models were developed
for each of the 13 HepSEs to enable their use in screening new
drug candidates computationally. The QSAR models developed
from the drugs related to each of these 13 HepSEs had high
prediction accuracy (.93%) in a leave-one-out cross-validation
(LOO-CV) process (Table 1).
Based on the 13 HepSE models, we further developed the
DILIps (step 3 of Figure 1, left box). Using the same classification
rule described above (i.e., drugs incriminated by any of the 13
HepSEs models are considered as DILI positives), we applied
DILIps to three external validation sets. The validation sets of
LTKB-BD and PfizerData contain drugs not used in developing
the 13 HepSE models. For the O’Brien et al. dataset [15], the
severe and moderate hepatotoxicity drugs were combined as DILI
positive drugs while the non-toxic drugs were defined as DILI
negative drugs (only the drugs not used by the 13 HepSE models
were included). As summarized in Table 2, the DILIps exhibited a
reasonable prediction performance for three external validation
sets with the prediction accuracy between 60–70%.
Development of the ‘‘Rule of Three’’ criterion in DILIps
Identifying the drugs of severe DILI potential with high
confidence has an important application since these drugs are
likely withdrawn from the market or restricted in use with black
box warning (BBW) due to the serious public health concern. We
assume that the number of models calling a drug causing DILI is
positively correlates with the drug’s severity for DILI and to the
extension of the confidence to make such a call. We generated a
union set based on the three validation datasets listed in Table 2.
We removed three drugs having an inconsistent DILI assignment
among three datasets (only three drugs were removed: moxisylyte,
carbidopa and terfenadine), i.e., positive in one dataset and
negative in another. This process resulted in 145 DILI positives
and 63 DILI negatives (see Supplementary Table S2). We used
this union set to assess how many HepSE models to be combined
likely identify high risk DILI drugs (i.e., withdrawal or BBW drugs)
with high positive predictive value (PPV).
Specifically, for each of the possible HepSE combination models
requiring a drug to be incriminated by
N
HepSE models (‘‘Rule of
N
’’), we calculate PPV and the number of drugs retained by each
of the HepSE combination models. As depicted in Figure 3, the
PPV reaches a maximum of 91.3% when
N
= 3. Additionally, the
percentage of high risk DILI drugs reached a local maximum at
N
= 3. Therefore, we established the RO3 criterion in the DILIps
for identifying drugs that might cause severe DILI with high
confidence (step 3 of Figure 1, right box). The number of drugs
meeting the RO3 is 23, dramatically decreased from 100 (RO1)
and 49 (RO2), which was expected when the optimization was
tilted toward increasing PPV. In order to identify the drugs of
severe DILI potential with high confidence, the trade-off was
accepted in the context of an application. Therefore, the RO3 was
selected to carry out further study.
DILI potential varies for different therapeutic categories
We applied the RO3 criterion to the drugs (small molecules
only) in DrugBank to investigate which therapeutic categories were
most likely associated with DILI (represented by the graph at the
Table 1. Performance of leave-one out cross-validation for the 13 HepSE models.
HepSE models #of drugs positive in HepSE Accuracy Sensitivity Specificity
bilirubinemia 88 0.96 0.76 0.99
cholecystitis 53 0.98 0.83 0.99
cholelithiasis 60 0.98 0.75 0.99
cirrhosis 27 0.99 0.89 0.99
elevated liver function tests 29 0.99 0.76 1.00
hepatic failure 132 0.95 0.82 0.97
hepatic necrosis 56 0.97 0.91 0.97
hepatitis 254 0.93 0.91 0.94
hepatomegaly 62 0.96 0.71 0.97
jaundice 274 0.93 0.89 0.95
liver disease 42 0.98 0.74 0.99
liver fatty 22 0.99 0.82 0.99
liver function tests abnormal 111 0.95 0.83 0.97
doi:10.1371/journal.pcbi.1002310.t001
Table 2. Performance of DILI prediction system (DILIps) on three literature datasets.
Datasets* The number of drugs for analysis (DILI positive drugs/DILI negative drugs) Accuracy Sensitivity Specificity
LTKB-BD 67/6 0.66 0.66 0.67
PfizerData 92/56 0.60 0.52 0.73
O’Brien et al. 25/15 0.70 0.56 0.93
*Only the drugs that did not overlap with the SIDER database were used.
doi:10.1371/journal.pcbi.1002310.t002
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 5 December 2011 | Volume 7 | Issue 12 | e1002310
right of step 4, Figure 1). Figure 4(a) shows the drug distribution
across 14 therapeutic categories as defined by Anatomic
Therapeutic Class (ATC) (http://www.whocc.no/atcddd/) with
the RO3 positive drugs highlighted in red. The enrichment of the
RO3 drugs in each therapeutic category was determined by
Fisher’s exact test. We found that two therapeutic categories (i.e.,
anti-infective for systemic use and musculoskeletal system drugs)
were significantly associated with drugs that cause DILI (p-
value = 5.00E-11 and 0.002, respectively). To confirm the findings,
we carried out the same analysis for drugs in the SIDER database
that met the RO3. As shown in Figure 4(b), the same two
therapeutic categories were also found to be significantly
associated with drugs that cause DILI (p-value = 8.94E-8 and
2.36E-7, respectively). Both results demonstrated that care must be
taken when drugs are developed with existing targets in these two
categories.
The findings are consistent with real-world observations; for
example, non-steroidal anti-inflammatory drugs (NSAIDs, a
subcategory of anti-infectives for systemic use) are often associated
with DILI. A good example is didanosine (VidexH) which is an
antiviral drug used to treat human immunodeficiency virus (HIV)
infection. On Jan 29
th
, 2010, the FDA notified healthcare
professionals and patients about a rare but serious complication
in the liver known as non-cirrhotic portal hypertension in patients
using the drug. Subsequently, a black box warning was added to
the drug label to warn doctors and consumers of this risk.
Didanosine can also cause lactic acidosis and severe hepatomegaly
with steatosis, and has resulted in several fatal cases (http://
dailymed.nlm.nih.gov/dailymed/drugInfo.cfm?id = 23496).
Associating the protein targets with DILI potential via
network analysis
It was important to determine if the drug target is related to the
drug’s likelihood of causing DILI. Accordingly, we investigated the
drugs that were RO3 positive from DrugBank in the target space
using network analysis as illustrated on the left side of Figure 1,
step 4. These drugs were associated with 134 human protein
targets. In the network analysis, we considered that two protein
targets are directly related (connected with an edge in network
analysis) if one or more drugs were associated with both targets.
As depicted in Figure 5, the network contains two large modules
(Modules #1 and #2) with several small modules. There are 72
targets in Module #1 associated with 125 RO3 positive drugs, and
23 targets in Module #2 associated with 8 drugs. We conducted
toxicity function and pathway analyses using Ingenuity Pathway
Analysis (IPA, http://www.ingenuity.com/) for both modules. In
each module, particularly Module #1, the biological functions
related to disease and disorder were investigated to assess if the
targets of the drugs meeting the RO3 have a relationship with
hepatic system diseases or disorders. As shown in Table 3, liver
injury and disease related functions enriched in Module #1 were
hepatic system disorder, jaundice, liver cancer, hepatocellular carcinoma, and
Figure 3. The evaluation of the ‘‘Rule of Three’’ (RO3). The predicted positive value, percentage of withdrawn or boxed warning (BW) drugs,
and the number of drugs meeting the ‘‘Rule of N’’ for different values of Nin the combined HepSE model.
doi:10.1371/journal.pcbi.1002310.g003
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 6 December 2011 | Volume 7 | Issue 12 | e1002310
Figure 4. The distribution of small molecular drugs in (a) DrugBank and (b) SIDER that satisfy the ‘‘Rule of Three’’ (RO3) at the first
level of Anatomical Therapeutic Chemical Classification System (ATC). The probability of the presence of DILI drugs is statistically significant
in two therapeutic categories (J and M).
doi:10.1371/journal.pcbi.1002310.g004
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 7 December 2011 | Volume 7 | Issue 12 | e1002310
hepatitis C. All the liver injury and disease functions are under the
hepatic system diseases branch of the top toxicity functions in IPA.
