NEW HORIZONS IN TOXICITY PREDICTION Sponsored by: LHASA Limited: Session Chairs: Professor
Robert Glen, Dr Alan
Wilson, Lexicon Genetics Incorporated Dr
David Hawkins, Lhasa Limited Board of Directors DAY ONE: 9:30 am Welcome and Opening
Remarks Session 1: Current Approaches for Toxicity Prediction
Current Status of Toxicity Prediction
What is currently available?
How are these tools being used by Regulators,
Industry & Academia ?
Unmet needs? 9:45 am Opening
Remarks Dr David Hawkins (Chair) 9:50 am Pharmaceutical
Regulatory Perspective
10:25 am
In Silico Tools & Guidance developed by the European Chemicals Bureau 11:00 am Break 11:30 am Modelling and informatics support for safety an metabolism studies in early drug discovery projects Dr Scott Boyer, Astra Zeneca 12:05
am Recent Developments in Toxico-Cheminformatics: Supporting a New Paradigm for Predictive Toxicology
Dr Ann Richard, US
Environmental Protection Agency 12:40 pm Lunch Session 2: Strengths and Limitations of
Current Toxicity Prediction Systems
How do different approaches compare (QSAR, Knowledge-Based,
Neural Networks)
Pros and Cons of different systems
Toxicity databases and tool boxes 2:00 pm Opening Remarks Professor Robert Glen (Chair) 2:10 pm The Role of
Predictive Data Mining in Understanding Toxicity Professor Chihae Yang, Leadscope Inc. 2:45 pm Consensus QSAR Models Dr Mark Cronin, 3:20 pm Break 3:55 pm QSAR Approaches,
Models & Statistics Relating to Toxicity Prediction Professor Douglas Hawkins, 4:30 pm Knowledge-Based Approaches Dr Nigel Greene, Pfizer 5:05 pm Models and Databases
for Genetic/Carcinogenic Toxicity Romualdo Benigni, National Institute of 5:40 pm- End - 7.30pm Symposium
Dinner: Peterhouse: College Campus DAY TWO: Session 3 Emerging Areas and Technologies in Toxicity Prediction
8:30 am Opening Remarks Dr Alan Wilson (Chair) 8:40 am Use of Biosimulation and Toxicogenomics to Predict the Efficacy of
Antiatherogenic Therapies, Cardiovascular Toxicity and Outcomes Dr Héctor De León, Entelos 9:15 am Emerging Areas in Toxicity Prediction: NIHS Perspective Dr Akihiko Hirose,
National Institute of Health Sciences, 9:50 am Coffee Break 10:20am Consumer Products
Industries Perspective Dr Cameron MacKay, Unilever - Safety & Environmental Assurance Centre 10:55 am Challenges In
Predicting Metabolism & Toxicity with Known & Abstract Targets Dr Fred Guengerich, Vanderbilt University 11:30 am Panel
Discussion: Future Needs and Challenges*
Identification of Gaps, Future Needs and Challenges
Increasing application of in silico modeling
Barriers to implementation 1:00 pm Closing Remarks
& End of Conference
Dr David Hawkins & Professor Robert
Glen (Chair)
Dr Edwin J. Matthews, Food & Drug
Administration
Abstract
US FDA and pharmaceutical industry are under increasing political, scientific, and fiscal pressure to modernize the pharmaceutical development and the regulatory review processes (see FDA's Critical Path). Predictive software programs are now recognized as cost-effective tools to facilitate this modernization and to improve upon the safety and efficacy of drug products. The programs are expected to provide rapid and reliable decision support information on the activities of substances detected in pre-clinical toxicology tests; absorption, distribution, metabolism, and excretion (ADME) properties of lead chemicals; and adverse effects observed in human clinical trials and post market surveillance.
Expectations for predictive software program performance are very high; they must have: (1) a high domain of applicability for FDA regulated substances, (2) both highly specific and highly sensitive predictions of activities, (3) a reasonable interpretation of the mode of action (MOA) of predicted activities, (4) measures of model performance through internal and external validation studies, (5) transparency of the prediction paradigm and weight of evidence employed for the training data-set algorithm, and (6) documentation of the data sources employed. These high expectations can best be met when the predictive software utilizes the collective knowledge of toxicology and clinical studies present in both publicly available sources and in proprietary studies in agency and industry archives.
