Prediction of Protein-Protein Binding Interfaces - Identification of Ligand Binding Site and Protein-Protein Interaction Area - page 109

enabling ZDOCK to estimate the strength of interactions through CHARMM
forcefields (MacKerell et al. 1998 ; Brooks et al. 2009 ) . The conformational search
procedure implemented in ZDOCK is often applied in docking protocols which
typically augment it with a scoring scheme. It has been successfully tested in sev-
eral editions of the CAPRI challenge (Wiehe et al. 2005, 2007 ; Chen et al. 2003a,
b, c ; Mintseris et al. 2005 ) .
In our study ZDOCK was used with its default settings. Proteins were docked on
a dense 6 Å grid, producing 54,000 unique alignments. In each complex the longer
chain (receptor) was immobilized while the other chain was rotated around the recep-
tor molecule. The resulting alignments were subsequently scored using ZDOCK,
reflecting surface complementarity, electrostatic energy and statistical data on atomic
contact potentials in the interface zone. The highest ranked structures were clustered
by applying the Root Mean Square Deviation (RMSD) distance criterion for all
heavy atom positions. All structures with RMSD lengths of less than 10.0 Å were
clustered together, while outliers were removed from pool. Such ranking-based clus-
tering enables relatively fast extraction of the most representative alignments.

6.3 Results

In order to validate the proposed complexation site identification method, a set of
homodimers has been prepared by scanning the PDB database for occurrences of
the “HOMODIMER” keyword. Structures which did not consist of exactly two
chains, or which occurred in complexes with DNA, were exempted from analysis.
In addition, the Needleman-Wunsch (Needleman and Wunsch 1970 ) alignment
algorithm was applied to verify sequence similarity (identity): chains differing by
more than 20 amino acids (through substitutions or insertions/deletions) were dis-
carded, resulting in a set of 208 acceptable homodimers. This selection was based
upon PDB as it existed in March 2010.
Results produced by the above described programs were evaluated in terms of
their accuracy, which can be expressed by four distinct ratios: TP (true positive), FP
(false positive), TN (true negative) and FN (false negative). The study was based on
the F-measure and MCC criteria presented in the previous chapter.
The “fuzzy oil drop” model identifies residues involved in complexation by
searching for local maxima (i.e. deficiencies) and minima (i.e. excesses) of the Δ H

profile, as compared to the idealized 3D Gauss distribution of hydrophobicity. This
identification is dependent on a predetermined set of thresholds (cutoff values),
which, in our study, was pegged at 80% of the peak value (for minima and maxima
of the profile). Thus, residues whose Δ H values were in excess of 80% of their

respective peaks (or troughs), were suspected of involvement in complexation. The
fraction of these residues which are actually involved in complexation constitutes
the true positive (TP) ratio. Similarly, the fraction of residues which the algorithm
suspects of involvement in complexation but which do not actually participate in
forming complexes is defined as the false positive (FP) ratio. This operation is