The other significant toxicity functions of Module #1 can be
found in Supplementary Table S3.
We also found that every drug in the two largest modules was
associated with more than three targets on average. Note that
drugs are prone to having multiple side effects if they interact with
multiple targets since different targets may invoke different side
Figure 5. Target network of corresponding drugs satisfying the ‘‘Rule of Three’’.
doi:10.1371/journal.pcbi.1002310.g005
Table 3. Summary of significant functions and pathways for module #1 identified in the network analysis of DILI targets using
IPA.
Functions Annotation p-Value Gene Name #Genes
hepatic system disorder 5.85E-25 ABCB1, ABCG2, ACSL4, ADRA1A, ADRA1B, ADRA1D, ALOX5, CASP3, CHRM1, CHRM2, CHRM3,
CHRM4, CHRM5, CHUK, DRD2, DRD5, HRH1, IKBKB, KIT, NR1I2, OPRM1, PDGFRA, PDGFRB,
PPARG, PTGS1, PTGS2, RXRA, SLC6A3, SLC6A4, TOP2A, VEGFA
31
jaundice 2.69E-15 CHRM1, CHRM2, CHRM3, CHRM4, CHRM5, CHUK, HRH1, IKBKB 8
liver cancer 5.58E-11 ABCB1, CA2, CASP3, HTR3A, KIT, LCK, PDGFRA, PDGFRB, PTGS1, PTGS2, RARA, SLC6A3,
SLC6A4, SRC, TOP2A, VEGFA
16
hepatocellular carcinoma 7.76E-08 CA2, CASP3, KIT, PDGFRA, PDGFRB, PTGS1, PTGS2, RARA, TOP2A, VEGFA 10
hepatitis C 1.04E-07 CASP3, DRD2, DRD5, OPRM1, PPARG, SLC6A4 6
doi:10.1371/journal.pcbi.1002310.t003
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 8 December 2011 | Volume 7 | Issue 12 | e1002310
effects [16,17]. We conducted text mining to verify the association
of 13 HepSEs and 134 targets identified by RO3 positive drugs.
We identified 45 proteins associated with eight HepSEs in a co-
occurrence analysis (Figure 6 and Supplementary Table S4). Most
of these targets are associated with hepatitis, while targets such as
PTGS2/COX-2 (prostaglandin-endoperoxide synthase 2) and
ABCD1 (ATP-binding cassette, sub-family, and member 1) are
related to multiple HepSEs.
Discussion
One application of translational science involves utilization of
clinical data to address challenges in drug discovery. The key
concept of this study is that the side effects observed in clinical
trials and post-marketing surveillance can be translated for use in
drug discovery. As a proof-of-concept study, we developed the
DILIps to address one of the most difficult clinical endpoints to
predict from preclinical studies, that is DILI. The DILIps contains
three distinct and sequential approaches. First, we identified 13
HepSEs based on the MedDRA ontology, which provided
excellent discrimination of a drug’s potential to cause DILI
(91% and 74% accuracy for LTKB-BD and PfizerData,
respectively). Secondly, HepSE-based QSAR models were devel-
oped by using all 888 drugs in SIDER, which were highly
predictive as compared to published models [14,18,19] and
offered a robust translation of clinical observation (i.e., side effects)
using in silico techniques to the drug discovery/preclinical testing
aspect of drug development. Next, we developed DILIps by
combining these 13 HepSE QSAR models, which yielded 60–70%
prediction accuracy for three independent validation sets. Lastly, a
RO3 criterion was implemented in DILIps, which had .91%
confidence for identification of drugs that might cause severe
DILI.
The DILIps is a modular system; each of its components can be
replaced by other methods or constructed using different variables.
For example, besides selecting 13 HepSEs from the hepatobiliary
disorders category in MedDRA, we also examined the effect of
including additional two DILI related terms from the investigation
SOC category, or selecting 14 DILI relevant terms as suggested by
an expert (Supplementary Table S5). Both yielded similar
performance compared to the 13 HepSE-based approach. Given
the fact that each MedDRA category is a stand-alone ontology and
other options did not yield exceptional performance, we choose
the terms under hepatobiliary disorders as representative types of DILI
in this study. For the second component of the DILIps, we
developed HepSE-based QSAR models because chemical struc-
ture data were readily available for the entire set of 888 drugs in
SIDER, providing a sufficiently large sample from which to build
the HepSE-based models. Other technologies, such as gene
expression microarrays, might be able to construct better HepSE
models. However, the data from these technologies was not
available for the complete set of SIDER drugs. With different
choices in components 1 and 2, the criterion in component 3 of
DILIps could be altered to optimize DILI classification using
different consensus approaches instead of RO3. Therefore, the
DILIps is subject to change and improvement when new data,
technology, and knowledge are available.
Development of predictive models for drugs that might cause
DILI in humans has been an active research field, with much of
the work being done using QSARs. However, the DILI labels used
in these studies are from different sources, some focused on case
reports and others developed using text mining. Furthermore, the
methods used to develop the models are also different. Thus, it is
difficult to compare these methods. For example, Greene et al. [14]
Figure 6. Text mining results to associate types of DILI
(columns) with protein targets (rows). The number of co-
occurrences (papers) between a target and a side effect type is
indicated in the cell. In each cell, the total number of reports as well as
the normalized value (shown in parenthesis) is provided. The
normalized value is the ratio of the number of co-occurrence reports
divided by the total number of reports for a protein target.
doi:10.1371/journal.pcbi.1002310.g006
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 9 December 2011 | Volume 7 | Issue 12 | e1002310
developed Derek for Windows (DfW), a knowledge-based expert
system, to predict a drug’s potential to cause DILI using the DILI
classification scheme developed by Pfizer. The system has 56%
overall accuracy with 73% specificity and 46% sensitivity.
Fourches et al. [18] applied text mining for DILI reported in
different species using MEDLINE abstracts, suggesting that the
concordance of liver effects is low (i.e., 39–44%) between different
species. They also developed QSAR models using a text mining
approach to define DILI classification with external prediction
accuracies ranging from 56 to 73%. Very recently, Ekins et al. [19]
developed a Bayesian model based on DILI endpoint from cellular
imaging predictions [7], which gave a concordance of 60%,
sensitivity of 56%, and specificity of 67%.
Development of DILI models for humans is always confronted
by two distinct but related challenges: (1) a comprehensive drug list
with DILI annotation is usually not available, and (2) there is no
authoritative assessment of whether a drug causes DILI or not. In
this study, we compiled three large datasets from our LTKB
project. We used only the drugs of the opposite extremes in DILI
classification (positive or negative in relationship to DILI) by
removing drugs with ambiguous call. The RO3 criterion of
DILIps reached .91% positive predictive value for a combined
drug list from these three literature datasets. We also applied
DILIps for the drugs with ambiguous call and the results are
available from Supplementary Table S6.
The translation of clinical observations to evaluation of drugs
earlier in the drug development pipeline is a goal of translational
medicine [20]. DILI is an endpoint influenced by several
important factors, and it is difficult to adequately predict with a
single model. The SIDER database has collected clinical
observation data (side effects) from drug labels and the scientific
literature, which allows the linkage of disease endpoints and
related symptom profiles. This, in turn, provides an opportunity to
combine drug information and patient information into a unified
prediction method, a focus of this study. The HepSEs provide a
new direction to predict DILI based on the consensus of multiple
clinical endpoints (side effects) using an in silico method.
Elucidation of therapeutic uses, drug targets, and pathways
related to DILI from a systematic perspective is of great use in
drug discovery and pharmacovigilance. By applying the RO3
criterion to the entire drug space defined by DrugBank, we
constructed a DILI landscape in terms of therapeutic and drug
target space. We do acknowledge that the findings from this excise
are dependent on the accuracy in annotation in DrugBank.
