Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5319_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
29.08.2026
Размер:
92 Мб
Скачать
. • D-QSAR: Correlation of Molecular Fields with Biological Properties


. Fig. 18.4 A grid is generated for the calculation of molecular elds
that broadly encompasses amolecule. In this image, the grid points are color-coded with increasing distance from the ligand (red < yel­low < green < blue < gray). The contributions from the chosen elds are calculated at all points of the lattice, which have agrid spacing of 1–2 Å. The eld contributions at each point in the grid (S1, S2,…Sn, E1, E2,…En) are entered into aspreadsheet. The analysis is carried out for all molecules in the dataset. The binding afnities are incor­porated into the spreadsheet as, for instance, −log (Ki). The eld contributions are weighted with appropriate coefcients (a, b,… z)
tion on amolecular scaffold is actually very important for binding, but remains insignicant in the analysis. This has to do with the fact that aQSAR analysis only makes relative comparisons within adataset. In other words, if aproperty is the same for all ligands in the dataset, then that property will not be able to explain any differences in the biological data.
and using aspecial statistical method, the partial least squares (PLS) analysis, they are related to the afnity. Amodel is obtained in the form of an equation that indicates at which grid points and with what weights the different eld contributions explain the biological activity. (7 https://sn.pub/jndLSI)
These examples are still easy to handle. The question can be asked whether atedious correlation method with a“detour” via molecular elds is really necessary. In practice, the situation is more complicated, especially when considering molecules with different scaffolds. The substituents do not fall exactly on top of one another in the molecular superposition. Their contribution must
18
Chapter  • Quantitative Structure–Activity Relationships
. Fig. 18.5 The Lennard-Jones potential (green) is amodel for de-
scribing the intermolecular interactions of two atoms without con­sidering their charge. Negative potential values correspond to mutual attraction, positive values correspond to arepulsion of the particles. If areciprocal distance becomes innite, the potential will approach zero. Upon approach, it goes through ashallow minimum due to al­ternating polarization. At even shorter distance, it very steeply rises towards positive innity because of atom–atom repulsions. The Cou­lomb potential (blue) considers only electrostatic interactions that for­mally reside as point charges on the atomic nuclei. It also approaches innity when the distance disappears for like-charged particles. For oppositely charged atoms, negatively innite values result. The hyper­bolic form of the Coulomb potential is considerably less steep, so that the particles can still “feel” one another at larger distances. Boundary values are set for potentials in aCoMFA analysis. AGaussian func­tion, which takes the course of abell-shaped curve (red, here only the right half of the “bell” is shown) describes the distance dependence of the interaction potential between the particles in the context of the CoMSIA model. As the distance disappears between the particles, the curve reaches its maximum value, which remains nite
be described as aeld in space, and only as such can it be evaluated. In any case, these examples underline the importance of careful planning of the analysis. The structures of the dataset must be chosen in such away that the variation of the substituents and their properties is maximized.
18.12 Results of aComparative Molecular
Field Analysis and Their Graphical Interpretation
When the full complexity of the eld contributions is considered in terms of amultidimensional matrix, asim­ple regression analysis cannot be applied to extract the interdependence of the variables, e.g., the binding afnity. Partial least squares (PLS) analysis is astatistical method that extracts relevant and explanatory factors, called PLS vectors, from large amounts of data. In CoMFA analysis, these vectors describe the range of elds that best cor­relate with the experimentally determined afnity. The result is an equation analogous to the results of classical QSAR methods. This equation shows to what extent par-
ticular grid points in each of the individual elds con­tribute to the relative differences in the binding afnities. Depending on how many grid points are to be evaluated in the analysis, the statistical signicance of the derived results must be strictly monitored. This signicance is checked by aspecial test: the cross validation.
This is done by randomly removing one or more compounds from the dataset. Amodel is constructed with the remaining compounds of the dataset and the afnities of the removed compounds are predicted with this model. The removal of compounds is repeated sev­eral times, in the simplest case until all compounds have been removed consecutively one after the other (leave- one-out validation). The quality of the overall prediction obtained is ameasure of the reliability and validity of the model. The obtained result is expressed by the q2 value, which can be calculated from the square of the deviation from the predicted values. It takes values from −∞ to +1. Avalue of +1 indicates aperfect model. All predictions exactly match the measured binding afnities. There is no deviation. Avalue of q2 = 0 indicates that the predictions of the model are no better than no model at all; they are as good as the average of all afnities. If q2 takes negative values, the model is worse than the average, which means it is worse than no model. Therefore, amodel can only be trusted if q2 is above 0.4–0.5.
