Small Molecule Docking from Theoretical Structural Models - Computational Modeling of Biological Systems - page 77

Structural coverage human proteins (RefSeq)

Unknown structures

Known experimental
structure
Model at 40% identity

Model at 30% identity

Fig. 1 Structural coverage of human proteins according to RefSeq without including splicing
variants

review the state-of-the-art of docking procedures, making special emphasis on the
potential use of ensembles of structural protein models derived from homology
modeling in high-throughput docking experiments.

2 Docking Algorithms

There is a plethora of docking algorithms and strategies that have been implemented
in a large variety of computer programs, some local and used by a restricted
community, and others commercially available that have a wide user community. It
is out of our scope to review all of them here, and we just outline the basic formalism
behind the most popular ones. The reader is addressed to excellent reviews to gain
a more complete view on current algorithms [ 10 , 13 - 16 ].
In principle, all docking algorithms follow a stepwise procedure: (1) several
estimates of the ligand-protein complex (binding poses) are proposed, and (2) these
poses are then ranked using a scoring function and offered to the user, who typically
focuses his/her attention to the best scored ones. Given that scoring functions
are fitted against experimental binding data, scoring values have “free energy of
binding” units. Therefore, they can be used to differentiate between good and bad
drug candidates and even to have an estimate of the binding free energy of the drug.
The differences between the different docking programs rely on (1) the method
used to explore the drug-binding landscape, (2) the method used to introduce
flexibility, and (3) the nature and the parameterization of the scoring function. For
example, DOCK [ 17 ], one of the first widely used docking programs, performs a
geometrically based docking of the ligands based on isomorphic subgraph matching
algorithms [ 18 ], which is later refined by considering the chemical nature of the
ligand and the binding site. Different scoring functions—mostly in the AMBER [ 19 ]
force-field—are used during the different stages of the fitting and ranking process,
including complex physical functions calling to atomistic force-field calculations
coupled to Generalized Born or Poisson-Boltzmann calculations. The popular
AUTODOCK program [ 20 ] offers a variety of optimizers including Monte Carlo
simulated annealing and different genetic algorithms using smoothed potential