Identification by golpe of 3D regions affecting the inhibition power of a series of glucose analogues

*Institut de Chimie Therapeutique, Ecole de Pharmacie, Universite de Lausanne, Lausanne, Switzerland tLaboratorium Fur Physikalische Chemie, ETH-Zentrum, Zurich, Switzerland Despite a decade of intensive studies using a battery of spectroscopic and theoretical techniques, the conformational behavior of cyclosporin A (CsA), an immunosuppressive cyclic undecapeptide widely used in clinical organ transplantation, is still unresolved. In the present study, the partition coefficient of CsA was measured in octanoliwater and heptane/water by centrifugal partition chromatography. By comparison with results from model compounds, it was deduced that the hydrogen-bonding capacity of CsA changes dramatically from apolar solvents (where it is internally Hbonded) to polar solvents (where it exposes its H-bonding groups to the solvent). Molecular dynamics simulations of long duration (up to 400 ps) in water and Ccl, confirm the tendency of CsA to undergo these solvent-dependent conformational changes, although the entire process appears to take a much longer time than can possibly be simulated by current computational means. These results are consistent with recent NMR studies of a water-soluble CsA derivative, [D-diaminobutyric acid]CsA, which indicate the presence of several different slowly interconverting conformations in aqueous solution.

REFERENCES 1 Theriault, Y., Logan, T.M., Meadows, I., Yu, L., Olejniczak, E.T., Holzman, T.F., Simmer, R.L., and Fesik, S.W. Nature 1993, 361, 88-91

AUTOMATED PHARMACOPHORE IDENTIFICATION AND ACTIVITY CLASSIFICATION USING APEX-3D E.R. Vorpagel BIOSYM Technologies,

Inc., San Diego, CA, USA

The utility of a new tool for identifying pharmacophores in a diverse set of biologically active structures was evaluated. Apex-3D is an expert system that uses a logico-structural approach to identify possible pharmacophores which are composed of various atom- or pseudoatom-centered descriptor centers, and the distances between these centers. Advanced statistical techniques and three-dimensional patternmatching algorithms are used to assign probabilities and reliabilities to the identified pharmacophores. This information is stored in the form of rules that can be used to predict the activity of novel structures. Qualitative activity classification and quantitative activity prediction are possible based on pharmacophores identified in novel structures. Several activity classes were studied, including antipicornavirus agents, angiotensin-converting enzyme inhibitors, angiotensin II receptor antagonists, ergosterol biosynthesis inhibitors, and allosteric hemoglobin modifiers. These were chosen because much structure-activity work has been

published, including several ligand-receptor crystal structures. Multiple conformations for compounds representing varying levels of activity for each of these activity classes were analyzed by Apex-3D. Knowledge bases containing rules based on pharmacophores identified for these biological activities were automatically generated by Apex-3D. Sets of pharmacophores identified for each activity class were visually inspected. For each activity class, pharmacophores had been identified which were consistent with either previously proposed models or actual ligand-receptor crystal structures. Apex-3D was also able to correctly classify compounds not included in the training set. For cases where good binding data was available, Apex-3D was able to predict binding activity with reasonable success. Apex3D provides a valuable tool for automated pharmacophore identification. It can use nonideal activity data, as well as good quantitative data. Synergy between Apex-3D and traditional QSAR will provide very powerful support for any scientist trying to efficiently develop new biologically active molecules.

IDENTIFICATION BY GOLPE OF 3D REGIONS AFFECTING THE INHIBITION POWER OF A SERIES OF GLUCOSE ANALOGUES G. Cruciani,” S. Clementi,* and K. Wood? *Dipartimento di Chimica, Universita di Perugia, Perugia, Italy *Laboratory of Molecular Biophysics, Oxford University, Oxford, UK Comparative molecular field analysis has become widely used in drug design and QSAR. The approach provides fast and cheap acquisition of a large number of quantitative descriptors, and uses PLS methods to correlate changes in biological activity with changes in chemical structure for series of drugs. At present, the state of the art in 3D-QSAR is represented by the CoMFA procedure of SYBYL. An appropriate alternative is the GRID program,’ which is limited to the three-dimensional (3D) description, but is supposed to give more reliable interaction energy values. One of the characteristics of the 3D-QSAR data matrix is the huge number of variables considered as the nonbonded interaction energies between one or more probes and each drug molecule. Some variables stand out as obvious, while others, which are less obvious, can often produce subltle changes in the final model or have a secondary but not easily distinguishable effect on the obvious variables. Moreover, it is well known that when variables are many, keeping irrelevant variables in the model is detrimental to the model’s predictive ability. A rapid screening method for selecting the most important variables (regions) is very desirable. Recent developments in statistics provide us with a new, interesting set of measures of validity that are based on simulating the predictive power of a model. By means of these tools, it is possible to estimate the set of parameters (variables or regions, and