Afurther step is required to check the predictive power of atrained model (one speaks of a “trained” model and, therefore, one refers to the dataset as a“training dataset”). This validation step requires atest dataset of molecules that are similar to the molecules in the training dataset, but were not used for the initial training. Binding afnities are predicted for these molecules. Only if acor­relation coefcient, calculated analogously to q2 with the training dataset, is of similar magnitude will the derived model have sufcient predictive power. This procedure is, therefore, atest of the model on independent data that were not used to derive the model.
The derived model can then be used to estimate the afnity of new compounds that have not yet been syn­thesized. The conformations of these compounds are generated and superimposed on the other structures. They must fall within the grid dimensions dened in the training set. Then their eld contributions are calculated. Using the correlation derived by CoMFA for the training set, it is possible to calculate which grid points are predic­tive of the binding afnity of new compounds.
CoMFA techniques establish acorrelation between activity data and molecular properties. From the relative comparison within atraining set, amodel can be derived that encompasses the properties of new molecules. Rel­evant predictions will only be obtained if the structural variations in the new molecule remain within the scope of the model. In other words, the model cannot make predictions about the inuence of substituents that occur in regions where there were no structural variations in the
. • Scope, Limitations, and Possible Expansions of the CoMFA Analysis


training set. CoMFA models interpolate between the eld contributions of molecules. Extrapolation to regions not covered by the dataset is not possible.
The results of aCoMFA analysis can be evaluated graphically, which is probably the most powerful aspect of the method for the medicinal chemist. From the model and the derived equation, it is known at which grid points eld contributions are obtained that provide asignicant explanation for the binding afnity. These contributions can be outlined for the different elds according to their importance. They indicate volume regions around the molecules in which changes in the eld contributions run parallel or opposite to the afnity changes in the data­set. These contour maps are an important support in the design of new compounds (Sect.18.14). They indicate where the properties of alead structure need to be varied to increase afnity.
18.13 Scope, Limitations, and Possible
Expansions of the CoMFA Analysis
Typically, only steric and electrostatic eld contributions are evaluated in CoMFA analyses. Ahydrophobic eld can quantify the size of the hydrophobic surfaces and, therefore, partially accounts for the entropic contribution to afnity. Since CoMFA evaluations yield relevant mod­els without the explicit use of hydrophobic elds, these eld contributions must be at least partially included in the Lennard-Jones and Coulomb elds. The lipophilic­ity of amolecule increases when an uncharged, sterically demanding group is enlarged, e.g., from methyl to bu­tyl. Here the changes in the steric eld contributions can correctly reect the lipophilic surface. Acorrelation with electrostatic properties is also possible. Hydrophobic moi­eties usually carry only small partial charges. Positively or negatively charged groups represent hydrophilic regions. In this way, the lipophilic and hydrophilic regions of the surface can be quantied by differences in charge.
The deviation that cannot be explained by aCoMFA model also includes, apart from experimental errors, all inadequately described binding contributions. These in­clude structural adaptations of the protein that are not identical for all compounds in the dataset. Entropic con­tributions resulting from (i)conformational xation of the drug molecules in the binding pocket, (ii)residual mobility of the ligands in the binding pocket, (iii)con- formational changes on the side of the protein, (iv)or signicant differences in the solvate structure remain­ing on the protein are all not considered in the elds. It should be noted, however, that such deviations will only be signicant if they differ from one ligand to another in the dataset.
In addition to these shortcomings, the elds them­selves cause some problems. Because of their mathe­matical functional form, they reach very large or very
small values near the surface or inside the molecules (. Fig. 18.5). Because the Lennard-Jones potential grows faster than the Coulomb potential as it ap­proaches the atoms, they both reach an arbitrarily set cutoff (Sect.18.10) at different distances from the mole­cules. The extremely steep Lennard-Jones potential can change its functional values from practically zero to the cut-off value within adistance of 2 Å, the commonly used grid spacing! These discontinuities and accordingly the lack of any variation within the cut-off regions near the ligand surfaces cause considerable problems in the evaluation. Moreover, they often produce “disrupted” and, therefore, difcult to interpret contour maps of the different elds.
