3.4.5 Comparison
to some PBSA or GBSA
Approaches Applied
to Other Systems
Table 2 summarizes the performance of some earlier PBSA and
GBSA approaches. Several achieve high correlations with experiments. Only five outperform the Null model and only one achieves
(for a much smaller dataset) the small errors reported here.
4 Concluding Notes
We conclude with a series of practical observations, or “lessons
learned” that we have derived from our work and that should
help readers apply the methods presented above to related systems:
1. High-throughput CPD of the PDZ protein for binding can be
handled in the same way as the peptide design illustrated in
Subheading 2. A slightly simpler approach was reported earlier
[37] that did not employ (yet-undiscovered) adaptive MC.
2. We recently reported redesign of the entire PDZ protein, the
first successful whole-protein design with a nonempirical,
physics-based energy function [60].
3. A key to the accurate PB/LIE predictions described in Subheading 3 was the use of the Eq. (4) empirical ansatz. A pure
PBSA or GBSA free energy is unlikely to succeed without
empirical weights.
4. The solute dielectric constant ϵ S for the PB component (Eq. 4
in Subheading 3) is an empirical parameter. Choosing a value of
8 is physically reasonable [61]. Choosing a different value (such
as 4) led to a nearly proportional change in the PB weight β
(Eq. 4) and thus similar results.
5. Another key to PB/LIE success was good conformational
sampling, thanks to moderately long MD runs and suppression
of slow (evidently unimportant) N-terminal fluctuations of the
peptide.
6. Our PB/LIE model (Subheading 3) made no attempt to
model contributions from conformational entropy changes;
we assume normal mode or quasi-harmonic models would
not be predictive for our systems and MD trajectory lengths.
7. The PB/LIE model could not handle very weak binders, which
were left out of the fit. Despite using MD, it also could not
handle the conformational changes between Sdc1-like peptides
and Cask-like peptides, which have distinctly different backbone arrangements. Therefore, these two data sets each had
their own reference complex [23].
8. The tradeoff for the excellent PB/LIE accuracy (Subheading
3.3) is the need to fit α, β, γ. Transferability to other PDZ
domains should be good but was not yet tested.
250
Nicolas Panel et al.
to some PBSA or GBSA
Approaches Applied
to Other Systems
Table 2 summarizes the performance of some earlier PBSA and
GBSA approaches. Several achieve high correlations with experiments. Only five outperform the Null model and only one achieves
(for a much smaller dataset) the small errors reported here.
4 Concluding Notes
We conclude with a series of practical observations, or “lessons
learned” that we have derived from our work and that should
help readers apply the methods presented above to related systems:
1. High-throughput CPD of the PDZ protein for binding can be
handled in the same way as the peptide design illustrated in
Subheading 2. A slightly simpler approach was reported earlier
[37] that did not employ (yet-undiscovered) adaptive MC.
2. We recently reported redesign of the entire PDZ protein, the
first successful whole-protein design with a nonempirical,
physics-based energy function [60].
3. A key to the accurate PB/LIE predictions described in Subheading 3 was the use of the Eq. (4) empirical ansatz. A pure
PBSA or GBSA free energy is unlikely to succeed without
empirical weights.
4. The solute dielectric constant ϵ S for the PB component (Eq. 4
in Subheading 3) is an empirical parameter. Choosing a value of
8 is physically reasonable [61]. Choosing a different value (such
as 4) led to a nearly proportional change in the PB weight β
(Eq. 4) and thus similar results.
5. Another key to PB/LIE success was good conformational
sampling, thanks to moderately long MD runs and suppression
of slow (evidently unimportant) N-terminal fluctuations of the
peptide.
6. Our PB/LIE model (Subheading 3) made no attempt to
model contributions from conformational entropy changes;
we assume normal mode or quasi-harmonic models would
not be predictive for our systems and MD trajectory lengths.
7. The PB/LIE model could not handle very weak binders, which
were left out of the fit. Despite using MD, it also could not
handle the conformational changes between Sdc1-like peptides
and Cask-like peptides, which have distinctly different backbone arrangements. Therefore, these two data sets each had
their own reference complex [23].
8. The tradeoff for the excellent PB/LIE accuracy (Subheading
3.3) is the need to fit α, β, γ. Transferability to other PDZ
domains should be good but was not yet tested.
250
Nicolas Panel et al.