J. Mol. Graphics,

1993, Vol. 11, December

261

dimensionality) able to maximize the predictive power of the model. GOLPE’ is an advanced variable selection procedure aimed at obtaining PLS regression models with the highest predictive ability. The variable selection is performed taking into account the effects of individual variables on the predictive power. GOLPE can be used in 3D-QSAR studies on the fields generated by CoMFA, as well as on any data set obtained by a variety of GRID probes, or similarity indices.’ Although this procedure seems to be a very powerful tool in evaluating the important variables which contribute in a positive way to the predictive ability, GOLPE results depend on data pretreatments. In this study, the different effects of data pretreatments on the predictive ability and the aspects of the selection of variables are discussed. The procedure was validated in a series of 36 glucose analogue inhibitor compounds whose X-ray structures bound to glycogen phosphorylase b enzyme has been determined,4 to overcome the aligment problem and in order to check the variables (regions) predicted as important for inhibition capability, which are well known from X-ray crystallographic studies. The results demonstrate how little GOLPE is affected by chance correlations; i.e., the procedure minimizes the risks of overfitting and overprediction. Moreover, a suitable data pretreatment can improve the predictive power, as well as the goodness and stability of all other chemometric regression models.

REFERENCES Boobbyer, D.N.A., Goodford, P.J., McWhinnie, P.M., and Wade, R.C. J. Med. Chem. 1989, 32, 1083- 1094 Baroni, M., Costantino, G., Cruciani, G., Riganelli, D., Valigi, R., and Clementi, S. Quant. Struct.-Act. Relat. 1993, 1, in press Good, A.C., So, S.S., and Richards, W.G. J. Med. Chem. 1993, 36, 433-438 Martin, J.L., Veluraja, K., Ross, K., Johnson, L.N., Fleet, G.W.J., Ramsden, N.G., Bruce, I., Ochard, M.G., Oikonomakos, N.G., Papageorgiou, A.C., Leonidas, D.D., and Tsitoura, H.S. Biochemistry. 1991, 30, 10101-10116

3D-QSAR: A COMPARISON OF TECHNIQUES APPLIED TO A SERIES OF BIOLOGICALLY ACTIVE MOLECULES OF DIVERSE STRUCTURE I.M. McLay and J.S. Mason Dagenham Research Centre, Dagenham, UK

Rhone

Poulenc

Rorer Ltd,

The QSAR of a series of biologically active molecules of diverse structure was considered using multiple linear regression (MLR) and partial least-squares projection to latent structures (PLS) techniques. The structures possess an invariant portion (Group A), ignored for the purposes of this work, linked by an aliphatic chain to a variant portion (Group B) used for the analyses. A MLR relationship was developed based on a combination of the “classical” QSAR parameters CMR and (cLogP)2, and the more unusual parameter pcos0, representing the dipole component in a defined direction relative to the bond linking Group A and Group B: Log( l/fC,,,) = 0.59( ? 0. 15)CMR + 0.38( ?O. I~)E.Lcos(@ - 0.16( &O.O7)cLogP’ SD0.41 F = 22R’ = 0.81

- 10.05

PLS relationships using molecular fields derived from Sybyl, GRID, and HINT (and combinations of these) were then developed using Sybyl-CoMFA and GOLPE. Predictive models were found using both SYBYL (stericielectronic C.r&) and GRID (CH,, NH,‘, and O-) probes; however, the GOLPE procedure, which preprocesses the data to remove variables that do not contribute to a predictive model, was found to be essential for derivation of the GRID model. The MLR and PLS relationships were compared by considering the standard error of prediction (SE) for the activity of a test set of 5 compounds excluded from the analysis. The results for the MLR and best PLS models are shown in Table 1. The number of PLS components (D) and the predictive Q2 (by the leave-one-out procedure) are also shown. Interestingly, a relationship (R’ 0.8) was found between CA4R and the object scores of the first PLS component derived for the GRID model. For this set of compounds, the PLS models were found to

Table 1. MLR and PLS models

Compound

Measured

SE

Q”

MLR Pred.

GolpeiGRID Pred

SybylPred.

0.4 0.55

0.35 0.92 2 Log(lC,,) 0.70 1.20 2.01 2.01 2.60

0.33 0.61 3 Log(lC,,) 0.60 1.42 1.97 1.89 2.60

D

bm,,) A B C D E

262

J. Mol. Graphics,

0.70 1.60 1.62 2.33 3.06

1993,

Vol. 11, December

bw,,) 0.57 1.16 1.82 2.44 3.68