The shortcomings of these elds have stimulated the search for alternative solutions. One method is to deter­mine the similarity of molecules by their steric and physi­cochemical properties in space. These are then correlated with binding afnities. This is done in the CoMSIA method (comparative molecular similarity indices anal- ysis). The molecules are superimposed similarly to the CoMFA method. How similar they are to one another is then measured in relation to aprobe, for example, acar­bon atom. For each molecule, the similarity to this probe is sampled at the intersections of asurrounding grid. The similarity measure between the probe and the molecule is dened as adistance-dependent function. AGaussian function is chosen for this purpose (. Fig.18.5). Unlike the hyperbolic course of the potentials described above, the Gaussian bell curve does not tend to innity for de­creasing distance values. Therefore, no cutoff values need to be set. At any grid point, asimilarity measure can be determined for alarge number of properties. Aprereq­uisite for the CoMSIA method is the description of the properties in terms of atomic values (e.g., partial charges, atomic volumes, lipophilicity, H-bonding properties). The same distance dependence is used for all proper­ties. Property-specic similarity elds are obtained and correlated with the binding afnities. The interpretation of the eld contributions is analogous to the CoMFA method. The main advantage of this method is the ease of interpretation of the resulting contour maps. If apartic­ular property in aregion of the superimposed molecules correlates signicantly with binding afnity, that region will be highlighted and this information can be easily translated into the design of anew molecule. In contrast, the CoMFA method contours only regions outside the molecules where aproperty reveals changes in eld con­tributions that positively or negatively affect afnity. However, by setting cut-off values, entire regions of these eld contributions are hidden, especially near ligand sur­faces (. Fig.18.5). Instead, the CoMSIA approach also contours at the atomic positions of the molecules that are responsible for the trends in the afnity changes. This gives medicinal chemists amuch more intuitive picture of where to modify molecules in an optimization process.
Chapter  • Quantitative Structure–Activity Relationships
18
3D-QSAR analyses were originally designed to es­tablish structure–activity relationships in cases where the structure of the target protein was not available as areference. Today, as more and more crystal structures of target proteins become available, the technique is in­creasingly used for cases where the protein reference is in deed known. The protein structure then helps to generate areasonable and relevant superposition of the training compounds to be compared in their biologically active conformations. It seems paradoxical to use the informa­tion about the protein environment only to superimpose the molecules and then to disregard this valuable data in the comparative eld analysis. Therefore, methods have been developed that take this information into account. Rebecca Wade’s group at the EMBL and HITS in Heidel­berg, Germany, has developed the COMBINE method. It uses aset of modeled protein–ligand complexes to cal­culate adata table. It contains the interaction energies between individual ligand atoms in the test molecules of the dataset and the amino acid residues and water molecules in the surrounding protein. The interpretation of this huge data table is achieved using atechnique sim­ilar to CoMFA methods. The graphical interpretation of the correlation model obtained by COMBINE indicates which regions of the protein are critical in explaining the afnity differences in the ligand dataset. These are very valuable details, but they are of limited help for the direct design of improved molecules with higher afnity.
The variant AFMoC (adaptation of fields for mo- lecular comparison), developed by Holger Gohlke in Marburg, allows the integration of information about the protein environment into the eld-based model. The advantages of the intuitive evaluation of the eld contributions with regard to the structural optimization of the ligands are not lost. To this end, the empirical scoring function DrugScore (Sect.17.10) is rst used to map onto the intersections of aCoMFA-like grid the values that aprotein environment would sense as an interaction at each of the grid points if sampled with aparticular atomic probe according to the functional form in DrugScore. The lattice is effectively “prepolar­ized” by the protein environment. The ligands of the training set are then placed on this lattice (using either adocking program or one of the superposition methods described). Whenever aligand places an atom type on aregion of the lattice where the protein considers that atom type to be advantageous, the eld contribution is increased. Otherwise, the interaction contribution on the lattice is reduced. In this way, adata table analogous to the CoMFA method is created for the entire training dataset. This table is then evaluated to generate aQSAR equation. The individual contributions can be graphi­cally visualized on the grid. They illustrate where certain atom types cause an increase or decrease in afnity, but now taking into account simultaneous information from the surrounding protein.
