14 Predictive Modeling of Tox21 Data
291
can lead to undesirable side effects and other liabilities [26]. CYP3A4 and CYP2D6
(the two most significant cytochrome P450 (CYP) isozymes among the CYP family
that are essential for drug metabolism) are also found within the set of 58 targets
[27]. This selected set of targets contains several metabolic pathways, a number of
cancer pathways, disease pathways, stress response pathways, and other signaling
pathways. Most of these targets/pathways, including the GPCRs and CYPs, are not
part of the current Tox21 suite of assays. This set of 58 targets can serve as a guide
for assay development in order to generate in vitro data that can better predict human
toxicity.
14.3.3 The Tox21 Data Challenge—New Methods for Data
Modeling
The high-quality concentration response datasets generated by the Tox21 program on
a wide spectrum of pathways and phenotypic toxicity endpoints provide a valuable
resource for predictive toxicity modeling. These data can not only serve as in vitro
signatures that could be used to predict in vivo toxicity endpoints [17, 19, 28–30] and
to prioritize chemicals for more in depth toxicity testing [31] that help to fulfill the
Tox21 goals, but also serve as a knowledge base to correlate chemical structures to
their biological activities for the QSAR (quantitative structure–activity relationship)
modeling community to build more robust models [24, 32]. There is no human
exposure and/or hazard data for 95% of the >80,000 chemicals registered for use in
the USA to inform society about their potential toxic effects [33]. In silico approaches,
such as QSAR models that infer biological activity from chemical structure similarity,
provide a viable alternative to fill in the experimental data gap [34, 35].
To encourage the mining and usage of the Tox21 data, NCATS launched the
Tox21 Data Challenge 2014 [36], to “crowdsource” data analysis by independent
researchers to obtain new models and methods that can predict the potential toxicity of compounds by disrupting cellular and biochemical pathways using chemical
structure data. The Tox21 10K qHTS data from 12 assays, seven nuclear receptor
and five stress response pathway assays, were selected based on data quality and hit
rate for the challenge. The competition attracted 125 participants representing 18
different countries, with 378 model submissions from 40 teams received for final
evaluation.
The winning models all achieved >80% accuracy (Fig. 14.3). Several models
exceeded 90% accuracy. High-quality winning models serve as a confirmation of the
ability of computational approaches to provide meaningful predictions of toxicity
responses in terms of pathway disruption upon environmental compound exposure,
and also as a validation of the quality of datasets produced from the Tox21 qHTS
assays. Consensus models constructed by combining the individual models from all
participating teams resulted in improved predictive performance, with some outperforming the winning models, showing the wisdom of the crowd. The winning models
291
can lead to undesirable side effects and other liabilities [26]. CYP3A4 and CYP2D6
(the two most significant cytochrome P450 (CYP) isozymes among the CYP family
that are essential for drug metabolism) are also found within the set of 58 targets
[27]. This selected set of targets contains several metabolic pathways, a number of
cancer pathways, disease pathways, stress response pathways, and other signaling
pathways. Most of these targets/pathways, including the GPCRs and CYPs, are not
part of the current Tox21 suite of assays. This set of 58 targets can serve as a guide
for assay development in order to generate in vitro data that can better predict human
toxicity.
14.3.3 The Tox21 Data Challenge—New Methods for Data
Modeling
The high-quality concentration response datasets generated by the Tox21 program on
a wide spectrum of pathways and phenotypic toxicity endpoints provide a valuable
resource for predictive toxicity modeling. These data can not only serve as in vitro
signatures that could be used to predict in vivo toxicity endpoints [17, 19, 28–30] and
to prioritize chemicals for more in depth toxicity testing [31] that help to fulfill the
Tox21 goals, but also serve as a knowledge base to correlate chemical structures to
their biological activities for the QSAR (quantitative structure–activity relationship)
modeling community to build more robust models [24, 32]. There is no human
exposure and/or hazard data for 95% of the >80,000 chemicals registered for use in
the USA to inform society about their potential toxic effects [33]. In silico approaches,
such as QSAR models that infer biological activity from chemical structure similarity,
provide a viable alternative to fill in the experimental data gap [34, 35].
To encourage the mining and usage of the Tox21 data, NCATS launched the
Tox21 Data Challenge 2014 [36], to “crowdsource” data analysis by independent
researchers to obtain new models and methods that can predict the potential toxicity of compounds by disrupting cellular and biochemical pathways using chemical
structure data. The Tox21 10K qHTS data from 12 assays, seven nuclear receptor
and five stress response pathway assays, were selected based on data quality and hit
rate for the challenge. The competition attracted 125 participants representing 18
different countries, with 378 model submissions from 40 teams received for final
evaluation.
The winning models all achieved >80% accuracy (Fig. 14.3). Several models
exceeded 90% accuracy. High-quality winning models serve as a confirmation of the
ability of computational approaches to provide meaningful predictions of toxicity
responses in terms of pathway disruption upon environmental compound exposure,
and also as a validation of the quality of datasets produced from the Tox21 qHTS
assays. Consensus models constructed by combining the individual models from all
participating teams resulted in improved predictive performance, with some outperforming the winning models, showing the wisdom of the crowd. The winning models