Comparison of a neural net-based QSAR algorithm (PCANN) with Hologram- and multiple linear regression-based QSAR approaches: application to 1,4-dihydropyridine-based calcium channel antagonists.

A QSAR algorithm (PCANN) has been developed and applied to a set of calcium channel blockers which are of special interest because of their role in cardiac disease and also because many of them interact with P-glycoprotein, a membrane protein associated with multidrug resistance to anticancer agents. A database of 46 1,4-dihydropyridines with known Ca2+ channel binding affinities was employed for the present analysis. The QSAR algorithm can be summarized as follows: (1) a set of 90 graph theoretic and information theoretic descriptors representing various structural and topological characteristics was calculated for each of the 1,4-dihydropyridines and (2) principal component analysis (PCA) was used to compress these 90 into the eight best orthogonal composite descriptors for the database. These eight sufficed to explain 96% of the variance in the original descriptor set. (3) Two important empirical descriptors, the Leo-Hansch lipophilic constant and the Hammet electronic parameter, were added to the list of eight. (4) The 10 resulting descriptors were used as inputs to a back-propagation neural network whose output was the predicted binding affinity. (5) The predictive ability of the network was assessed by cross-validation. A comparison of the present approach with two other QSAR approaches (multiple linear regression using the same variables and a Hologram QSAR model) is made and shows that the PCANN approach can yield better predictions, once the right network configuration is identified. The present approach (PCANN) may prove useful for rapid assessment of the potential for biological activity when dealing with large chemical libraries.

A knowledge-based approach in designing combinatorial or medicinal chemistry libraries for drug discovery. 1. A qualitative and quantitative characterization of known drug databases.

The discovery of various protein/receptor targets from genomic research is expanding rapidly. Along with the automation of organic synthesis and biochemical screening, this is bringing a major change in the whole field of drug discovery research. In the traditional drug discovery process, the industry tests compounds in the thousands. With automated synthesis, the number of compounds to be tested could be in the millions. This two-dimensional expansion will lead to a major demand for resources, unless the chemical libraries are made wisely. The objective of this work is to provide both quantitative and qualitative characterization of known drugs which will help to generate "drug-like" libraries. In this work we analyzed the Comprehensive Medicinal Chemistry (CMC) database and seven different subsets belonging to different classes of drug molecules. These include some central nervous system active drugs and cardiovascular, cancer, inflammation, and infection disease states. A quantitative characterization based on computed physicochemical property profiles such as log P, molar refractivity, molecular weight, and number of atoms as well as a qualitative characterization based on the occurrence of functional groups and important substructures are developed here. For the CMC database, the qualifying range (covering more than 80% of the compounds) of the calculated log P is between -0.4 and 5.6, with an average value of 2.52. For molecular weight, the qualifying range is between 160 and 480, with an average value of 357. For molar refractivity, the qualifying range is between 40 and 130, with an average value of 97. For the total number of atoms, the qualifying range is between 20 and 70, with an average value of 48. Benzene is by far the most abundant substructure in this drug database, slightly more abundant than all the heterocyclic rings combined. Nonaromatic heterocyclic rings are twice as abundant as the aromatic heterocycles. Tertiary aliphatic amines, alcoholic OH and carboxamides are the most abundant functional groups in the drug database. The effective range of physicochemical properties presented here can be used in the design of drug-like combinatorial libraries as well as in developing a more efficient corporate medicinal chemistry library.

An information-intensive approach to the molecular pharmacology of cancer.

Since 1990, the National Cancer Institute (NCI) has screened more than 60,000 compounds against a panel of 60 human cancer cell lines. The 50-percent growth-inhibitory concentration (GI50) for any single cell line is simply an index of cytotoxicity or cytostasis, but the patterns of 60 such GI50 values encode unexpectedly rich, detailed information on mechanisms of drug action and drug resistance. Each compound's pattern is like a fingerprint, essentially unique among the many billions of distinguishable possibilities. These activity patterns are being used in conjunction with molecular structural features of the tested agents to explore the NCI's database of more than 460,000 compounds, and they are providing insight into potential target molecules and modulators of activity in the 60 cell lines. For example, the information is being used to search for candidate anticancer drugs that are not dependent on intact p53 suppressor gene function for their activity. It remains to be seen how effective this information-intensive strategy will be at generating new clinically active agents.