We identified two therapeutic categories (i.e., anti-infectives for
systemic use and musculoskeletal system drugs) in which the drugs
have a high risk for causing DILI. This is consistent with the
general understanding that, for example, NSAIDs (a subcategory
of anti-infectives for systemic use) are often associated with DILI
and have been subject to a broad range of studies looking into
drug-specific, therapeutic class-specific, and genetic-specific effects
[21]. Another possibility is that these drugs may have higher
exposure rates; they are widely used by many people over
prolonged periods, which may inadvertently increase the risk of
DILI. The RO3 positive criterion was able to identify most ‘‘bad
actors’’ among NSAIDs including celecoxib, diclofenac, diflunisal,
ibuprofen, leflunomide, and rofecoxib. Most of them are PTGS2
(COX-2) protein inhibitors. This gene is also involved in several
hepatic system pathways such as hepatic system disorder,liver cancer,
and hepatocellular carcinoma. COX (Cyclooxygenase) is an enzyme
that is responsible for formation of important biological mediators
called prostanoids. Pharmacological inhibition of COX can
provide relief from symptoms of inflammation and pain. However,
more and more reports indicated that the selective inhibition
profile of COXs can cause certain serious adverse drug reactions.
A classic example is rofecoxib (brand name VioxxH), which was
withdrawn in 2004 because of the risk of heart attack caused by
selective inhibition of COX-2. Rofecoxib was also associated with
DILI [22]. Another example is lumiracoxib, a selective COX-2
inhibitor developed for the symptomatic treatment of osteoarthritis
and acute pain. Concern over hepatotoxicity has contributed to
the withdrawal or non-approval of lumiracoxib in most major
drug markets worldwide [23]. Therefore, the study of the
relationship between drug target and DILI, such as COX
selectivity and DILI, may provide new insights into DILI at a
molecular level [24].
We also found that DILI drugs often involve multiple targets,
which is often associated with drugs applied in multiple
therapeutic categories [25]. Drugs interacting with multiple targets
are considered ‘‘dirty’’ since they have a potential to initiate
different adverse reactions. On the other hand, these drugs may
also hold the potential to be repositioned for use outside of their
original therapeutic indications. One such example is diclofenac,
which is used to relieve pain, tenderness, swelling and stiffness
caused by osteoarthritis, rheumatoid arthritis, and ankylosing
spondylitis. Diclofenac is labeled with four different ATC codes
(i.e., four different therapeutic uses) and associated with a number
of targets categorized by DrugBank, including prostaglandin G/H
synthase 1 and 2, the cytochrome P450 family (2C18/2E1/2C19/
1A2/2C8/2D6/2C9/3A4/1A1/2B6), the UDP-glucuronosyl-
transferase family (1–1,2B7), prostaglandin G/H synthase 1, etc.
Several case-control studies have been carried out to investigate
the role of polymorphisms in the gene encoding regions of the
aforementioned drug-metabolizing enzymes and transporters to
determine susceptibility to diclofenac-induced hepatotoxicity
[26,27,28,29,30]. Diclofenac has been withdrawn in several
countries due to liver injury and other adverse drug reactions,
including ulcers, bleeding, and ulcerations in the stomach and
intestinal linings [31]. Diclofenac induced liver injury causes a
number of side effect patterns, including cirrhosis, hepatic failure,
hepatic necrosis, hepatitis, jaundice, all of which were included in
our set of 13 HepSEs.
DILI is associated with two distinct but related parameters: drug
properties and patient susceptibility. Some drugs are more likely to
cause DILI, while some patients are more likely to show DILI.
The DILIps is primarily capable of addressing the former
challenge with an aim to enhance DILI identification in drug
discovery. Identifying genetic variations and their associated
protein products that contribute to DILI is another important
research area, but one that requires the costly and time-consuming
collection of samples from large numbers of affected individuals.
Study of the genetic risk factors to DILI or other conditions usually
involves the identification of genes associated with key disease
mechanisms and immunological reactions using genotyping
approaches. The network analysis conducted in this study
connected DILI drugs with pathways and targets and might
contribute to the identification of mechanisms that relate a
patient’s genetic predisposition and DILI. There are a small
number of genetic risk factors identified for DILI, most are
associated with a drug interaction with a specific HLA (human
leukocyte antigen system) polymorphism within the major
histocompatibility complex (MHC) such as lumiracoxib (HLA-
DRB1*15:01) [23], antituberculosis chemotherapy (HLA-
DQB1*02:01) [32], ticlopidine (HLA-A*33:03) [33], ximelagatran
(HLA-DRB1*07:01) [34], flucloxacillin (HLA-B*57:01) [21], and
amoxicillin-clavulanate (HLA-DRB1*15:01) [35]. Other genetic
risk factors such as those involving drug metabolizing enzymes are
exemplified by CYP2C8*4 (diclofenac), CYP2E1*1A (isoniazid),
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 10 December 2011 | Volume 7 | Issue 12 | e1002310
GSTT1-M1 (troglitazone), and UGT2B7*2 (diclofenac) are also
reported [36,37,38].
Drug safety is a key area of focus in the FDA. Modernizing
safety evaluation has been advocated by the FDA’s recent initiative
on advancing regulatory science with a proposal of incorporating
both in vitro and in silico methodologies in drug development and
safety assessment [39]. The DILIps follows the same philosophy
that underlies this new initiative at the FDA. It could be a
predictive system for FDA to utilize and reference when
hepatotoxicity issues arise during the various stages of the
regulatory review process. It could also serve as a proof-of-concept
approach of using predictive systems for drug safety to support the
FDA’s regulatory science. While the DILIps was developed for
DILI, its methodology can be applied equally well to address other
drug safety issues, such as renal and cardiovascular toxicity.
Materials and Methods
Preparation of datasets
SIDER database. SIDER is computer-readable database of
side effects which connects 888 drugs with 1450 different side
effect terms [12]. The side effects were extracted from drug labels
in either Structured Product Labeling (SPL) or Portable Document
Format (PDF) documents. The standardized Coding Symbols for a
Thesaurus of Adverse Reaction Terms (COSTART), a part of the
Unified Medical Language System (UMLS) Metatheasaurus, was
used as the basic lexicon of side effects. In this study, we
downloaded the entire database from http://sideeffects.embl.de/.
We then constructed a matrix with 888 drugs corresponding to
1450 side effects with supplementing the chemical structure data.
DrugBank. DrugBank (http://www.drugbank.ca) is a richly
annotated database of drugs and drug target information [40,41]. It
contains extensive information about nomenclature, chemistry,
structure, function, mode of action, pharmacology, pharma-
cokinetics, metabolism, and pharmaceutical properties of both
small molecule and large molecule (biotech) drugs. The updated
DrugBank 3.0 contains 6,800 drug entries including 1,400 FDA-
approved small molecule drugs, 132 FDA-approved biotech
(protein/peptide) drugs, 82 nutraceuticals and 5,200 experimental
drugs. In additional, more than 4,300 non-redundant protein (i.e.
drug target) sequences are linked to these drug entries [42].
In this study, information about 6620 small molecule drugs
(1,400 FDA drugs and 5,200 experimental drugs) was retrieved
including chemical structure, approval status, therapeutic catego-
ries and protein targets for use to generate the DILI landscape in
terms of therapeutic uses and drug targets.
LTKB benchmark dataset (LTKB-BD). As a part of the
LTKB project, a research team from the FDA’s National
Center for Toxicological Research has developed the LTKB-
BD dataset that contains 287 drugs with DILI annotation based
on the FDA-approved drug labels. The data are available
from http://www.fda.gov/ScienceResearch/BioinformaticsTools/
LiverToxicityKnowledgeBase/ucm226811.htm [13]. The drugs are
classified into three categories: those of Most-DILI-Concern, Less-
DILI-Concern, and No-DILI-Concern. In this study, only those in
the Most-DILI-Concern (gemtuzumab was excluded since it is a
biotechnology product) and No-DILI-Concern categories were
used. The dataset was divided into two sets. One set overlapped with
the SIDER database and contained 69 drugs of Most-DILI-
Concern and 59 No-DILI-Concern. This was used to evaluate the
performance of HepSEs. The rest of the LTKB-BD contained 67
drugs of Most-DILI-Concern and 6 of No-DILI-Concern that were
not in SIDER and were used to validate the DILIps performance
(Supplementary Table S6).
