6 Mode-of-Action-Guided, Molecular …
111
geometrical parameters covering molecular surfaces and fields as well as parameters
calculated in quantum chemistry.
QSAR approaches have also evolved from 1D to 6D [86, 87] and have the following properties: 1D-QSAR correlates biological activity with physiochemical properties, such as pK a and logP [88]; 2D-QSAR correlates activity with structural patterns,
such as connectivity indices, and 2D-pharmacophores. [89]; 3D-QSAR correlates
activity with non-covalent interaction fields surrounding the ligand (e.g., CoMFA
[90]) or receptor (COMBINE [91]) in an alignment-dependent (CoMFA [90] and
COMBINE [91]) or independent (WHIM [92] and COMPASS [93]) fashion; 4DQSAR adds ensemble sampling of conformation, orientation and protonation state as
the fourth dimension [94, 95]; 5D-QSAR allows for a multiple representation of the
topology of the quasi-atomistic receptor surrogate (i.e. the fifth dimension), leading to
less biased induced-fit models [96]; and 6D-QSAR further allows for the simultaneous consideration of different solvation models (i.e., the sixth dimension), reflecting
varying solvent accessibility [97]. However, all these QSAR strategies either do not
consider ligand-receptor interaction (1D- and 2D-QSAR), or restrain the flexibility
of the ligand and/or receptor by modeling the binding interaction in a predefined grid
box (3D- to 6D-QSAR). It was not until recently that molecular docking and MD
simulation were performed to infer optimal conformations with minimal binding free
energy in order to compute structural protein–ligand interaction fingerprints (SPLIF)
[98] and 3D-D Moments/WHIM descriptors [99], respectively.
To the best of our knowledge, QSAR modeling approaches have given little or no
consideration to the dynamic nature in chemical-target biomacromolecule interactions, leading to limited success of QSAR models in toxicity prediction. Our novel
MoA-guided and molecular modeling-based approach is geared to overcome the
aforesaid drawbacks. Despite being an ongoing effort, this approach is believed to
improve the accuracy and efficiency of predictive toxicology, which is supported by
our preliminary results. There are three aspects of novelty in our approach: (1) precategorized reference chemical libraries organized by their documented MoA/MIE
[50], (2) the generation of >5000 dyPLIDs that give full consideration to the flexibility of both ligands (small chemical molecules) and receptors (toxicity biomacromolecular targets) [51], and (3) the application of machine learning, especially deep
learning, in the development of prediction models [48, 49].
6.5 Conclusion and Future Directions
The motivation for developing the novel approach presented in this chapter was to
improve in silico toxicity characterization and risk assessment of existing chemicals
as well as prediction of adverse biological effects for emerging or novel chemicals
undergoing development. Ultimately, our work may lead to the following outcomes:
(1) a reduction in animal use for toxicity testing, (2) early detection of toxicological
properties, and (3) an increase in the likelihood of launching a sustainable “green”
product without incurring undesirable human health and environmental risks. Specif-
111
geometrical parameters covering molecular surfaces and fields as well as parameters
calculated in quantum chemistry.
QSAR approaches have also evolved from 1D to 6D [86, 87] and have the following properties: 1D-QSAR correlates biological activity with physiochemical properties, such as pK a and logP [88]; 2D-QSAR correlates activity with structural patterns,
such as connectivity indices, and 2D-pharmacophores. [89]; 3D-QSAR correlates
activity with non-covalent interaction fields surrounding the ligand (e.g., CoMFA
[90]) or receptor (COMBINE [91]) in an alignment-dependent (CoMFA [90] and
COMBINE [91]) or independent (WHIM [92] and COMPASS [93]) fashion; 4DQSAR adds ensemble sampling of conformation, orientation and protonation state as
the fourth dimension [94, 95]; 5D-QSAR allows for a multiple representation of the
topology of the quasi-atomistic receptor surrogate (i.e. the fifth dimension), leading to
less biased induced-fit models [96]; and 6D-QSAR further allows for the simultaneous consideration of different solvation models (i.e., the sixth dimension), reflecting
varying solvent accessibility [97]. However, all these QSAR strategies either do not
consider ligand-receptor interaction (1D- and 2D-QSAR), or restrain the flexibility
of the ligand and/or receptor by modeling the binding interaction in a predefined grid
box (3D- to 6D-QSAR). It was not until recently that molecular docking and MD
simulation were performed to infer optimal conformations with minimal binding free
energy in order to compute structural protein–ligand interaction fingerprints (SPLIF)
[98] and 3D-D Moments/WHIM descriptors [99], respectively.
To the best of our knowledge, QSAR modeling approaches have given little or no
consideration to the dynamic nature in chemical-target biomacromolecule interactions, leading to limited success of QSAR models in toxicity prediction. Our novel
MoA-guided and molecular modeling-based approach is geared to overcome the
aforesaid drawbacks. Despite being an ongoing effort, this approach is believed to
improve the accuracy and efficiency of predictive toxicology, which is supported by
our preliminary results. There are three aspects of novelty in our approach: (1) precategorized reference chemical libraries organized by their documented MoA/MIE
[50], (2) the generation of >5000 dyPLIDs that give full consideration to the flexibility of both ligands (small chemical molecules) and receptors (toxicity biomacromolecular targets) [51], and (3) the application of machine learning, especially deep
learning, in the development of prediction models [48, 49].
6.5 Conclusion and Future Directions
The motivation for developing the novel approach presented in this chapter was to
improve in silico toxicity characterization and risk assessment of existing chemicals
as well as prediction of adverse biological effects for emerging or novel chemicals
undergoing development. Ultimately, our work may lead to the following outcomes:
(1) a reduction in animal use for toxicity testing, (2) early detection of toxicological
properties, and (3) an increase in the likelihood of launching a sustainable “green”
product without incurring undesirable human health and environmental risks. Specif-