We at the FDA are meeting these objectives through collaborations with a consortium of software developers. Harvesting the knowledge in FDA and EPA archival data is in progress and non-proprietary portions of these data have already been made publicly available in part. Methods to achieve high domain of applicability, specificity, sensitivity, and accurate MOA predictions in validated QSAR models and expert systems have been reported and new improvements are continuing to be realized.
Dr
Abstract
To promote the availability of reliable computer-based estimation methods for the regulatory assessment of chemicals, the European Chemicals Bureau (ECB) within the European Commission's Joint Research Centre (JRC) has been developing a range of user-friendly and publicly accessible software tools.
Toxtree predicts various kinds of toxic effect by applying decision tree approaches. The set of decision trees currently includes the Cramer classification scheme, the Verhaar scheme, the BfR rulebases for irritation and corrosion, and the Benigni-Bossa scheme for mutagenicity and carcinogenicity. Additional rulebases are under development.
Toxmatch generates quantitative measures of chemical similarity. These can be used to compare datasets and to calculate pairwise similarity between compounds. Consequently, Toxmatch can be used to compare model training and test sets, to facilitate the fomration of chemical categories, and to support the application of chemical read-across.
DART (Decision Analysis by Ranking Techniques) was developed to make ranking methods available to scientific researchers. DART is designed to support the ranking of chemicals according to their environmental and toxicological concern and is based on the most recent ranking theories. Different kinds of order ranking methods, roughly classified as total and partial-order ranking methods are implemented.
Finally, the ECB is developing a web-based inventory of (Q)SAR models (the JRC QSAR Model Database) which will help to identify relevant (Q)SARs for chemicals undergoing regulatory review. The JRC QSAR Model Database will provide publicly-accessible information on QSAR models and will enable any developer or proponent of a (Q)SAR model to submit this information by means of a QSAR Model Reporting Format (QMRF). The in silico tools developed by the ECB can be found on the ECB website (http://ecb.jrc.it/qsar/).
Access to metabolism and toxicology data is critical to
effective decision making in early drug discovery projects.
Often in such projects little is known about the therapeutic
target and usually even less is known about potential metabolism
or adverse effects of the chemical series being investigated.
Simply providing unstructured metabolism- and safety-related
information on targets and chemical series to project teams
trying to make decisions is not adequate due to the varied
nature and quality of metabolism and toxicology data. This
presentation gives examples of how relevant data can be
structured, mined and in some cases modelled to enhance
decision-making. Project examples will be presented of QSAR
models and their interpretation, including characterization of
the underlying assay error for better interpretation of the
model results, development of SAR systems that support
decision-making and enhance awareness around such endpoints as
metabolism/P450 activation, mutagenesis, hERG and reactive
intermediates. In general, metabolism and toxicology data should
be structured depending on, 1) its intended use, 2) its overall
quality and 3) its internal data structure (text vs. numerical)
to assure its optimum use. Brief examples of the varying data
types and their usage in project decision making will be
presented along with some strategies for hypothesis generation
around adverse events using a combined approach of molecular
modelling/virtual screening and text mining. Together, these
tools, built to be appropriate to the various data types,
represent a basic toolkit for the toxicologist and drug
metabolism scientist needing to make meaningful contributions to
chemical design decisions made in early drug discovery projects.
Abstract
EPA's National Center for Computational Toxicology is building capabilities to support a new paradigm for toxicity screening and prediction through the harnessing of legacy toxicity data, creation of data linkages, and generation of new high-content and high-thoughput screening data. In association with EPA's ToxCastTM, ToxRef DB, and ACToR projects, the DSSTox project provides cheminformatics support and, in addition, is improving public access to quality structure-annotated chemical toxicity information in less summarized forms than traditionally employed in SAR modeling, and in ways that facilitate data-mining and data read-across. The latest DSSTox version of the Carcinogenic Potency Database file (CPDBAS) illustrates ways in which various summary definitions of carcinogenic activity can be employed in modeling and data mining. The DSSTox Structure-Browser provides structure searchability across all published DSSTox toxicity-related inventory, and is enabling linkages between previously isolated toxicity data resources associated with environmental and industrial chemicals. The public DSSTox inventory also has been integrated into PubChem, allowing a user to take full advantage of PubChem structure-activity and bioassay clustering features. Phase I of the ToxCastTM project is generating high-throughput screening data from several hundred biochemical and cell-based assays for a set of 320 chemicals, mostly pesticide actives with rich toxicology profiles. Incorporating traditional SAR concepts into this new data-rich world poses conceptual and practical challenges, but also holds great promise for improving predictive capabilities. This work was reviewed by EPA and approved for publication, but does not necessarily reflect EPA policy.