Pfizer hepatotoxicity dataset (PfizerData). Another
independent test set comes from part of the Derek for Windows
(DfW) system [14], which is a knowledge-based expert system
designed to assess the potential toxicity of a chemical from its
structure. A total of 626 compounds were classified into four
categories based on case reports, including evidence of human
hepatotoxicity (HH), no evidence of hepatotoxicity in any species
(NE), weak evidence (,10 case reports) of human hepatotoxicity
and evidence for animal hepatotoxicity but not tested in humans.
In this study, only HH and NE drugs were used, As a result, there
were 406 drugs remaining; 168 HH (positive) and 90 NE (negative)
overlapped with the SIDER database. The other 92 HH (positive)
and 56 NE (negative) that were not contained in the SIDER
database were selected as another independent test set
(Supplementary Table S6).
O’Brien et al. dataset. O’Brien et al. classified drugs into
four categories according to the severity of human hepatotoxicity
based on the frequency of an observed increase in ALT and other
evidence [15]. In this study, the categories of ‘‘Severely’’ and
‘‘Moderately’’ hepatotoxic drugs were considered DILI positive
drugs while non-toxic drugs were considered DILI negative, and
those that did not overlap with the SIDER database were
employed. The ratio of positive to negative drugs was 25/15
(Supplementary Table S6).
Data analysis method
Identification and assessment of Hepatotoxic Side Effects
(HepSEs). This section is shown as step 1 in Figure 1. There are
1450 different side effects listed in the SIDER database. We
identified 473 side effects for HepSE identification, with each side
effect associated with more than 20 drugs. We used MedDRA to
identify HepSEs. MedDRA is an ontology that provides a
controlled vocabulary describing adverse events. The 473 side
effect terms were mapped to the System Organ Class (SOC) level
of hepatobiliary disorders in MedDRA to extract the HepSEs
(Supplementary Table S1). The drugs in LTKB-BD which
overlapped with drugs in SIDER (128 total drugs) as well as
those in PfizerData which overlapped with drugs in SIDER (258
total drugs) were employed to assess the performance of HepSEs.
If a drug was associated with any HepSE as observed in the
SIDER database, it was considered as DILI positive. To determine
if the predictive performance of the 13 HepSE models was better
than would be expected by chance alone we randomly selected a
set of Mside effects (M= 3,5,…,21) and used these to predict DILI
potential. The selection process for each Mwas repeated 20,000
times, and the average performance of each Mwas compared to
the performance of the 13 HepSEs.
DILI prediction system (DILIps). Development of DILIps
consists of two steps (steps 2 and 3 of Figure 1). In step 2, all of the
drugs were transformed into well-established functional class
fingerprints (FCFP_6), structural fingerprint developed by Pipeline
Pilot 8.0 from SciTegic (http://accelrys.com/). It has been shown
in other studies that Bayesian models built using circular
fingerprints work very well in virtual screening tasks
[14,43,44,45,46,47]. Then, multiple-category naı
¨ve Bayesian
classifiers were trained for each of the selected HepSE
endpoints. In the training set, leave-one-out cross-validation
(LOO-CV) was employed to investigate the model performance.
For each model, a receiver operating characteristic plot (ROC
plot) was drawn to select the best Bayesian score (cut-off value) to
distinguish DILI drugs and non-DILI drugs.
In step 3, the independent test sets were submitted to the 13
HepSE models to calculate the Bayesian scores and give the
prediction results: For each HepSE endpoint, the predicted
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 11 December 2011 | Volume 7 | Issue 12 | e1002310
Bayesian scores (PB-SCORE
i
,i=1, 2n) compared to cut-off
Bayesian score obtained in step 2. If PB-SCORE
i
.cut-off value,
the drug was considered positive for this endpoint and vice versa.
A drug was considered to have the potential to cause DILI if any of
the HepSE endpoints was called as positive (the left side of
Figure 1, step 3). In the right side of Figure 1 (step 3), consensus
prediction strategies were used to investigate the effectiveness of
combining results from multiple HepSEs into a single prediction.
A ‘‘Rule of
N
’’ strategy was evaluated, where 13 separate
consensus prediction strategies were examined with each predict-
ing a drug as causing DILI if
N
(
N
= 1,2,…,13) HepSEs were
positive for that drug.
DILI landscape. Three sets of analysis were conducted,
which is summarized in step 4 of Figure 1. The right side of
Figure 1 is to assess the relationship of therapeutic use and DILI
potential of RO3 drugs. The Anatomical Therapeutic Chemical
(ATC) codes [48] for small molecule drugs which meet the ‘‘Rule
of Three’’ were extracted for this analysis.
The right side of Figure 1 is to assess the association of protein
targets and DILI potential of RO3 drugs. The protein targets
associated with small molecule drugs which meet the RO3
criterion were obtained from DrugBank 3.0. There are 4437
different protein targets from different organisms, and only the
human protein targets were selected. The protein target network
was built by considering two protein targets as connected if at least
one drug was associated with both targets. Two large modules
were identified using the SCAN algorithm, which is used to find
modules in the network [49]. The protein targets in these two large
modules were submitted to Ingenuity Pathway Analysis (IPA)
software (http://www.ingenuity.com/products/pathways_analysis.
html) for pathway analysis. In addition, a text mining with co-
occurrence analysis [50] was also employed to verify the protein
target and HepSE relationship from the network analysis. In this
analysis, the number of papers in PubMed that links a target with a
HepSE in a co-occurrence analysis was extracted. Since some
proteins are more extensively studied than others, the number of
papers associating the protein to the HepSE was normalized by
dividing the number of co-occurrence reports by the total number of
reports of the related protein.
Supporting Information
Figure S1 The distribution of 888 drugs over 473 side
effects and vice versa.
(TIF)
Table S1 Information for the 473 side effects. Informa-
tion included: (1) the number of drugs involved; (2) system organ
classes (SOC) number and annotation of MedDRA; and (3) the
distribution of 473 side effects in the SOC levels of MedDRA.
(XLS)
Table S2 Information of the validation set from three
literature datasets (i.e., LTKB-BD, PfizerData, and
O’Brien et al.) for drugs that do not overlap with SIDER.
(XLS)
Table S3 The top toxicity functions of Module 1.
(XLS)
Table S4 The literature proof about the co-occurrence
between the HepSE terms and protein target. The
EntrezGene ID and PubMed ID (which can be linked to PubMed
directly) are provided.
(XLS)
Table S5 The QSAR models performance for the 13
HepSEs and an additional two terms from the Investi-
gation category, and for 14 HepSEs as suggested by an
expert.
(XLS)
Table S6 Information for datasets. Datasets include: (1)
SIDER, (2) LTKB-BD, (3) PfizerData, (4) O’Brien et al., and (5)
Small molecules in DrugBank.
(XLS)
Author Contributions
Conceived and designed the experiments: ZL WT. Performed the
experiments: ZL QS. Analyzed the data: ZL. Contributed reagents/
materials/analysis tools: ZL WT. Wrote the paper: ZL QS RK HF WT
DD.
References
1. Lee WM, Squires RH, Jr., Nyberg SL, Doo E, Hoofnagle JH (2008) Acute liver
failure: Summary of a workshop. Hepatology 47: 1401–1415.
2. Olson H, Betton G, Robinson D, Thomas K, Monro A, et al. (2000)
Concordance of the toxicity of pharmaceuticals in humans and in animals.
Regul Toxicol Pharmacol 32: 56–67.
3. Kaplowitz N (2005) Idiosyncratic drug hepatotoxicity. Nat Rev Drug Discov 4:
489–499.
4. McKenzie R, Fried MW, Sallie R, Conjeevaram H, Di Bisceglie AM, et al. (1995)
Hepatic failure and lactic acidosis due to fialuridine (FIAU), an investigational
nucleoside analogue for chronic hepatitis B. N Engl J Med 333: 1099–1105.
