521
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
estimation of ligand binding [92, 103]. Prediction of accurate binding affinities is
out of the scope of the scoring functions implemented in established docking applications [103]. In contrast to rigorous physics-based approaches aiming at the prediction of free energy of binding, scoring functions are fitting functions designed to
be fast and transferable, thereby allowing high-throughput screening of databases.
Three types of scoring functions can be identified in general: force field-based,
knowledge-based and empirical scoring functions. Classically, they all treat molecular interactions as pair-wise additive terms on atom level aiming to score a
ligand pose at high-throughput speed. In force field-based methods, the energy of
the protein–ligand interaction is covered by van-der-Waals terms as well as electrostatic contributions from molecular mechanics. The ligand’s force field energy
is also included in some scoring functions to consider the conformational strain of
the pose [104]. It has been shown that empirical scoring functions including entropy
penalties for the ligand fixation and desolvation terms are more accurate to estimate
binding affinities [105]. Knowledge-based scoring functions derive the terms for
interaction potential based on experimental geometry distributions. Coefficients are
directly fitted to experimental data to reproduce binding affinities of protein–ligand
complexes with known X-ray structures [106]. For a review on scoring functions
and docking algorithms we recommend the review by Kitchen et al. as an overview
article [98]. Recent advances in scoring functions are the inclusion of non-additive,
cooperative effects, e.g., for hydrogen bond networks established upon ligand binding [107]. Also in the field of solvent effects some progress can be expected by
inclusion of entropic and enthalpic contributions depending on interaction with explicitly investigated water molecules by WaterMap [108, 109].
However, the universal scoring function has not yet been developed and the optimal choice of the best function is project-dependent [103]. We propose to consider
results of unbiased docking challenges which are periodically repeated to reflect the
current state of the technique. One example is the challenge organized by the Community Structure-Activity Resource (CSAR) around Heather Carlson [92]. Focusing on the application to multiple structures, the speed of the docking and scoring
algorithm as well as the capability of the program to handle multiple structures in
parallel might be an issue. When docking to multiple structures derived from an
MD simulation, the major steps remain the same as for docking to rigid structures.
However, several docking programs offer a workflow especially tailored for parallel docking to optimize computing time. For example, in GOLD up to 20 protein
conformations can be processed by the “ensemble docking” feature and allow the
user to weight the scores for individual protein conformations.
15.2.2.2 Preparation of Structures for Molecular Docking
As important as the choice of the program and scoring function is the preparation
of the ligand structures or the database of potential ligands used for a virtual screening [110]. The protein structure must be prepared appropriately as for rigid docking
[50]. The question of proper protonation of protein residues should already have
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
estimation of ligand binding [92, 103]. Prediction of accurate binding affinities is
out of the scope of the scoring functions implemented in established docking applications [103]. In contrast to rigorous physics-based approaches aiming at the prediction of free energy of binding, scoring functions are fitting functions designed to
be fast and transferable, thereby allowing high-throughput screening of databases.
Three types of scoring functions can be identified in general: force field-based,
knowledge-based and empirical scoring functions. Classically, they all treat molecular interactions as pair-wise additive terms on atom level aiming to score a
ligand pose at high-throughput speed. In force field-based methods, the energy of
the protein–ligand interaction is covered by van-der-Waals terms as well as electrostatic contributions from molecular mechanics. The ligand’s force field energy
is also included in some scoring functions to consider the conformational strain of
the pose [104]. It has been shown that empirical scoring functions including entropy
penalties for the ligand fixation and desolvation terms are more accurate to estimate
binding affinities [105]. Knowledge-based scoring functions derive the terms for
interaction potential based on experimental geometry distributions. Coefficients are
directly fitted to experimental data to reproduce binding affinities of protein–ligand
complexes with known X-ray structures [106]. For a review on scoring functions
and docking algorithms we recommend the review by Kitchen et al. as an overview
article [98]. Recent advances in scoring functions are the inclusion of non-additive,
cooperative effects, e.g., for hydrogen bond networks established upon ligand binding [107]. Also in the field of solvent effects some progress can be expected by
inclusion of entropic and enthalpic contributions depending on interaction with explicitly investigated water molecules by WaterMap [108, 109].
However, the universal scoring function has not yet been developed and the optimal choice of the best function is project-dependent [103]. We propose to consider
results of unbiased docking challenges which are periodically repeated to reflect the
current state of the technique. One example is the challenge organized by the Community Structure-Activity Resource (CSAR) around Heather Carlson [92]. Focusing on the application to multiple structures, the speed of the docking and scoring
algorithm as well as the capability of the program to handle multiple structures in
parallel might be an issue. When docking to multiple structures derived from an
MD simulation, the major steps remain the same as for docking to rigid structures.
However, several docking programs offer a workflow especially tailored for parallel docking to optimize computing time. For example, in GOLD up to 20 protein
conformations can be processed by the “ensemble docking” feature and allow the
user to weight the scores for individual protein conformations.
15.2.2.2 Preparation of Structures for Molecular Docking
As important as the choice of the program and scoring function is the preparation
of the ligand structures or the database of potential ligands used for a virtual screening [110]. The protein structure must be prepared appropriately as for rigid docking
[50]. The question of proper protonation of protein residues should already have