Toxicity data from various sources are integrated into a rigorously designed database using the ToxML schema. The public database sources include FDA submission data from approved new drug applications, food contact notifications, generally recognized as safe food ingredients, as well as chemicals from NTP and CCRIS databases. The data from various sources are then combined according to ToxML criteria. The resulting database is used for data mining analysis to investigate, for example, genotoxic and non-genotoxic carcinogens. Data analysis methods include both knowledge building and applying informatics, leading to predictive data mining. Extraction and validation of structural alerts, grouping compounds by mode-of-action driven alerts, and building predictive models within this space using chemical and biological descriptors are discussed.
Abstract
In silico approaches to predicting toxicity can range from the relatively simple to the remarkably complex. At the simplest level are structure-activity relationships (SARs) or the qualitative relationship between a chemical sub-structure and a toxicological event. An understanding of the structural basis for toxicology may allow for the formation of (mechanistically based) categories, and hence allow for read across or extrapolations.
Quantitative structure-activity relationships (QSARs) vary from simplistic linear regression analyses to wholly multivariate and non-linear models such as neural networks. Within all these approaches, there is a capability to form a consensus from predictions both within models (the so-called consensus-QSAR approaches) and from between models. Forming a consensus from populations of (regression-based) QSAR models can have advantages in improving statistical fit by bringing in outliers, but the loss of transparency of models may be viewed as a disadvantage. Particularly exciting for the prediction of toxicity is the formation of a consensus between different modelling approaches, e.g. combination of information from read-across, (Q)SARs, experts systems as well as other information such as in chemico reactivity or in vitro data.
Approaches to predict toxicity, in the context of forming a consensus using weight-of-evidence within integrated testing strategies, will be presented.
Abstract
Quantitative Structure Activity/Property/Toxicity Relationship
modelling has a long history. The first statistical tools used
were multiple linear regression for a ordinal measure of
activity, and logistic regression and linear or quadratic
discriminant analysis for a binary measure of activity. These
methods generally work quite well when n, the number of
available compounds for calibrating the model, is large, and p,
the number of predictors, is small. They do rely on additivity
and linearity assumptions which, with small p, can be diagnosed
and remedied graphically.
More recently, many QSAR data sets have used values of p in the
hundreds to thousands. At this point, traditional graphical
methods of diagnosing interaction and non-linearity are
ineffective and, even if the traditional models can be fitted,
they tend to work poorly because of the impact of collinearity.
Current QSAR methodologies for 'large p' problems
address different deficiencies of the traditional methods and in
different ways. Partial least squares (PLS), ridge regression
(RR), the elastic net (EN) and principal component regression
(PCR) keep the assumption of a linear relationship, but address
the collinearity issues associated with large p, particularly if
coupled with small n. Neural nets and support vector machines
(SVMs) concentrate on curing possible nonlinearities in the
relationship between the dependent and the predictors, but are
impacted by the 'large p' problem. Nearest
neighbours models have minimal assumptions about the nature of
the relationship, but rely on compounds being relatively evenly
spread around in the predictor space. Recursive partitioning
(RP) methods assume neither additivity nor linearity, and can
handle arbitrarily large p values, but require large n.
There is therefore no 'one size fits all'
statistical methodology.
Feature selection - picking out a relatively few
predictors that account for the relationship - is
desirable both for making better predictions and for
understanding underlying mechanisms, but is a potential source
of large biases.
Having fitted a model, one needs to evaluate its usefulness for
predicting the toxicity of as-yet-unstudied compounds whose
chemical structure is known. Unless very large samples are
available for the fitting and verification, this is best done
using some form of cross-validation which allows all available
compounds to be used for both the modelling and the model
assessment steps.
Abstract
The Structure-Activity Relationships paradigm provides a wide
range of tools that can be exploited in the research on, and
regulation of toxic chemicals. Such tools have different degrees
of approximation / uncertainty, and apply to different scopes.
On one side, there are coarse-grain approaches such as the
Structure Alerts (SA) (often accompanied by modulating factors).