5. Navarro VJ, Senior JR (2006) Drug-related hepatotoxicity. N Engl J Med 354:
731–739.
6. Shi Q, Hong H, Senior J, Tong W (2010) Biomarkers for drug-induced liver
injury. Expert Rev Gastroenterol Hepatol 4: 225–234.
7. Xu JHJ, Henstock PV, Dunn MC, Smith AR, Chabot JR, et al. (2008) Cellular
imaging predictions of clinical drug-induced liver injury. Toxicol Sci 105:
97–105.
8. Hosomi H, Fukami T, Iwamura A, Nakajima M, Yokoi T (2011) Development
of A Highly Sensitive Cytot oxicity Assay System for CYP3A4-mediated
Metabolic Activation. Drug Metab Dispos 39: 1388–1395.
9. Deng XM, Luyendyk JP, Ganey PE, Roth RA (2009) Inflammatory Stress and
Idiosyncratic Hepatotoxicity: Hints from Animal Models. Pharmacol Rev 61:
262–282.
10. Boelsterli UA, Lim PLK (2007) Mitochondrial abnonnalities - A link to
idiosyncratic drug hepatotoxicity? Toxicol Appl Pharmacol 220: 92–107.
11. Anderson N, Meier T, Borlak Jr. (2009) Toxicogenomics Applied to Cultures of
Human Hepatocytes Enabled an Identification of Novel Petasites hybridus
Extracts for the Treatment of Migraine with Improved Hepatobiliary Safety.
Toxicol Sci 112: 507–520.
12. Kuhn M, Campillos M, Letunic I, Jensen LJ, Bork P (2010) A side effect
resource to capture phenotypic effects of drugs. Mol Syst Biol 6: 6.
13. ChenM, Vijay V, Shi Q, Liu Z, Fang H, et al. (2011) FDA-approved drug labeling
for the study of drug-induced liver injury. Drug Discov Today 16: 697–703.
14. Greene N, Fisk L, Naven RT, Note RR, Patel ML, et al. (2010) Developing
Structure-Activity Relationships for the Prediction of Hepatotoxicity. Chem Res
Toxicol 23: 1215–1222.
15. O’Brien PJ, Irwin W, Diaz D, Howard-Cofield E, Krejsa CM, et al. (2006) High
concordance of drug-induced human hepatotoxicity with in vitro cytotoxicity
measured in a novel cell-based model using high content screening. Arch
Toxicol 80: 580–604.
16. Yildirim MA, Goh KI, Cusick ME, Barabasi AL, Vidal M (2007) Drug-target
network. Nat Biotechnol 25: 1119–1126.
17. Hopkins AL (2009) Drug Discovery: Predicting promiscuity. Nature 462:
167–168.
18. Fourches D, Barnes JC, Day NC, Bradley P, Reed JZ, et al. (2010)
Cheminformatics Analysis of Assertions Mined from Literature That Describe
Drug-Induced Liver Injury in Different Species. Chem Res Toxicol 23: 171–183.
19. Ekins S, Williams AJ, Xu JJ (2010) A Predictive Ligand-Based Bayesian Model
for Human Drug-Induced Liver Injury. Drug Metab Dispos 38: 2302–2308.
20. (2009) Translational medicine. Nat Rev Drug Discov 8: 593–593.
21. Daly AK, Day CP (2009) Genetic Association Studies in Drug-Induced Liver
Injury. Semin Liver Dis 29: 400–411.
22. Yan B, Leung Y, Urbanski SJ, Myers RP (2006) Rofecoxib-induced
hepatotoxicity: A forgotten complication of the coxibs. Can J Gastroen terol
20: 351–355.
23. Singer JB, Lewitzky S, Leroy E, Yang F, Zhao XJ, et al. (2010) A genome-wide
study identifies HLA alleles associated with lumiracoxib-related liver injury. Nat
Genet 42: 711–U794.
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 12 December 2011 | Volume 7 | Issue 12 | e1002310
24. Aithal GP (2011) Hepatotoxicity related to antirheumatic drugs. Nat Rev
Rheumatol 7: 139–150.
25. Nacher JC, Schwartz JM (2008) A global view of drug-therapy interactions.
BMC Pharmacol 8: 5–13.
26. Daly AK, Donaldson PT, Bhatnagar P, Shen YF, Pe’er I, et al. (2009) HLA-
B(star)5701 genotype is a major determinant of drug-induced liver injury due to
flucloxacillin. Nat Genet 41: 816–U871.
27. Thibaudeau J, Lepine J, Tojcic J, Duguay Y, Pelletier G, et al. (2006)
Characterization of common UGT1A8, UGT1A9, and UGT2B7 variants with
different capacities to inactivate mutagenic 4-hydroxylated metabolites of
estradiol and estrone. Cancer Res 66: 125–133.
28. Duguay Y, Baar C, Skorpen F, Guillemette C (2004) A novel functional
polymorphism in the uridine diphosphate-glucuronosyltransferase 2B7 promoter
with significant impact on promoter activity. Clin Pharmacol Ther 75: 223–233.
29. Daly AK, Aithal GP, Leathart JBS, Swainsbury RA, Dang TS, et al. (2007)
Genetic susceptibility to diclofenac-induced hepatotoxicity: Contribution of
UGT2B7, CYP2C8, and ABCC2 genotypes. Gastroenterology 132: 272–281.
30. Jensen LE, Wall AM, Cook M, Hoess K, Thorn CF, et al. (2004) A common
ABCC2 promoter polymorphism is not a determinant of the risk of spina bifida.
Birth Defects Res Part A 70: 396–399.
31. Morison WL, Baughman RD, Day RM, Forbes PD, Hoenigsmann H, et al.
(1998) Consensus workshop on the toxic effects of long-term PUVA therapy.
Arch Dermatol 134: 595–598.
32. Sharma SK, Balamurugan A, Saha PK, Pandey RM, Mehra NK (2002)
Evaluation of clinical and immunogenetic risk factors for the development of
hepatotoxicity during antituberculosis treatment. Am J Respir Crit Care Med
166: 916–919.
33. Hirata K, Takagi H, Yamamoto M, Matsumoto T, Nishiya T, et al. (2008)
Ticlopidine-induced hepatotoxicity is associated with specific human leukocyte
antigen genomic subtypes in Japanese patients: a preliminary case-control study.
Pharmacogenomics J 8: 29–33.
34. Kindmark A, Jawaid A, Harbron CG, Barratt BJ, Bengtsson OF, et al. (2008)
Genome-wide pharmacogenetic investigation of a hepatic adverse event without
clinical signs of immunopathology suggests an underlying immune pathogenesis.
Pharmacogenomics J 8: 186–195.
35. O’Donohue J, Oien KA, Donaldson P, Underhill J, Clare M, et al. (2000) Co-
amoxiclav jaundice: clinical and hist ological features and HLA class II
association. Gut 47: 717–720.
36. Russmann S, Jetter A, Kullak-Ublick GA (2010) Pharmacogenetics of Drug-
Induced Liver Injury. Hepatology 52: 748–761.
37. Andrade RJ, Robles M, Ulzurrun E, Lucena MI (2009) Drug-induced liver
injury: insights from genetic studies. Pharmacogenomics 10: 1467–1487.
38. Wilke RA, Lin DW, Roden DM, Watkins PB, Flockhart D, et al. (2007)
Identifying genetic risk factors for serious adverse drug reactions: current
progress and challenges. Nat Rev Drug Discov 6: 904–916.
39. Hamburg MA (2011) Advancing Regulatory Science. Science 331: 987.
40. Wishart DS, Knox C, Guo AC, Cheng D, Shrivastava S, et al. (2008) DrugBank:
a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res 36:
D901–D906.
41. Wishart DS, Knox C, Guo AC, Shrivastava S, Hassanali M, et al. (2006)
DrugBank: a comprehensive resource for in silico drug discovery and
exploration. Nucleic Acids Res 34: D668–D672.