Similar eld analyses can also be used to correlate and predict selectivity differences between ligands. Many enzymes exist as isoforms. They therefore have similari­ties in their binding pockets. As aconsequence, ligands show graduated afnities or “selectivity proles” for
-
these isoforms. If aligand is to be optimized to improve selectivity, the positions at which achange in aprop­erty leads to an improved prole must be known. A3D QSAR model is constructed for each isoenzyme. Either the difference of the afnity values can be calculated and used for the model as the values to be predicted, or al­ternatively two correlation models can be constructed and the eld contributions at each grid point are sub- tracted from one another. The models obtained from either approach can be interpreted graphically. Contour plots show where and how molecules should be mod­ied to improve their selectivity for one or the other isoenzyme.
18.14 A Glimpse Behind the Scenes:
Comparative Molecular Field Analysis of Carbonic Anhydrase Inhibitors
Today, comparative eld analyses are part of the stan­dard repertoire in drug discovery. As an example, the binding of inhibitors to carbonic anhydraseI (CAI) andII (CAII) will be examined. The biological func­tion of these enzymes is described in detail in Sect.25.7. The sequence identity of the two isoforms is 60%. The ligands in the training dataset are derived from the parent structures shown in . Fig.18.6. First, asuperposition model is created by docking the ligands into the pro­tein (. Fig.18.7). The funnel-shaped binding pocket of the enzyme is occupied by ligands in avariety of ways. A good correlation model is obtained with the three methods, CoMFA, CoMSIA, and AFMoC. The models also achieve convincing predictive power on atest dataset independent of the training set.
The contours for the acceptor properties with respect to the inhibition of carbonic anhydraseII are shown in
. Fig.18.8. Molecules in the dataset that exhibit an ac-
ceptor function in the areas marked in red have lower potency. On the other hand, an acceptor function in the blue area improves potency. Compound 18.2, which has both acceptor functions of an SO2 group oriented in the detrimental red area, is aweak CAII inhibitor. Moreover its NH group is in the blue region, which should be oc­cupied by an acceptor. Compound 18.3, which is about four orders of magnitude more potent, leaves the area that was occupied by the oxygen atoms in 18.2 empty, and orients its thiadiazole ring in the direction of the de­sirable acceptor function. It achieves considerably better inhibition of the target enzyme.
. • A Glimpse Behind the Scenes: Comparative Molecular Field Analysis of Carbonic Anhydrase Inhibitors

. Fig. 18.6 The scaffolds of inhibitors that were used in different eld analyses to establish afnity (pKi[CAII]) and selectivity models (pKi[-
CAII] − pKi[CAI] = ∆pKi[CAII − CAI]) to describe the inhibition of the carbonic anhydrase CAI and CAII. Different substituents were varied at the positions that are marked as R1 and R2

. Fig. 18.7 The superposition of inhibitors from the dataset in the
funnel-shaped binding pocket of carbonic anhydraseII; the zinc ion is shown as the blue-gray sphere, carbon atoms are light yellow, oxygen red, nitrogen blue, sulfur orange, and hydrogen light cyan
As with acceptor properties, contour maps can be generated for steric, electrostatic, hydrophobic, and hy­drogen-bond donor properties. Their evaluation helps to identify where certain properties improve or decrease binding afnity. Such correlation analyses help the syn­thetic chemist plan the optimization of lead structures.
Contour maps based on a CoMFA analysis for steric properties that cause adifference in selectivity between CAI and CAII are shown in . Fig.18.9. Placing an in­hibitor next to the green areas improves selectivity for
. Fig. 18.8 Contour map (CoMSIA) for the description of the bind-
ing contributions of H-bond acceptor properties. Inhibitors that oc­cupy the red-contoured areas with H-bond acceptor groups do not inhibit carbonic anhydraseII (CAII) well; however, the occupancy of the blue areas with acceptor groups leads to increasing values. Both oxygen atoms of the sulfonamide group of 18.2 occupy the red-con­toured area, which is unfavorable for acceptor properties. On the other hand, 18.3 leaves these areas unoccupied and places its basic nitrogen in the vicinity of the blue-contoured region, which is favorable for the occupancy by acceptor groups. This explains the markedly better inhi­bition of CAII by 18.3
CAI. On the other hand, occupancy next to the yellow areas improves the selectivity for CAII. Compound
18.4 binds unselectively with the same afnity to both isoforms, but 18.5 can clearly discriminate between the two. The model shown is derived purely from the cor­relation of ligand-binding data. The relative alignment of the molecules in the dataset is achieved by docking into
Chapter  • Quantitative Structure–Activity Relationships
18
. Fig. 18.9 Top left The selectivity can be improved with regard to
