Docking Predictions of Protein-Protein Interactions and Their Assessment: The CAPRI Experiment - Identification of Ligand Binding Site and Protein-Protein Interaction Area - page 97

functions used in docking did not yield reliable binding energies. They had been
developed to identify the correct mode of assembly of two proteins known to interact,
not to determine whether or not they form a stable complex, and this was beyond
their capacity. A parallel study showed a very poor correlation between experimental
binding energies and values calculated with several scoring procedures (Kastritis
and Bonvin 2010 ), with the same conclusion that the latter could not predict af fi nity.

5.6.3 A Structure Af fi nity Benchmark

The binding energy of a complex, or more correctly its Gibbs free energy of disso-
ciation ΔG d derived from the equilibrium constant K d , is a convenient measure of
af fi nity. K d is known from biophysical measurements in solution for many protein-
protein complexes that have been studied by crystallography, and a number of
authors have attempted to derive ΔG d from these structures. The first were Horton
and Lewis ( 1992 ). They collected data on 16 protein-protein complexes of known
structure (mostly protease/inhibitor complexes at that time), and found that a model
based on just the size and chemical composition of the interface yielded ΔG calc val-
ues that were within 1 or 2 kcal.mol −1 of the measured ΔG exp . However, there was an
exception: their model predicted a very similar affinity for BPTI binding to trypsin
and trypsinogen, whereas the experimental values differed by 10 kcal.mol −1 . Horton
and Lewis knew the reason why, and their paper discusses it. Trypsinogen, an inac-
tive precursor of trypsin, has flexible surface loops that become ordered when BPTI
binds (Bode et al. 1978 ). As a result, its affinity for the inhibitor is orders of magni-
tude less than trypsin, where no such change occurs, even though the two complexes
with BPTI are nearly identical in structure.
Like trypsin, most of the proteases and inhibitors of the Horton-Lewis set bind as
rigid bodies, with no major conformation change to affect their thermodynamic
stability. Later studies of the affinity/structure relationship in protein-protein com-
plexes employed larger data sets and more elaborate models of ΔG calc . But as none
took into account the structure of the free proteins, they all ignored the role of con-
formation changes, and also the large effect that experimental conditions, especially
pH, can have on K d . Not surprisingly, the correlation between ΔG calc and ΔG exp was
poor in these studies. In addition, errors accumulated in the structure/affinity sets
that served to optimize or test the models, as each study re-used data collated by
previous ones. Many of the experimental values in the sets were incorrect, some
grossly so; for instance, trypsinogen/BPTI and trypsin/BPTI were given the same
ΔG exp , a 10 kcal.mol −1 error. There was an obvious need for a validated test set, and
in 2010, I teamed with three other groups to assemble a benchmark set of binary
complexes that would have (a) experimental structures for both the complex and its
components; (b) a reliable K d measured under well-defined conditions. The 176
complexes of the Weng docking benchmark satisfied condition (a). They were an
obvious starting point, and we undertook to scan the biochemical literature in search
of a K d for them.