42. Knox C, Law V, Jewison T, Liu P, Ly S, et al. (2011) DrugBank 3.0: a
comprehensive resource for ‘Omics’ research on drugs. Nucleic Acids Res 39:
D1035–D1041.
43. Bender A, Scheiber J, Glick M, Davies JW, Azzaoui K, et al. (2007) Analysis of
pharmacology data and the prediction of adverse drug reactions and off-target
effects from chemical structure. Chemmedchem 2: 861–873.
44. Crisman TJ, Parker CN, Jenkins JL, Scheiber J, Thoma M, et al. (2007)
Understanding false positives in reporter gene assays: in silico chemogenomics
approaches to prioritize cell-based HTS data. J Chem Inf Model 47: 1319–1327.
45. Glen RC, Bender A, Arnby CH, Carlsson L, Boyer S, et al. (2006) Circular
fingerprints: Flexible molecular descriptors with applications from physical
chemistry to ADME. Idrugs 9: 199–204.
46. Nettles JH, Jenkins JL, Bender A, Deng Z, Davies JW, et al. (2006) Bridging
chemical and biological space: ‘‘Target fishing’’ using 2D and 3D molecular
descriptors. J Med Chem 49: 6802–6810.
47. Nidhi, Glick M, Davies JW, Jenkins JL (2006) Prediction of biological targets for
compounds using multiple-category Bayesian models trained on chemogenomics
databases. J Chem Inf Model 46: 1124–1133.
48. (2011) ATC The Anatomical Therapeutic Chemical (ATC) classification. WHO
Collaborating Centre for Drug Statistics and Methodology. Available: http://
www.whocc.no/atcddd/.
49. Xu XYN, Feng Z, Schweiger TA (2007) SCAN: a structrural clustrering
algorithm for networks. In Proceedings of the 13th ACM SIGKDD international
conference on Knowledge Discovery and Data Mining. San Jose, California,
USA. pp 824–833.
50. Cohen AM, Hersh WR, Dubay C, Spackman K (2005) Using co-occurrence
network structure to extract synonymous gene and protein names from
MEDLINE abstracts. BMC Bioinformatics 6: 103.
Drug-Induced Liver Injury Prediction System
PLoS Computational Biology | www.ploscompbiol.org 13 December 2011 | Volume 7 | Issue 12 | e1002310

Supplementary resources (7)

... Therefore, statistical models are the most common. Approximately 75% of the existing predictive models for liver toxicity have been developed using an array of different statistical methodologies, including discriminant analysis [26], Bayesian models [27,28], Artificial Neutral Networks (ANN) [26], k-Nearest Neighbor (kNN) [29,30], Random Forest (RF) [29,31], Decision Tree (DT), Support Vector Machine (SVM) [32], and specialist QSAR software [33]. ...
... Liu et al. [28] utilized the clinical and postmarketing data from the computer-readable side effect resource (SIDER) database [70] and identified 13 types of hepatotoxic side effects (HepSEs) based on MedDRA ontology, including bilirubinemia, cholecystitis, cholelithiasis, cirrhosis, elevated liver function tests, hepatic failure, hepatic necrosis, hepatitis, hepatomegaly, jaundice, liver disease, fatty liver and liver function test abnormalities [28]. Firstly, these 13 side effects were used to discriminate drugs that do and do not cause DILI using the Liver Toxicity Knowledge Base Benchmark Dataset (LTKB-BD) [71] and the PfizerData [37]. ...
... Liu et al. [28] utilized the clinical and postmarketing data from the computer-readable side effect resource (SIDER) database [70] and identified 13 types of hepatotoxic side effects (HepSEs) based on MedDRA ontology, including bilirubinemia, cholecystitis, cholelithiasis, cirrhosis, elevated liver function tests, hepatic failure, hepatic necrosis, hepatitis, hepatomegaly, jaundice, liver disease, fatty liver and liver function test abnormalities [28]. Firstly, these 13 side effects were used to discriminate drugs that do and do not cause DILI using the Liver Toxicity Knowledge Base Benchmark Dataset (LTKB-BD) [71] and the PfizerData [37]. ...
Chapter
In this chapter, we review the state of the art of predicting human hepatotoxicity using in silico techniques. There has been significant progress in this area over the past 20 years but there are still some challenges ahead. Principally, these challenges are our partial understanding of a very complex biochemical system and our ability to emulate that in a predictive capacity. Here, we provide an overview of the published modeling approaches in this area to date and discuss their design, strengths and weaknesses. It is interesting to note the diversity in modeling approaches, whether they be statistical algorithms or evidenced-based approaches including structural alerts and pharmacophore models. Irrespective of modeling approach, it appears a common theme of access to appropriate, relevant, and high-quality data is a limitation to all and is likely to continue to be the focus of future research.
... It is shown that SCAN can help find meaningful clustering results in biological data [12,18,20,21], web data [17,22,23,26,28,29], and social networks [16,24]. It can also be used for community detection [24], image segmentation [25], and fraud detection on blockchain data [7]. ...
... Similarly, [ ] needs to be updated asˆ( , ) is also changed with vertex as the key in [ ] (Line 7). After that, it updates the bucket index for vertex by invoking the UpdateBucket procedure (Lines [16][17][18][19][20][21][22][23]. Given the old similarity score and the updated similarity score , it first identifies the bucket * of and bucket * of . ...
Article
Full-text available
Graph clustering is a fundamental data mining task that clusters vertices into different groups. The structural graph clustering algorithm ( SCAN ) is a widely used graph clustering algorithm that derives not only clustering results, but also special roles of vertices like hubs and outliers. In this paper, we consider structural graph clustering on dynamic graphs under Jaccard similarity. The state-of-the-art index-based solution focuses on static graphs and incurs prohibitive update costs to maintain indices. Lately, an efficient approximate dynamic structural graph clustering algorithm Dyn-StrClu under Jaccard similarity is proposed. However, their solution needs to fix input parameters while parameter settings of SCAN usually need to be fine-tuned to achieve good clustering results. Motivated by these limitations, we present a study on devising effective index structures for SCAN algorithm on dynamic graphs. Similar to the state-of-the-art dynamic scheme, our main idea to reduce the time complexity is still by bringing approximation to clustering results. However, our solution does not need to fix the input parameters. To achieve this, our solution includes two key components. The first is to maintain a bottom- k sketch for each vertex so that the similarities of affected vertices can be easily updated. The second key is a bucketing strategy that allows us to update clustering results and roles of vertices efficiently. Our theoretical analysis shows that our proposed algorithm achieves O (log n · log M + m / pf ) expected update cost and guarantees to return approximate clustering results with probability 1 - pf after up to M updates. Extensive experiments show that our solution is up to two orders of magnitude faster than the state-of-the-art index-based solution while still achieving high-quality clustering results.
... [8] Systematic and quantitative studies of adverse side effects have become increasingly important due to rising concerns of drug-induced toxicity and post-marketing withdrawal. [9] Concerns over drug induced hepatotoxicity have risen significantly from 10% to 50% of all acute liver failures. [10] It is time for drug developers to design new and accurate models to assess the hepatotoxic effect before costly human clinical trials. ...
Article
Hepatotoxicity implies chemical-driven liver damage. More than 900 drugs have been implicated in causing liver injury and it is the most common reason for a drug to be withdrawn from market. Drug-induced liver injury is responsible for 5% of all hospital admissions and 50% of all acute liver failures. More than 75 percent of cases of idiosyncratic drug reactions result in liver transplantation or death. Systematic and quantitative studies of hepatotoxicity have become increasingly important due to rising concerns of drug induced hepatotoxicity. Drugs frequently interact with more than one target, with hundreds of these targets linked to the side effects of clinically used therapeutics. This is based on the hypothesize that drugs with same side effects are likely to have similar targets. Developing a computational model to predict drug induced hepatotoxicity will provide a screening tool for hepatotoxicity thereby minimizing the number of hepatotoxic drugs released to the market. The study was aimed at exploring the various molecular and structural mechanisms for drug induced hepatotoxicity using computer simulation techniques; pharmacophore studies, PASSONLINE target identification and molecular docking simulation techniques. Hydrogen bond acceptor and hydrophobic interactions were the features common to hepatotoxic drugs, macrophage stimulating factor 1 and glutamate dehydrogenase were the common hepatotoxic targets. The hepatotoxic drugs demonstrated excellent binding affinities against the common targets and superimpose with each other and the co-crystalized ligand in the active pocket of each of the targets. These findings imply that hepatotoxic drugs potentiate the effects of these targets and might be molecular mechanism responsible for the hepatotoxic associated with drugs.