carbonic anhydraseII (CAII) inhibition by sterically lling the region next to the yellow-contoured area. Filling the green area with sterically demanding group causes an increase in selectivity with regard to CAI. Top left and top right Compound 18.4 occupies virtually no area that is particularly selectivity discriminating; the compound is not isoen-
the binding pocket of the protein. Therefore, the protein environment around this binding pocket should be ex­amined more closely to see whether the derived contours are reasonable. Comparing the amino acid replacements between the two isoforms, it is apparent that CAI has two large residues, Phe91 and Leu 131, which restrict the lower left portion of the binding pocket more than in CAII. The inhibitors have less space in CAI than in CAII. In fact, the comparative eld analysis generates ayellow contour in this region (near position91), the occupancy of which should be favorable for potent inhibition of CAII. CAII also provides alarge space for inhibitors near position 204, which is occupied by the less crowded Leu 204 in CAII instead of Tyr 204 in CAI. Ayellow contour is visible, indicating a favorable occupancy of this site. Inhibitor 18.5, which is much more potent at
zyme specic. On the other hand, 18.5 occupies ayellow-contoured area neighboring position 204 (bottom left), which causes aselectivity enhancement for CAII. Compound 18.5 inhibits CAII decidedly more potently than CAI. All contour maps shown are based on data analy­sis by CoMFA; the adjacent protein residues were added to the images from corresponding crystal structures
CAII, orients its pentauorophenyl group exactly in this region (. Fig.18.9, right). In the vicinity of position 131 (Leu 131/Phe 131), ayellow and agreen region appear directly adjacent to each other but spatially separated, the occupancy of which is favorable for either CAI or CAII inhibitors. Compound 18.4, which can hardly distinguish between the two isoforms, occupies the upper edge of both regions equally well. Moreover, it leaves virtually all regions unoccupied which, for steric reasons, should lead to abetter inhibition of either CAI or CAII. This explains why this compound shows no particular selectivity.
Finally, the binding of the well-discriminating com­pound 18.6 should be considered (. Fig.18.10). The evaluation of the acceptor properties of the ligands in the training dataset shows that the occupancy of the red contour regions with H-bond acceptor groups shifts the
. • Synopsis


. Fig. 18.10 Compound 18.6 inhibits carbonic anhydraseI (CAI;
left, green) signicantly less potently than CAII (right, yellow). The sulfone oxygen atom on the left hand side (circled in magenta) falls close to ared contoured area (occupancy with H-bond acceptor group favorable), the lling of which causes an increase in the selectivity for CAII binding. Interestingly, Gln 92 is found in this region in both isoforms. However, it is only in CAII that this group is available to
selectivity in favor of CAII. Filling the blue contours with this property results in an increase in potency with respect to CAI. Compound 18.6 places its oxygen atoms of the endocyclic SO2 group near the red CAII selective regions. Again, it is important to note that the model shown is de­rived purely from the correlation of ligand-binding data, and that in the following the information from the protein isoforms is used only to understand the obtained correla­tion model. The protein structures show aglutamine res­idue adjacent to position92 in both CAI and CAII. This amino acid can accept an H-bond from the inhibitor via the NH2 group of its carboxamide group. However, only CAII allows these structural conditions. Gln92 is adjacent to Asn69 and Glu58 in CAI. The carboxamide group of Glu92 forms acontinuous H-bonding network with these residues and with His94. Therefore, the NH group is no longer available for interactions with abound inhibitor. This is reected in the lower binding afnity of inhibitors that have an acceptor function at this position, such as
18.6. The situation is completely different for CAII. The adjacent functional groups of Glu69 and Arg58 form an internal salt bridge. Therefore, they are not available as
accept an H-bond from the inhibitor that will contribute to binding afnity (right, black dotted line). The comparable residue Gln92 in CAI is involved in anetwork of H-bonds to neighboring amino acids Asn69 and Glu58. Therefore, it is not available as abinding partner, and adecrease in the afnity for CAI is the consequence. The shown contour maps are based aCoMSIA analysis and superimposed onto crystal structures of CAI and CAII
H-bonding partners for Gln92. The carboxamide group of Gln92 involves His94 in an H-bond via its carbox­amide CO group, and its NH2 group is now available as an acceptor functionality to interact with abound ligand. This results in signicantly enhanced binding to CAII and is expressed as aselectivity advantage. It is important to remember that the applied CoMSIA analysis was trained only on ligand-binding data. Therefore, it did not know anything about the properties of the neighboring amino acids. Nevertheless, it produces acorrelation model that can be easily explained by the unexpected binding behav­ior of the adjacent amino acids in the two isoforms.
