6 Mode-of-Action-Guided, Molecular …
109
[73], are employed to build homology models for the target biomacromolecule. The
stereochemical quality of the homology models is assessed using PROCHECK [74],
whereas the local and global model quality is estimated using the QMEAN scoring function [75]. If a homologous protein of known 3D structure cannot be found,
the de novo modeling approach QUARK [76] is employed to produce ab initio a
template-free predicted model for the target of interest.
6.3.3 Molecular Docking and MD Simulation for dyPLID
Generation
With the 3D structure of a target biomacromolecule in hand, FINDSITE [77], a
threading algorithm, or POCKET [78], a cavity detection program, is used to identify
putative active binding sites in the target biomacromolecule if the sites are unknown.
Then, AutoDock Vina [79] is utilized to dock a chemical into the binding sites in the
target biomacromolecule. The top-scoring binding pose with a favorable calculated
binding energy is selected and further refined using Amber18 (http://ambermd.org/),
an MD simulation program package [80, 81]. The obtained MD trajectories and the
VMD program [82] are used to calculate the dyPLIDs as the quantitative measurements for possible target-chemical interactions. The binding energy is recalculated
after MD simulations, and the refined binding energy estimate is included in the set
of target-chemical interaction descriptors.
6.3.4 In Vitro Toxicity Prediction Mode Libraries
We are currently developing toxicity classification and quantification models using
machine learning (including classical SVM, random forest, and deep learning algorithms such as deep neural networks [48, 83]) methods. For instance, using the androgen receptor bioassay dataset for more than ten thousand chemicals made available
through the Tox21 data challenge [84] (https://tripod.nih.gov/tox21/challenge/), we
have built a set of models to classify these chemicals into agonists, antagonists,
inactive ligands, and inconclusive compounds (i.e., neither active nor inactive) (G.
Idakwo et al. Manuscript under review) and to quantitatively predict the degree of
agonism or antagonism for 273 active compounds [51].
6.3.5 Web Portal for ChemMoA/TsTKb
It has been our intention to make the toolkits we developed publicly accessible so that
our research findings and products can be disseminated to the relevant communities
109
[73], are employed to build homology models for the target biomacromolecule. The
stereochemical quality of the homology models is assessed using PROCHECK [74],
whereas the local and global model quality is estimated using the QMEAN scoring function [75]. If a homologous protein of known 3D structure cannot be found,
the de novo modeling approach QUARK [76] is employed to produce ab initio a
template-free predicted model for the target of interest.
6.3.3 Molecular Docking and MD Simulation for dyPLID
Generation
With the 3D structure of a target biomacromolecule in hand, FINDSITE [77], a
threading algorithm, or POCKET [78], a cavity detection program, is used to identify
putative active binding sites in the target biomacromolecule if the sites are unknown.
Then, AutoDock Vina [79] is utilized to dock a chemical into the binding sites in the
target biomacromolecule. The top-scoring binding pose with a favorable calculated
binding energy is selected and further refined using Amber18 (http://ambermd.org/),
an MD simulation program package [80, 81]. The obtained MD trajectories and the
VMD program [82] are used to calculate the dyPLIDs as the quantitative measurements for possible target-chemical interactions. The binding energy is recalculated
after MD simulations, and the refined binding energy estimate is included in the set
of target-chemical interaction descriptors.
6.3.4 In Vitro Toxicity Prediction Mode Libraries
We are currently developing toxicity classification and quantification models using
machine learning (including classical SVM, random forest, and deep learning algorithms such as deep neural networks [48, 83]) methods. For instance, using the androgen receptor bioassay dataset for more than ten thousand chemicals made available
through the Tox21 data challenge [84] (https://tripod.nih.gov/tox21/challenge/), we
have built a set of models to classify these chemicals into agonists, antagonists,
inactive ligands, and inconclusive compounds (i.e., neither active nor inactive) (G.
Idakwo et al. Manuscript under review) and to quantitatively predict the degree of
agonism or antagonism for 273 active compounds [51].
6.3.5 Web Portal for ChemMoA/TsTKb
It has been our intention to make the toolkits we developed publicly accessible so that
our research findings and products can be disseminated to the relevant communities