An approach to rapid estimation of relative binding affinities of enzyme inhibitors: application to peptidomimetic inhibitors of the human immunodeficiency virus type 1 protease.

This report describes a method for rapid assessment of the binding affinities of a series of analogous ligands to an enzyme. This approach is based on two variables (scores), representing (i) the enthalpy of binding and (ii) the strength of hydrophobic interaction. The method is then used to evaluate the binding of 11 different peptidomimetic inhibitors to the HIV-1 protease. Three-dimensional structures of these enzyme-inhibitor complexes are modeled based on the crystal structures of HIV-1 protease complexes with the known inhibitors. These structures are minimized using the AMBER force field, and the scores of binding enthalpy for each of the ligands are calculated. A second score to represent the hydrophobic interaction between a pair of atoms uses an exponential function of distance between the atoms and the product of their atomic hydrophobicity constants. This exponential function is used to assess the hydrophobic interaction energy between an enzyme and its inhibitor and also to compute and display a 'molecular hydrophobicity map' as a 3D visualization tool. These methods are then applied to obtain trends in relative binding affinities of pairs of analogous inhibitors. Calculated scores agree well with corresponding results from thermodynamic cycle perturbation (TCP) simulations as well as experimental binding data. Since the proposed calculations are computationally cheaper and faster than TCP calculations, it is suggested that these scores can form the basis for rapid, preliminary theoretical screening of proposed derivatives of an inhibitor prior to TCP analysis, synthesis, and testing.

Calculation of relative differences in the binding free energies of HIV1 protease inhibitors: a thermodynamic cycle perturbation approach.

An iterative computer-assisted drug design (CADD) method that combines molecular mechanics, dynamics, thermodynamic cycle perturbation (TCP) calculations, molecular design, synthesis, and biochemical testing of peptidomimetic inhibitors and crystallographic structure determination of the protein-inhibitor complexes has been successfully applied to the design of novel inhibitors for the HIV1 protease. The first "designer" compound in this series (I) was designed by replacing the C-terminal Val-Val methyl ester of a known hydroxyethylene inhibitor with a diphenhydramine amide derivative in which two phenyl groups fill the p2' and p3' side-chain binding pockets in the HIV1 protease. Subsequent testing showed modest inhibition (Ki = 1.67 microM). Concurrently, molecular mechanics calculations on designed analogs indicated the feasibility of replacement of a phenyl ring with an indole ring (II). Synthesis and biochemical testing resulted in better inhibition potency for II. X-ray crystal structure determination of HIV1 protease complexed with I and II provided structural information for subsequent design and TCP calculations. A TCP protocol was established and validated for the mutation of I-->II. TCP results showed a net gain of 2.1 (+/- 0.9) kcal/mol in replacing II with I, which agreed with experimental result within an error margin of 0.8 kcal/mol. TCP calculations for six other mutations (I-->III, II-->III, IV, V, VI, and VII) were performed prior to synthesis and testing. These results allowed for the prioritization of design ideas for synthesis. In all cases where experimental results are available, TCP calculations showed good agreement. These results demonstrate that the TCP approach can be used with medicinal chemistry and crystallography for screening the proposed derivatives of a lead compound prior to synthesis, thus potentially reducing the time for the discovery of new drugs.

Predictive statistics and artificial intelligence in the U.S. National Cancer Institute's Drug Discovery Program for Cancer and AIDS.

The National Cancer Institute's drug discovery program screens more than 20,000 chemical compounds and natural products a year for activity against a panel of 60 tumor cell lines in vitro. The result is an information-rich database of patterns that form the basis for what we term an "information-intensive" approach to the process of drug discovery. The first step was a demonstration, both by statistical methods (including the program COMPARE) and by neural networks, that patterns of activity in the screen can be used to predict a compound's mechanism of action. Given this finding, the overall plan has been to develop three large matrices of information: the first (designated A) gives the pattern of activity for each compound tested against each cell line in the screen; the second (S) encodes any of a number of types of 2-D or 3-D structural motifs for each compound; the third (T) indicates each cell's expression of molecular targets (e.g., from 2-dimensional protein gel electrophoresis). Construction and updating of these matrices is an ongoing process. The matrices can be concatenated in various ways to test a variety of specific hypotheses about compounds screened, as well as to "prioritize" candidate compounds for testing. To aid in these efforts, we have developed the DISCOVERY program package, which integrates the matrix data for visual pattern recognition. The "information-intensive" approach summarized here in some senses serves to bridge the perceived gap between screening and structure-based drug design.

