510
T. Borowski and M. Szaleniec
proteins and are more suited for studies of extreme pK a shifts of individual residues
[34]. For the practical purposes, however, fast methods that utilize force field and
Poisson–Boltzmann or generalized Born approach, or even empirical methods, are
commonly used. It is not our goal to discuss here details of different approaches, but
rather we would like to point out some potential pitfalls and dangers associated with
use of such tools. For a comprehensive comparison of different methods, see reviews
by Awoonor-Williams et al. and Nielsen et al. [35, 36].
As a measure of accuracy of pK a prediction methods, we can use the isoelectric
point (pI) , which can be both readily measured, e.g. by means of DLS (Dynamic Light
Scattering) [37] or isoelectrofocusing [38], as well as calculated by pK a -predicting
programs. Such comparison seems to be a valid assessment tool as the computational methods are parameterized based on the known data set of pK a shifts of
selected residues (Asp or Glu) and very seldom pI values of proteins are used in
parametrization. In our small test, we have used two computational methods to calculate pK a and pI values: a very popular empirical method of Propka3.1 [39] and a
CHARMM-based GB/IMC (Generalized Born/iterative mobile clustering) protocol
implemented in Discovery Studio [40]. These have been applied to four proteins
for which structures are either known from X-ray diffraction studies (bovine αchymotrypsin (PDB: 5CHA), human serum albumin HSA (PDB: 1AO6), fibrynogen
(PDB: 1M1J)), a homology model of 3-ketosteroid dehydrogenase (KTSD) from
Sterolibacteria denitrificans, also known as AcmB, as well as two protein models
developed using molecular modelling techniques in conjunction with global secondary structural information derived from Raman spectroscopy data: α-casein and
β-casein [41, 42]. As can be seen from the data gathered in Table 1, for most of
the tested proteins the prediction methods perform relatively well (error below 1
pH unit). In some cases, however, the errors are way beyond the theoretical RMSD
for individual pK a (0.79–1.3 for PROPKA3, 0.5 for GB/IMC) of the training set
used to develop these two methods. Large discrepancies are observed for fibrynogen and AcmB with the error for the latter being bigger than 3–4 units. The largest
error observed for the AcmB homology model might indicate that some parts of it
are less-than-optimally folded. On the other hand, homology models are frequently
used to study new enzymes and this case illustrates a potential problem one may
encounter. pH reaction optimum of AcmB is around 6.5, and for this range of pH,
the computational methods predict the model to be positively charged (with the predicted pI ca. 8–9). Meanwhile, the experimental data indicate that in reality this
enzyme has pI in the 4.7–5.2 range, and hence, the protein is negatively charged at
the optimum pH. As a result, one might obtain qualitatively incorrect charge of the
homology model for the optimal pH, which in turn might influence the quality of
both MD simulation results and, most importantly, possibly also energy profile of
the catalytic reaction. The latter can be expected especially when the active site is
close to the protein surface and the reaction involves large charge separation events.
Even though the actual source of error in this particular case is still not known (it
might be due to failure of the homology modelling protocol, pK a prediction or both),
we advise a healthy dose of scepticism towards results obtained with pK a predictors
T. Borowski and M. Szaleniec
proteins and are more suited for studies of extreme pK a shifts of individual residues
[34]. For the practical purposes, however, fast methods that utilize force field and
Poisson–Boltzmann or generalized Born approach, or even empirical methods, are
commonly used. It is not our goal to discuss here details of different approaches, but
rather we would like to point out some potential pitfalls and dangers associated with
use of such tools. For a comprehensive comparison of different methods, see reviews
by Awoonor-Williams et al. and Nielsen et al. [35, 36].
As a measure of accuracy of pK a prediction methods, we can use the isoelectric
point (pI) , which can be both readily measured, e.g. by means of DLS (Dynamic Light
Scattering) [37] or isoelectrofocusing [38], as well as calculated by pK a -predicting
programs. Such comparison seems to be a valid assessment tool as the computational methods are parameterized based on the known data set of pK a shifts of
selected residues (Asp or Glu) and very seldom pI values of proteins are used in
parametrization. In our small test, we have used two computational methods to calculate pK a and pI values: a very popular empirical method of Propka3.1 [39] and a
CHARMM-based GB/IMC (Generalized Born/iterative mobile clustering) protocol
implemented in Discovery Studio [40]. These have been applied to four proteins
for which structures are either known from X-ray diffraction studies (bovine αchymotrypsin (PDB: 5CHA), human serum albumin HSA (PDB: 1AO6), fibrynogen
(PDB: 1M1J)), a homology model of 3-ketosteroid dehydrogenase (KTSD) from
Sterolibacteria denitrificans, also known as AcmB, as well as two protein models
developed using molecular modelling techniques in conjunction with global secondary structural information derived from Raman spectroscopy data: α-casein and
β-casein [41, 42]. As can be seen from the data gathered in Table 1, for most of
the tested proteins the prediction methods perform relatively well (error below 1
pH unit). In some cases, however, the errors are way beyond the theoretical RMSD
for individual pK a (0.79–1.3 for PROPKA3, 0.5 for GB/IMC) of the training set
used to develop these two methods. Large discrepancies are observed for fibrynogen and AcmB with the error for the latter being bigger than 3–4 units. The largest
error observed for the AcmB homology model might indicate that some parts of it
are less-than-optimally folded. On the other hand, homology models are frequently
used to study new enzymes and this case illustrates a potential problem one may
encounter. pH reaction optimum of AcmB is around 6.5, and for this range of pH,
the computational methods predict the model to be positively charged (with the predicted pI ca. 8–9). Meanwhile, the experimental data indicate that in reality this
enzyme has pI in the 4.7–5.2 range, and hence, the protein is negatively charged at
the optimum pH. As a result, one might obtain qualitatively incorrect charge of the
homology model for the optimal pH, which in turn might influence the quality of
both MD simulation results and, most importantly, possibly also energy profile of
the catalytic reaction. The latter can be expected especially when the active site is
close to the protein surface and the reaction involves large charge separation events.
Even though the actual source of error in this particular case is still not known (it
might be due to failure of the homology modelling protocol, pK a prediction or both),
we advise a healthy dose of scepticism towards results obtained with pK a predictors