Docking Predictions of Protein-Protein Interactions and Their Assessment: The CAPRI Experiment - Identification of Ligand Binding Site and Protein-Protein Interaction Area - page 95

predictors' ones. These models came from the same docking searches, but the
predictors' procedures scored them low, whereas some of the schemes developed by
scorer groups adequately identified them as correct (Lensink and Wodak 2010 ) .
A scoring function can be physical-chemical (force fields, solvation energies), or
empirical, combining terms from different origins with weights optimized on sets of
positive and negative examples. It may include non-structural information derived
from the comparison of homologous sequences, from point mutants or other genetic
or biochemical experiments. However, such information is often ambiguous, and
sometimes misleading. If external information is used to screen models during or
after the search, it should be treated as a flexible restraint rather than a rigid con-
straint. The HADDOCK procedure efficiently incorporates such information into a
search algorithm that can also handle data from other sources, NMR experiments for
instance (Dominguez et al. 2003 ; de Vries et al. 2007, 2010 ; Stratmann et al. 2011 ) .

5.5.3 Flexibility and the Docking Benchmark

Developing scoring functions is an active field of research in many fields of science,
but in docking, the main difficulty remains flexibility. The structures deposited in
the Protein Data Bank illustrate many kinds of conformation changes, the docking
benchmark of Weng and colleagues, also. The benchmark is a set of PDB entries
assembled to test docking procedures. It contained only 59 complexes in its first
version (Chen et al. 2003b ), but now has entries for 176 protein-protein complexes
and their unbound components; one-third display significant backbone movements
with root-mean-square amplitudes that range from 1.5 to 10 Å (Hwang et al. 2010 ,
and Table 5.1 ).
The complexes of the benchmark are implicated in all sorts of biological pro-
cesses. Antigen/antibody and enzyme/inhibitor complexes are no longer a majority.
Signal transduction and cellular trafficking (exemplified by Arf6 in Fig. 5.1 ) are
well represented, and the protein-protein complexes involved these processes offer
many examples of flexible recognition. Conformation changes mediate signal trans-
duction in many ways: they may change the affinity of a protein for a small ligand,
another protein or DNA, enhance or inhibit a catalytic activity, the GTPase activity
of a G-protein for instance, mask or reveal a group that governs the cellular localiza-
tion of the protein or its attachment to a membrane. Their variety is immense, com-
parable in principle to the variety of macromolecular interactions seen in nature,
which neither the docking benchmark nor the PDB itself, are close to cover.
Moreover, entire classes of interactions are missing: those that involve membrane
proteins and intrinsically disordered proteins (IDP), for instance. IDP are implicated
in many macromolecular interactions (Dunker et al. 2005, 2008 ; Tompa et al. 2009 ) ,
and they undergo disorder-to-order transitions when they interact with other compo-
nents. Simulating such transitions in the context of docking will remain a challenge
for many years.