Neural computing in cancer drug development: predicting mechanism of action.

Described here are neural networks capable of predicting a drug's mechanism of action from its pattern of activity against a panel of 60 malignant cell lines in the National Cancer Institute's drug screening program. Given six possible classes of mechanism, the network misses the correct category for only 12 out of 141 agents (8.5 percent), whereas linear discriminant analysis, a standard statistical technique, misses 20 out of 141 (14.2 percent). The success of the neural net indicates several things. (i) The cell line response patterns are rich in information about mechanism. (ii) Appropriately designed neural networks can make effective use of that information. (iii) Trained networks can be used to classify prospectively the more than 10,000 agents per year tested by the screening program. Related networks, in combination with classical statistical tools, will help in a variety of ways to move new anticancer agents through the pipeline from in vitro studies to clinical application.

New joint prediction algorithm (Q7-JASEP) improves the prediction of protein secondary structure.

The classical problem of secondary structure prediction is approached by a new joint algorithm (Q7-JASEP) that combines the best aspects of six different methods. The algorithm includes the statistical methods of Chou-Fasman, Nagano, and Burgess-Ponnuswamy-Scheraga, the homology method of Nishikawa, the information theory method of Garnier-Osgurthope-Robson, and the artificial neural network approach of Qian-Sejnowski. Steps in the algorithm are (i) optimizing each individual method with respect to its correlation coefficient (Q7) for assigning a structural type from the predictive score of the method, (ii) weighting each method, (iii) combining the scores from different methods, and (iv) comparing the scores for alpha-helix, beta-strand, and coil conformational states to assign the secondary structure at each residue position. The present application to 45 globular proteins demonstrates good predictive power in cross-validation testing (with average correlation coefficients per test protein of Q7, alpha = 0.41, Q7, beta = 0.47, Q7,c = 0.41 for alpha-helix, beta-strand, and coil conformations). By the criterion of correlation coefficient (Q7) for each type of secondary structure, Q7-JASEP performs better than any of the component methods. When all protein classes are included for training and testing (by cross-validation), the results here equal the best in the literature, by the Q7 criterion. More generally, the basic algorithm can be applied to any protein class and to any type of structure/sequence or function/sequence correlation for which multiple predictive methods exist.

Relative differences in the binding free energies of human immunodeficiency virus 1 protease inhibitors: a thermodynamic cycle-perturbation approach.

Peptidomimetic inhibitors of the human immunodeficiency virus 1 protease show considerable promise for treatment of AIDS. We have, therefore, been seeking computer-assisted drug design methods to aid in the systematic design of such inhibitors from a lead compound. Here we report thermodynamic cycle-perturbation calculations (using molecular dynamics simulations) to compute the relative difference in free energy of binding that results when one entire residue (valine) is deleted from one such inhibitor. In particular, we studied the "alchemic" mutation of the inhibitor Ac-Ser-Leu-Asn-(Phe-Hea-Pro)-Ile-Val-OMe (S1) to Ac-Ser-Leu-Asn-(Phe-Hea-Pro)-Ile-OMe (S2), where Hea is hydroxyethylamine, in two different (R and S) diastereomeric configurations of the hydroxyethylene group. The calculated (averaged for R and S) difference in binding free energy [3.3 +/- 1.1 kcal/mol (mean +/- SD); 1 cal = 4.184 J] is in good agreement with the experimental value of 3.8 +/- 1.3 kcal/mol, obtained from the measured Ki values for an equilibrium mixture of R and S configurations. Precise testing of our predictions will be possible when binding data become available for the two disastereomers separately. The observed binding preference for S1 is explained by the stronger ligand-protein interaction, which dominates an opposing contribution arising from the large desolvation penalty of S1 relative to S2. This calculation suggests that the thermodynamic cycle-perturbation approach can be useful even when a relatively large change in the ligand is simulated and supports the use of the thermodynamic cycle-perturbation algorithm for screening proposed derivatives of a lead inhibitor/drug prior to their synthesis.