... A pharmaceutical survey suggested a very low correlation was observed between preclinical liver toxicity and clinical outcome with low predictive human values, 50% (monkeys), 27% (dogs), and 33% (rats) for small molecule drugs (Liu et al., 2011). At present, there are various complex in vitro liver models that are being developed, such as micropatterned models, hepatocyte-derived spheroids, 3D bioprinted liver, stem cell derived liver organoids and liver-on-chip models to address this issue (Baudy et al., 2020;Ehrlich et al., 2019;Kanabekova et al., 2022;Deng et al., 2019). ...
Article
Immune responses are heavily involved in the regulation and pathogenesis of human diseases, including infectious diseases, inflammatory and autoimmune conditions, cancer, neurological disorders and cardiometabolic syndromes. The immune system has been considered as a double-edged sword serving as a powerful host defense mechanism against infection and cancerous cells and causing detrimental tissue damage when the immune response is exaggerated or uncontrollable. One of the challenges in studying the efficacy and toxicity of drugs that target or modulate the immune system is the lack of suitable preclinical human models that are predictive of human response. Recent advancements in human microphysiological systems (MPS) have provided a promising in vitro platform to evaluate the response of immune organs ex vivo, to investigate the interaction of immune cells with non-lymphoid tissue cells, and to reduce the reliance on animals in preclinical studies. The development, regulation, trafficking, and responses of immune cells have been extensively studied in preclinical animal models and clinically, providing a wealth of knowledge by which to evaluate new in vitro models. Therefore, the application of immunocompetent MPS in drug discovery and development should first verify that the immune response in an MPS model recapitulates complexity of the human immune physiology. This manuscript reviews biological functions of immune organ systems and tissue-resident immune cells and discusses contexts-of-use for commonly used immunocompetent and immune organ MPS models. Current perspective and recommendations are provided to guide the continued development of immune organ and immunocompetent MPS models and their application in drug discovery and development.
... Machine learning is an alternative analytical approach that can handle complex relationships between variables in very large data sets. Machine-learning algorithms have been used to predict DILI and are superior to traditional statistical models because they handle non-linearity and complex feature interaction (Cruz-Monteagudo et al., 2008;Ekins et al., 2010;Rodgers et al., 2010;Liu et al., 2011;Low et al., 2011;Zhu et al., 2014). However, few studies have investigated the associations between a combination of two drugs and the risk of DILI. ...
Article
Full-text available
Drug-induced liver injury (DILI) is a common adverse drug reaction, with abnormal elevation of serum alanine aminotransferase (ALT). Several clinical studies have investigated whether a combination of two drugs alters the reporting frequency of DILI using traditional statistical methods such as multiple logistic regression (MLR), but this model may over-fit the data. This study aimed to detect a synergistic interaction between two drugs on the risk of abnormal elevation of serum ALT in Japanese adult patients using three machine-learning algorithms: MLR, logistic least absolute shrinkage and selection operator (LASSO) regression, and extreme gradient boosting (XGBoost) algorithms. A total of 58,413 patients were extracted from Nihon University School of Medicine’s Clinical Data Warehouse and assigned to case (N = 4,152) and control (N = 54,261) groups. The MLR model over-fitted a training set. In the logistic LASSO regression model, three combinations showed relative excess risk due to interaction (RERI) for abnormal elevation of serum ALT: diclofenac and famotidine (RERI 2.427, 95% bootstrap confidence interval 1.226–11.003), acetaminophen and ambroxol (0.540, 0.087–4.625), and aspirin and cilostazol (0.188, 0.135–3.010). Moreover, diclofenac (adjusted odds ratio 1.319, 95% bootstrap confidence interval 1.189–2.821) and famotidine (1.643, 1.332–2.071) individually affected the risk of abnormal elevation of serum ALT. In the XGBoost model, not only the individual effects of diclofenac (feature importance 0.004) and famotidine (0.016), but also the interaction term (0.004) was included in important predictors. Although further study is needed, the combination of diclofenac and famotidine appears to increase the risk of abnormal elevation of serum ALT in the real world.
... Two thousand eight hundred and seventy-five compounds were obtained from 4 DILI-related databases, and 6067 compounds were from 11 datasets in published DILI-related literatures [23][24][25][26][27][28][29][30][31][32][33]. After excluding the duplicate compounds, 2365 compounds was obtained. ...
Article
Full-text available
Backgrounds Traditional Chinese medicine and Western medicine combination (TCM-WMC) increased the complexity of compounds ingested. Objective To develop a method for screening hepatotoxic compounds in TCM-WMC based on chemical structures using artificial intelligence (AI) methods. Methods Drug-induced liver injury (DILI) data was collected from the public databases and published literatures. The total dataset formed by DILI data was randomly divided into training set and test set at a ratio of 3:1 approximately. Machine learning models of SGD (Stochastic Gradient Descent), kNN (k-Nearest Neighbor), SVM (Support Vector Machine), NB (Naive Bayes), DT (Decision Tree), RF (Random Forest), ANN (Artificial Neural Network), AdaBoost, LR (Logistic Regression) and one deep learning model (deep belief network, DBN) were adopted to construct models for screening hepatotoxic compounds. Result Dataset of 2035 hepatotoxic compounds was collected in this research, in which 1505 compounds were as training set and 530 compounds were as test set. Results showed that RF obtained 0.838 of classification accuracy (CA), 0.827 of F1-score, 0.832 of Precision, 0.838 of Recall, 0.814 of area under the curve (AUC) on the training set and 0.767 of CA, 0.731 of F1, 0.739 of Precision, 0.767 of Recall, 0.739 of AUC on the test set, which was better than other eight machine learning methods. The DBN obtained 82.2% accuracy on the test set, which was higher than any other machine learning models on the test set. Conclusion The DILI AI models were expected to effectively screen hepatotoxic compounds in TCM-WMC.
... In this respect, it remains challenging to detect and predict chemical-induced liver toxicity. Indeed, standard animal studies conducted during routine drug development usually pick up about half of all human hepatotoxic compounds, while human-based in vitro testing identifies up to 60% of the in vivo human hepatotoxic drugs (Liu et al., 2011). Adverse outcome pathways (AOPs) are promising tools in that regard, as they may help to predict chemical-induced liver toxicity in a more accurate and mechanistically-anchored way. ...
... In this respect, it remains challenging to detect and predict chemical-induced liver toxicity. Indeed, standard animal studies conducted during routine drug development usually pick up about half of all human hepatotoxic compounds, while human-based in vitro testing identifies up to 60% of in vivo human hepatotoxic drugs (Liu et al., 2011). ...
Article
Full-text available
The workshop titled "Application of evidence-based methods to construct mechanism-driven chemical assessment frameworks" was co-organized by the Evidence-based Toxicology Collaboration and the European Food Safety Authority (EFSA), and hosted by EFSA at its headquarters in Parma, Italy on October 2 and 3, 2019. The goal was to explore integration of systematic review methods with mechanistic evidence evaluation. Participants were invited to work on concrete products to advance the exploration of how evidence-based approaches can support the development and application of adverse outcome pathways (AOP) in chemical risk assessment. The workshop discussions were centered around three related themes: 1) assessing certainty in AOPs, 2) literature-based AOP development and 3) integrating certainty in AOPs and non-animal evidence into decision frameworks. Several challenges, mostly related to methodology, were identified, and largely determined the workshop recommendations. The workshop recommendations included the comparison and potential alignment of processes used to develop AOP and systematic review methodology, including the translation of vocabulary of evidence-based methods to AOP and vice versa, the development and improvement of evidence mapping and text mining methods and tools, as well as a call for a fundamental change in chemical risk and uncertainty assessment methodology if to be conducted based on AOPs and new approach methodologies (NAM). The usefulness of evidence-based approaches for mechanism-based chemical risk assessments was stressed, particularly the potential contribution of the rigor and transparency inherent to such approaches in building stakeholders' trust for implementation of NAM evidence and AOP into chemical risk assessment.
