526
S. von Grafenstein et al.
averaging over the binding energies of the individual states would be physically rigorous as the free energy of binding is expected to be a function of the whole thermodynamic ensemble [119]. Still, classical docking algorithms and scoring functions
do not fulfill the needs for this approach. This theory relies on a thermodynamic
ensemble of docking poses whereas most docking algorithms are optimized to generate a single pose with optimal score [119]. Moreover, scoring functions are fitted
to experimentally-derived thermodynamic data as free energy of binding or related
binding constants (K i ) intrinsically representing ensemble properties. Therefore,
scoring functions inherently include ensemble properties and are trained to mirror
these effects in the score of a single pose. In contrast, this theory needs alternative
scoring functions which estimate the free energy of binding directly via averaging
over an ensemble. Such scoring functions might rather focus on an estimate of the
enthalpy of binding for each individual receptor conformation in contrast to a global
free energy of binding. In addition, these scoring functions would need a training
based on ensemble data rather than single X-ray structures.
15.2.2.4 Validation of Molecular Docking Results
Having decided for a merging strategy of the multiple hit list, the final scoring value
used for ranking can be evaluated with the same criteria and measures as a classical
docking run. Performance metrics of molecular docking rely on experimental data
and can be categorized in: (1) evaluation of the binding pose for ligands with known
native pose, (2) ranking of known ligands by the experimental activity, (3) discrimination of active ligands and non-active ligands or decoy molecules.
First, identification of a native binding pose similar to the ligand’s structure in an
X-ray structures is in the simplest case a redocking of a ligand into the empty binding pocket. Similarity is classically measured in terms of root mean square deviation (RMSD) of (heavy) atom positions of the ligand. The recovery of a near-native
pose is a confirmation of a successful sampling of the pose. Moreover, successful
pose identification alone is not sufficient: a scoring function must guide most docking algorithms and it should be capable to identify the best pose as the top solution
or within the top solutions. Therefore, some studies give the RMSD of the highest
ranked as well as the first occurrence of near-native pose within the solutions. The
interdependence of posing and scoring is a problem. The posing algorithm SKATE
tries to provide exhaustively enumerated poses which can be used for evaluation
and optimization of scoring without interference on pose sampling and show an attempt to decouple posing and scoring [120].
In rigid docking, values below 2 Å are classically considered near-native structures and a successful docking pose. The tolerance is related to the resolution of the
X-ray experiments. Considering docking algorithms covering flexibility, the generation of a ligand pose from an alternative crystal structure is a more appropriate
measure for posing. Especially when dealing with X-ray ensembles the exclusion of
the specific ligand complex within the ensemble should be considered. It reduces the
challenge to a classical redocking challenge if the right combination of ligand and
S. von Grafenstein et al.
averaging over the binding energies of the individual states would be physically rigorous as the free energy of binding is expected to be a function of the whole thermodynamic ensemble [119]. Still, classical docking algorithms and scoring functions
do not fulfill the needs for this approach. This theory relies on a thermodynamic
ensemble of docking poses whereas most docking algorithms are optimized to generate a single pose with optimal score [119]. Moreover, scoring functions are fitted
to experimentally-derived thermodynamic data as free energy of binding or related
binding constants (K i ) intrinsically representing ensemble properties. Therefore,
scoring functions inherently include ensemble properties and are trained to mirror
these effects in the score of a single pose. In contrast, this theory needs alternative
scoring functions which estimate the free energy of binding directly via averaging
over an ensemble. Such scoring functions might rather focus on an estimate of the
enthalpy of binding for each individual receptor conformation in contrast to a global
free energy of binding. In addition, these scoring functions would need a training
based on ensemble data rather than single X-ray structures.
15.2.2.4 Validation of Molecular Docking Results
Having decided for a merging strategy of the multiple hit list, the final scoring value
used for ranking can be evaluated with the same criteria and measures as a classical
docking run. Performance metrics of molecular docking rely on experimental data
and can be categorized in: (1) evaluation of the binding pose for ligands with known
native pose, (2) ranking of known ligands by the experimental activity, (3) discrimination of active ligands and non-active ligands or decoy molecules.
First, identification of a native binding pose similar to the ligand’s structure in an
X-ray structures is in the simplest case a redocking of a ligand into the empty binding pocket. Similarity is classically measured in terms of root mean square deviation (RMSD) of (heavy) atom positions of the ligand. The recovery of a near-native
pose is a confirmation of a successful sampling of the pose. Moreover, successful
pose identification alone is not sufficient: a scoring function must guide most docking algorithms and it should be capable to identify the best pose as the top solution
or within the top solutions. Therefore, some studies give the RMSD of the highest
ranked as well as the first occurrence of near-native pose within the solutions. The
interdependence of posing and scoring is a problem. The posing algorithm SKATE
tries to provide exhaustively enumerated poses which can be used for evaluation
and optimization of scoring without interference on pose sampling and show an attempt to decouple posing and scoring [120].
In rigid docking, values below 2 Å are classically considered near-native structures and a successful docking pose. The tolerance is related to the resolution of the
X-ray experiments. Considering docking algorithms covering flexibility, the generation of a ligand pose from an alternative crystal structure is a more appropriate
measure for posing. Especially when dealing with X-ray ensembles the exclusion of
the specific ligand complex within the ensemble should be considered. It reduces the
challenge to a classical redocking challenge if the right combination of ligand and