Docking Predictions of Protein-Protein Interactions and Their Assessment: The CAPRI Experiment - Identification of Ligand Binding Site and Protein-Protein Interaction Area - page 92

the third, a bacterial protein kinase co-crystallized with its substrate, the small protein
HPr. Fifteen predictor groups submitted a total of 193 models, and the CAPRI asses-
sors led by S. Wodak, compared them to the X-ray structures. The assessors
found that the submissions contained good models of the two antigen/antibody
complexes, but not of the HPr/kinase complex (Méndez et al. 2003 ) . They neverthe-
less decided that a few of the HPr/kinase models were “acceptable”: their geometry
was poor, but most of the residues in the contact regions were correctly predicted,
which could in principle help designing experiments. Predicting the residues in con-
tact was not a big feat in that case, since the location of the kinase active site and the
serine residue phosphorylated on HPr were known from the literature. Moreover,
the poor geometry of the models had an obvious origin: in the X-ray structure of the
complex, the rotation of a a-helix in the kinase modified the shape of the substrate
binding site and the way it bound HPr (Fieulaine et al. 2002 ) . Thus, rigid-body
docking was able to locate the correct epitopes on the two viral antigens, and place
them correctly at the antibody combining sites, but it failed on HPr/kinase due to a
conformation change, albeit one of limited amplitude.

5.4.2 Success and Failure in Blind Predictions

This pattern was repeatedly observed in later prediction rounds (Méndez et al. 2005 ;
Lensink et al. 2007 ; Lensink and Wodak 2010 ; Janin 2005, 2010 ). In the 10 years
that followed the Charleston meeting, CAPRI has had 22 rounds, with a total of 43
targets and an average of 45 predictor groups, each submitting ten models of each
target. In addition to protein-protein complexes, the targets have been a protein-
RNA complex and four oligomeric proteins. For each target, the predictors were
given the coordinates of the unbound components, or of an homolog protein that
could be used for model building, and they had 3-6 weeks to make their prediction
and submit their models. A majority (70%) of the targets obtained good quality
models. Almost all those that displayed only small backbone movements did, and in
most cases, the good models came from several groups using different docking
procedures. Figure 5.1 shows an example. Target T37, drawn here after the X-ray
structure of Isabet et al. ( 2009 ), is a complex between the G-protein Arf6, a member
of the Ras family of small GTPases, and the LZ2 segment of JIP4 (JNK-interacting
protein 4), an effector of Arf6. LZ2 was known to form a leucine zipper, and it had
to be model built from its amino acid sequence before docking on Arf6. A standard
leucine zipper yields a rather accurate model of its structure in the complex, while
Arf6 undergoes little change in the interaction. Correspondingly, the submissions
contained a number of good quality models of LZ2/Arf6, submitted by nine different
groups (Lensink and Wodak 2010 ) .
On the other hand, CAPRI predictions have yielded at best “acceptable” models
of the targets in which the backbone changes were large, or the homology models of
poor quality. Prediction yielded no valid model at all in six cases. In two, the failure
could be traced to misleading biochemical information rather than the structure
itself, in the other four, to large conformation changes. Moreover, some of the targets