Article
In the previous study, we developed the generalized drug-induced liver injury (DILI) prediction model-ResNet18DNN to predict DILI based on multi-source combined DILI dataset and achieved better performance than that of previously published described DILI prediction models. Recently, we were honored to receive the invitation from the editor to response the Letter to Editor by Liu Zhichao, et al. We were glad that our research has attracted the attention of Liu's team and they has put forward their opinions on our research. In this response to Letter to the Editor, we will respond to these comments.
Chapter
Liver toxicity is a major adverse drug reaction that accounts for drug failure in clinical trials and withdrawal from the market. Therefore, predicting potential liver toxicity at an early stage in drug discovery is crucial to reduce costs and the potential for drug failure. However, current in vivo animal toxicity testing is very expensive and time consuming. As an alternative approach, various machine learning models have been developed to predict potential liver toxicity in humans. This chapter reviews current advances in the development and application of machine learning models for prediction of potential liver toxicity in humans and discusses possible improvements to liver toxicity prediction.
Article
Full-text available
Drug-induced liver injury (DILI) is the most common adverse event causing drug nonapprovals and drug withdrawals. Using drugs as test agents and measuring a panel of cellular phenotypes that are directly linked to key mechanisms of hepatotoxicity, we have developed an in vitro testing strategy that is predictive of many clinical outcomes of DILI. Mitochondrial damage, oxidative stress, and intracellular glutathione, all measured by high content cellular imaging in primary human hepatocyte cultures, are the three most important features contributing to the hepatotoxicity prediction. When applied to over 300 drugs and chemicals including many that caused rare and idiosyncratic liver toxicity in humans, our testing strategy has a true-positive rate of 50–60% and an exceptionally low false-positive rate of 0–5%. These in vitro predictions can augment the performance of the combined traditional preclinical animal tests by identifying idiosyncratic human hepatotoxicants such as nimesulide, telithromycin, nefazodone, troglitazone, tetracycline, sulindac, zileuton, labetalol, diclofenac, chlorzoxazone, dantrolene, and many others. Our findings provide insight to key DILI mechanisms, and suggest a new approach in hepatotoxicity testing of pharmaceuticals.
Article
Introduction Lumiracoxib is a selective COX-2 inhibitor that was developed for the symptomatic treatment of osteoarthritis and acute pain. Concerns over liver injury, primarily at chronic doses >100 mg once daily, led to the withdrawal of lumiracoxib in most major drug markets worldwide. A genome-wide association study was performed to identify genetic markers associated with lumiracoxib-related liver injury. Methods Using DNA (n=10 057) collected during the Therapeutic Arthritis Research and Gastrointestinal Event Trial (TARGET), a multistaged case-control pharmacogenetic study was conducted to identify genetic markers associated with the risk of developing elevations of liver aminotransferases (ALT/AST) during treatment with lumiracoxib. TARGET was a 52-week, gastrointestinal clinical safety outcomes study which demonstrated that lumiracoxib (400 mg once daily) reduced the risk of a definite or probable upper GI ulcer complication by 79% compared to NSAIDs (naproxen 500 mg twice daily and ibuprofen 800 mg three times daily) in non-aspirin treated osteoarthritis patients. Results Several single nucleotide polymorphisms (SNPs) from the MHC class II region on chromosome 6 were found to be associated with liver aminotransferase elevations five times the upper limit of normal (>5×ULN) in patients treated with lumiracoxib (top SNP p=2.8E-10). These findings were replicated in an independent set of patients with elevated aminotransferases (>3×ULN) (top SNP p=4.4E-12). To further refine the association results, HLA genotyping and analysis were conducted and a very strong association was identified (top HLA allele p=6.8E-25). Conclusion The study described here identified a highly significant association between HLA alleles and lumiracoxib-related liver injury. These results offer the potential to improve the safety profile of lumiracoxib by excluding patients at elevated risk for developing liver injury.
Article
This report summarizes the results of a multinational pharmaceutical company survey and the outcome of an International Life Sciences Institute (ILSI) Workshop (April 1999), which served to better understand concordance of the toxicity of pharmaceuticals observed in humans with that observed in experimental animals. The Workshop included representatives from academia, the multinational pharmaceutical industry, and international regulatory scientists. The main aim of this project was to examine the strengths and weaknesses of animal studies to predict human toxicity (HT). The database was developed from a survey which covered only those compounds where HTs were identified during clinical development of new pharmaceuticals, determining whether animal toxicity studies identified concordant target organ toxicities in humans. Data collected included codified compounds, therapeutic category, the HT organ system affected, and the species and duration of studies in which the corresponding HT was either first identified or not observed. This survey includes input from 12 pharmaceutical companies with data compiled from 150 compounds with 221 HT events reported. Multiple HTs were reported in 47 cases. The results showed the true positive HT concordance rate of 71% for rodent and nonrodent species, with nonrodents alone being predictive for 63% of HTs and rodents alone for 43%. The highest incidence of overall concordance was seen in hematological, gastrointestinal, and cardiovascular HTs, and the least was seen in cutaneous HT. Where animal models, in one or more species, identified concordant HT, 94% were first observed in studies of 1 month or less in duration. These survey results support the value of in vivo toxicology studies to predict for many significant HTs associated with pharmaceuticals and have helped to identify HT categories that may benefit from improved methods.
Conference Paper
Network clustering (or graph partitioning) is an important task for the discovery of underlying structures in networks. Many algorithms find clusters by maximizing the number of intra-cluster edges. While such algorithms find useful and interesting structures, they tend to fail to identify and isolate two kinds of vertices that play special roles - vertices that bridge clusters (hubs) and vertices that are marginally connected to clusters (outliers). Identifying hubs is useful for applications such as viral marketing and epidemiology since hubs are responsible for spreading ideas or disease. In contrast, outliers have little or no influence, and may be isolated as noise in the data. In this paper, we proposed a novel algorithm called SCAN (Structural Clustering Algorithm for Networks), which detects clusters, hubs and outliers in networks. It clusters vertices based on a structural similarity measure. The algorithm is fast and efficient, visiting each vertex only once. An empirical evaluation of the method using both synthetic and real datasets demonstrates superior performance over other methods such as the modularity-based algorithms.
Article
A large number of case-control association studies on genetic susceptibility to drug-induced liver injury, involving both candidate gene and genome-wide association approaches, have now been reported. The strongest associations have been observed for human leukocyte antigen (HLA) class I and II genes and N-acetyltransferase 2 (NAT2). The associations with HLA class I and II genes are drug specific, though some apparently unrelated compounds show genetic associations with the same alleles. The underlying mechanism for the HLA association is likely to involve T-cell responses to either drug-protein adducts or to drug alone, but needs further investigation. The NAT2 association relates to liver injury induced by isoniazid, with most published studies finding an increased risk of injury in slow acetylators lacking NAT2 enzyme activity, presumably because of the accumulation of toxic metabolites. Other associations with genes relevant to drug disposition, innate immunity, oxidative stress, and mitochondrial function have also been reported, though these still need to be confirmed by replication in independent cohorts.
Article
Drug-induced liver injury (DILI) is a leading cause of drugs failing during clinical trials and being withdrawn from the market. Comparative analysis of drugs based on their DILI potential is an effective approach to discover key DILI mechanisms and risk factors. However, assessing the DILI potential of a drug is a challenge with no existing consensus methods. We proposed a systematic classification scheme using FDA-approved drug labeling to assess the DILI potential of drugs, which yielded a benchmark dataset with 287 drugs representing a wide range of therapeutic categories and daily dosage amounts. The method is transparent and reproducible with a potential to serve as a common practice to study the DILI of marketed drugs for supporting drug discovery and biomarker development.