18.15 Synopsis
The concept of quantitative structure–activity re-
-
lationships is not new. It was rst described in the
nineteenth century qualitatively, and later more quan-
titatively by Hansch and Fujita. It is an attempt to
describe structure–activity relationships with mathe-
matical models.
Chapter  • Quantitative Structure–Activity Relationships
18
Across aseries of structurally closely related test com-
-
pounds, the equieffective dose that induces apartic­ular biological effect is related in alinear or squared dependence on the logarithm of the octanol/water partition coefcient and the Hammett constant, which describes the electronic properties of substitu­ents at agiven scaffold. Amathematical correlation model is computed by regression analysis.
3D QSAR methods have been developed to consider
-
and correlate the spatial structure of active sub­stances beyond molecular topology.
The mutually aligned test molecules are embedded
-
in aregularly spaced grid and their properties are ex­plored with an interaction probe. The probe is placed systematically at all grid points and amolecular inter­action eld is computed around the aligned molecules by using adistance-dependent property potential.
Usually, Lennard-Jones and Coulomb potentials are
-
evaluated, and the generated data table for all mole­cules of the training dataset is correlated by the par­tial least squares technique.
The derived CoMFA correlation model can be used
-
to predict the biological properties of novel ligands not included in the training dataset. Strict criteria to monitor the statistical signicance of the derived cor­relations must be met.
Other property elds beyond Lennard-Jones and
-
Coulomb potentials with mathematically different functional forms can be applied. With respect to the prediction of binding afnity, it has to be regarded that hydrophobic properties also implicitly reect an entropic contribution to binding that is particularly difcult to regard in property elds.
QSAR analysis only performs arelative comparison
-
of molecules with regard to the considered biological property. Any dependence on aparticular descrip­tor across acompound series can only be expected if the property related to this descriptor is varied in the series. QSAR methods only interpolate and never extrapolate beyond the scope of molecular properties reected by the training set.
Anumber of alternative 3D QSAR approaches have
-
been developed. They use different types of elds (e.g., for similarity analysis, CoMSIA) or try to incor­porate protein information into the analysis (COM­BINE and AFMoC). Comparative molecular eld analyses can be evaluated
-
graphically. Results are displayed as contours around the molecules and indicate where the change of apar­ticular property runs either parallel or opposite to the changes in the biological property in the dataset.
The graphical information can be directly translated
-
into the design of modied molecules and, thus, sup­port the medicinal chemist in optimizing agiven lead structure in asystematic fashion.

Bibliography and Further Reading

General Literature
C. A. Ramsden, Eds., Quantitative Drug Design, Vol. 4: Comprehen-
sive Medicinal Chemistry, C. Hansch, P. G. Sammes and J. B. Tay-
lor, Eds., Pergamon Press, Oxford (1990) H. Kubinyi, QSAR: Hansch Analysis and Related Approaches, VCH,
Weinheim (1993) H. van de Waterbeemd, Chemometric Methods in Molecular Design,
VCH, Weinheim (1995) H. van de Waterbeemd, Advanced Computer-Assisted Techniques in
Drug Discovery, VCH, Weinheim (1995) C. Hansch and A. Leo, Exploring QSAR. Fundamentals and Appli-
cations in Chemistry and Biology, 2 Volumes, American Chemical
Society, Washington (1995) H. Kubinyi, Ed., 3D-QSAR in Drug Design: Theory, Methods, and
Applications, ESCOM, Leiden (1993) H. Kubinyi, G. Folkers and Y.C. Martin, 3D QSAR in Drug Design,
Vol. 1–3, Kluwer/ESCOM, Dordrecht, Boston, London (1998)