Together with being a repository of the science on the chemical
biological interactions at the basis of chemical
carcinogenicity, the SAs have a crucial role in practical
applications for risk assessment, for: a) description of sets of
chemicals; b) preliminary hazard characterization; c) formation
of categories for e.g., regulatory purposes; d) generation of
subsets of congeneric chemicals to be analyzed subsequently with
Quantitative Structure-Activity Relationships (QSAR) methods; e)
priority setting .
On the other side, there are the fine-tuned QSARs for congeneric
classes of chemicals. Recently, a range of good quality, local
QSARs for mutagenicity and carcinogenicity have been assessed in
our laboratory, and challenged for their predictivity in respect
to real external test sets (i.e., chemicals never considered by
the authors while developing their models). The QSARs for
potency (applicable only to toxic chemicals) generated
predictions 30 to 70 % correct, whereas the QSARs for
discriminating between active and inactive chemicals were 70 to
100 % correct in their external predictions: thus the latter can
be used with good reliability for applicative purposes. The same
study showed that internal, statistical validation methods,
which are often assumed to be good diagnostics for predictivity,
did not correlate well with the predictivity of the QSARs when
challenged in external prediction tests.
Non local QSARs, designed for non congeneric sets of chemicals,
showed a high degree of variability supposedly linked mainly to
the definition of the applicability domain.
The crucial role of mechanistic knowledge in the process of
applying Structure-Activity considerations to risk assessment
should be strongly emphasized. Mechanistic knowledge provides a
ground for interaction and dialogue between model developers,
toxicologists and regulators, and permits the integration of the
(Q)SAR results into a wider regulatory framework, where
different types of evidence and data concur or complement each
other as a basis for making decisions and taking actions.
Whenever possible, (Q)SARs with clear links to the underlying
toxicity mechanisms should be preferably applied.
Abstract
Despite the progress made in the treatment of cardiovascular (CV)
disease, the high failure rate of therapies entering late-stage
clinical development and the unexpected appearance of adverse CV
events highlight the need for predictive technologies to complement
drug discovery and clinical trial design. We have developed a unique
set of biosimulation and gene expression profiling tools aimed at the
early identification of effective drug candidates with low toxicity.
We assembled the CV PhysioLab® platform, a large-scale
mathematical model of human lipid metabolism and CV pathology, to
evaluate the potential efficacy of alternate therapeutic approaches.
The model uses differential equations to represent interactions of
cells and biomolecules linked to key CV clinical outcomes (e.g.
myocardial infarction). Finite-element modeling is used to simulate
the temporal changes in the structure of atherosclerotic plaques that
lead to rupture. The structural stability of the plaque can be linked
to an estimated risk of a CV event. Virtual patients and patient
populations are used to represent different pathophysiological
hypotheses and to analyze the impact of phenotypic variability in
response to therapies and drug-induced toxicity. We also developed
DrugMatrix®, an extensive toxicogenomic database of microarray
expression data linked to classic preclinical and clinical toxicology
measurements, to identify predictive gene expression profiles.
Hypotheses generated from these profiles can be simulated in the CV
PhysioLab platform to identify biomarkers predictive of adverse
events. Together, these technologies can significantly increase the
likelihood of identifying effective compounds with low toxicity
profiles early in the drug development process.
Abstract
Huge kinds of chemicals exist in our environment, and the risks
of majority industrial chemicals have not been evaluated. We
urgently need to develop a high throughput evaluation system for
the human risk of these chemicals. The (quantitative) structure
activity relationships ((Q)SAR) approach has possibility to
resolve this issue, but individual QSAR system has not so
powerful to judge administratively in the default settings.
Recently, NIHS has started to develop the flow to assess
chemical genotoxicity in combination of three in silico systems.
We selected DEREK Windows, MultiCASE and ADMEWorks, because of
different modes for analyses. At first, we customized each
system for mutagenisity prediction, i.e., bacterial gene
mutation and in vitro chromosomal aberration assays. In
combination approach, the concordance between in vitro and in
silico assays on bacterial gene mutation reached to around 94%,
although the applicability decreased to 55%. Next, we tried the
similar approach for developing the chromosomal aberration
predicting system. The performance of chromosomal aberration
prediction is still lower than that of bacterial mutagenicity
prediction and the further development are required.