Special Literature
S. H. Unger and C. Hansch, On Model Building in Structure-Activity
Relationships. A Reexamination of Adrenergic Blocking Activity
of β-Halo-β-arylalkylamines, J. Med. Chem., 16, 745–749 (1973) J. M. Blaney, C. Hansch, C. Silipo and A. Vittoria, Structure-Activity
Relationships of Dihydrofolate Reductase Inhibitors, Chem. Rev.,
84, 333–407 (1984) C. Hansch and T. E. Klein, Quantitative Structure-Activity Relation-
ships and Molecular Graphics in Evaluation of Enzyme-Ligand
Interactions, Methods Enzymol.,202, 512–543 (1991) H.-D. Höltje and L. B. Kier, Nature of Anionic or α-Site of Cholines-
terase, J. Pharm. Sci., 64, 418–420 (1975) R. D. Cramer and M. Milne, Abstracts of the ACS Meeting, April
1979, COMP 44 R. D. Cramer, D. E. Patterson, and J. D. Bunce, Comparative Molecu-
lar Field Analysis (CoMFA). 1. Effect of Shape on Binding of Ste-
roids to Carrier Proteins, J. Am. Chem. Soc., 110, 5959–5967 (1988) S. A. DePriest, D. Mayer, C. B. Naylor and G. R. Marshall, 3D-QSAR
of Angiotensin-Converting Enzyme and Thermolysin Inhibitors:
A Comparison of CoMFA Models Based on Deduced and Exper-
imentally Determined Active Site Geometries, J. Am. Chem. Soc.,
115, 5372–5384 (1993) P. J. Goodford, A Computational Procedure of Determining Energet-
ically Favorable Binding Sites on Biologically Important Macro-
molecules, J. Med. Chem., 28, 849–857 (1985) G. E. Kellogg and D. J. Abraham, Key, Lock and Locksmith: Comple-
mentary Hydropathic Map Predictions of Drug Structure from a
Known Receptor-Receptor Structure from Known Drugs, J. Mol.
Graphics, 10, 212–217 (1992) T. Hüfner-Wulsdorf and G. Klebe, Protein-Ligand Complex Solva-
tion Thermodynamics: Development, Parameterization and Test-
ing of GIST-based Solvent Functionals, J. Chem. Inf. Model., 60,
1409–1423 (2020) G. Klebe, U. Abraham and T. Mietzner, Molecular Similarity Indices
in a Comparative Analysis (CoMSIA) of Drug Molecules to Cor-
relate and Predict Their Biological Activity, J. Med. Chem., 37,
4130–4146 (1994) A. R. Ortiz, M.T. Pisabarro, F. Gago and R.C. Wade, Prediction of
Drug Binding Afnities by Comparative Binding Energy Analysis,
J. Med. Chem., 38, 2681–2691 (1995) H. Gohlke and G. Klebe, DrugScore Meets CoMFA: Adaptation of
Fields for Molecular Comparison (AFMoC) or How to Tailor
Knowledge-based Pair-Potentials to a Particular ProteinJ. Med.
Chem., 45, 4153–4170 (2002)
. • Bibliography and Further Reading
A. Weber, M. Böhm, C. T. Supuran, A. Scozzafava, C. A. Sotriffer and
G.Klebe, 3D QSAR Selectivity Analyses of Carbonic Anhydrase Inhibitors: Insights for the Design of Isozyme Selective Inhibitors, J. Chem. Inf. Model., 46, 2737–2760 (2006)
A. Hillebrecht, C. T. Supuran and G. Klebe. Integrated Approach Us-
ing Protein and Ligand Information to Analyze Afnity and Se­lectivity Determining Features of Carbonic Anhydrase Isozymes, ChemMedChem, 1, 839–853 (2006)


From In Vitro to In Vivo:
Optimization of ADME and Toxicology Properties
Contents
19.1 Rate Constants of Compound Transport – 292
19.2 Absorption of Organic Molecules: Model and Experimental Data – 294
19.3 The Role of Hydrogen Bonds – 294
19.4 Distribution Equilibria of Acids and Bases – 295


19.5
19.6 What Is the Optimal Lipophilicity of aDrug? – 298
19.7 Computer Models and Rules to Predict ADME
19.8 From In Vitro to In Vivo Activity – 300
19.9 Compartmentalization: Natural Ligands
19.10 Specicity and Selectivity of Drug Interactions – 301
19.11 Of Mice and Men: The Value of Animal Models – 302
19.12 Toxicity and Adverse Eects – 304
19.13 Animal Protection and Alternative Test Models – 306
19.14 Synopsis – 306
Absorption Proles of Acids and Bases – 296
Parameters – 299
Are Often Unspecic – 300
Bibliography and Further Reading – 308
© The Author(s), under exclusive license to Springer-Verlag GmbH, DE, part of Springer Nature 2024 G. Klebe, Drug Design, https://doi.org/10.1007/978-3-662-68998-1_19