In addition to these genotoxicity studies, repeated dose rat
toxicity studies are commonly used as screening tests for
evaluating the human risk of industrial chemicals. However, no
appropriate in silico general toxicity evaluating system is
available at present. We analyzed the toxicity profiles of
hundreds of the twenty-eight days repeated dose studies. The
hepato- and/or renal-toxicities were selected as major
endpoints. We decided to focus to develop the predicting system
of these two types of toxicity, by searching new sub-structural
alerts for DEREK Windows and using the discriminant-based QSAR
model builder, such as the Leadscope Predictive Data Miner
(Leadscope Inc.).
NIHS also joined to the multi-institutional project of the
developing for the repeated-dose toxicity knowledge-base system,
which could assist toxicological expert judgment, or support
preliminary governmental decisions. The system consists of three
parts, the detailed sub-chronic toxicity studies database, the
toxicity mechanisms database, and the metabolite predicting
system. The project is lead by National Institute of Technology
and Evaluation (NITE), Tohoku University, Kwansei Gakuin
University, Fujitsu Co. Ltd., and NIHS. The project is
financially supported by the New Energy and Industrial
Technology Development Organization (NEDO). The parts of this in
silico knowledge-base system will be integrated to the OECD
(Q)SAR Application Tool Box, and will support the categorical
approach on HPV chemicals evaluation. We are also developing
repeated-dose toxicity (Q)SAR system in future.
Abstract
Assuring the safety of consumer products without the need to conduct animal tests is a considerable challenge. The mouse local lymph node assay (LLNA) is now used widely to generate data for assessing the risk of chemical-induced skin sensitization. However, changes in EU legislation [i.e. 7th Amendment to the EU Cosmetics Directive] have made developing non-animal approaches to provide the data for skin sensitization risk assessment a key business need. In collaboration with Entelos Inc, Unilever has developed a large-scale in silico model (comprised of nonlinear ODEs) of skin sensitization induction to characterise and quantify the contribution of implicated pathways to the overall biological response. Such knowledge is crucial in guiding the development of in vitro assay development for use in consumer safety risk assessment.
The model describes the response in mouse over a 7-day period following exposure to dinitrochlorobenzene (a well known contact allergen) and includes both epidermal and lymph node cellular processes implicated in skin sensitisation such as cytokine responses, cell surface marker regulation, cellular migration and proliferation. In order to populate the model, in vivo and in vitro data from the published literature were used. Some of the data, such as epidermal cytokine release in response to chemical insult, were used to build focused submodels of the biology. Mouse local lymph node assay data (an assay for the identification of contact allergens that covers topical exposure through to measurement of the resulting immune response) were used in order to ensure that, acting together, these submodels could model the full system response effectively.
The modelling uncovered a previously underappreciated pathway in skin sensitisation and showed it to be key to the sensitisation response. Additionally, the modelling revealed a number of gaps in both the current mechanistic knowledge and the available data. Unilever is using the model to focus and guide its future research in the area of skin sensitisation.
Abstract
The costs of drug development and environmental risk assessment continue to increase, and the availability of better in silico and in vitro methods has the potential to yield better and more economical predictions. Metabolism issues are key to some toxicities, e.g. acetaminophen. However, an accurate assessment of the fraction of all drug and other toxicities is due to metabolism is unavailable and, even when metabolism is agreed to be central, the ensuing biological events are not well understood.
A need exists for better biomarkers and assays to predict not only bioactivation but also off-target toxicity, immune-related toxicity, and idiosyncratic reactions. Although much has been learned about the enzymes involved in the metabolism of many drugs and other xenobiotics, predictions are not trivial. Actual protein structures are very preferable to homology models, but even when these are available they often do not predict products accurately.
Issues include non-product complexes and the limitations on obtaining structures of enzymes relevant to the key catalytic steps. Predicting toxicity in silico is difficult in that relatively few molecules are recognized in off-target and other toxicities. Even when known, the structures may still be models (e.g. hERG channel). In most toxicity events, the mechanism of action is unknown at the molecular level and relatively few biological structures can be used. Thus, many aspects of predictive toxicology still involve structure-activity relationships among chemicals, but the issues of relevance of the parameters used in comparisons is critical in judging the usefulness of the analyzer.
(Supported in part by U.S. Public Health Service grants R01 ES010546 and P30 ES000267.)