Analysis of the in vitro antitumor activity of novel purine-6-sulfenamide, -sulfinamide, and -sulfonamide nucleosides and certain related compounds using a computer-aided receptor modeling procedure.

The comparative antileukemic activities of 21 novel nucleosides were determined in vitro by using cultured L1210 cells and analyzed for structure-related efficacy by a computer-aided receptor modeling method (REMOTEDISC) as recently described (Ghose, A. K.; et al. J. Med. Chem. 1989, 32, 746). The algorithm can be classified as a 3D-QSAR method and consists of the following steps: selection of a reference structure from the low-energy conformations of the active compounds; an automated superposition of the low-energy conformations of the other compounds so that there is maximum matching (or overlapping) of the atom-based physicochemical properties; construction of the binding-site cavity from the location of the atoms of the superimposed molecules; and determinations of the relative importance of the various physicochemical properties at different regions of the site cavity using reverse stepwise regression analysis. The model was based on the minimum energy conformation of (R,S)-2-amino-9-beta-D-ribofuranosylpurine-6-sulfinamide (sulfinosine, 5), an effective antileukemic agent in vivo, in the data set. The model fit the biological data with a standard deviation of 0.363, a correlation coefficient of 0.933 and a explained variance of 0.815. The method targeted a syn conformation as the probable active form and the 2'-OH, 5'-OH as well as C2-NH2 group of the purine ring as favoring the stability of the syn conformation, thereby establishing the major contributions of these three molecular entities to overall antitumor activity.

Mapping the binding site of the nucleoside transporter protein: a 3D-OSAR study.

The nucleoside transporter is an intrinsic membrane protein that mediates salvage of nucleosides from the extracellular medium. In this report, its binding sites have been characterized by a 3D-QSAR (three-dimensional structure-directed quantitative structure-activity relationships) receptor mapping technique. REMOTEDISC. The algorithm is applied to a set of 19 nucleoside analogues, each of which binds to the transporter. The methodology includes: (i) conformational analysis of each ligand; (ii) estimation of physicochemical properties of each ligand at the atomic level; (iii) structural comparison of the low energy conformation of each ligand in the series with a reference structure on the basis of physicochemical property matching; (iv) construction of a predicted binding site cavity from the alignments of step (iii); and (v) multiple regression analysis of the binding data with respect to the 3-dimensional physicochemical descriptors in different 'site-pockets' of the binding cavity. The pharmacophore model that emerges consists of the geometry of the binding site cavity and the relative weights of various properties in different pockets for each of the ligands considered. The study suggests that binding free energy is sensitive to the composition, size and hydrophobicity of the heterocyclic base in the ligand. Though both syn and anti conformations are tried as active forms, the anti conformation gives a better solution and is chosen for modeling the binding site cavity. The best model obtained divides the binding site into six pockets and uses nine independent variables, fitting the observed data with a correlation coefficient of 0.94, a standard deviation of 0.22 and an explained variance of 0.80. Results of our model are consistent with a hypothesis that the 5'-OH group hydrogen bonds with the receptor. This model provides tentative design criteria for development of new nucleoside drugs and transport inhibitors. The model will undoubtedly continue to evolve (i) as the 3D-QSAR algorithm is further refined, and (ii) as data on additional nucleoside analogues become available.

Secondary structure of the human membrane-associated folate binding protein using a joint prediction approach.

The secondary structures of the human membrane-associated folate binding protein (FBP) and bovine soluble FBP are assessed by a joint prediction approach that combines neural network models, information theory, homology modeling and the Chou-Fasman methods. Two new profile maps are used to characterize the non-regular secondary structure and to assist in assigning buried and exposed parts of secondary structure: (i) the loop potential profile and (ii) the long range contact profile. Approximately half of human FBP is predicted to form regular secondary structure (alpha-helices-35% or beta-sheets - 12%, excluding the transmembrane helices) and the rest is predicted to form coil, turns or loops. The bovine milk soluble FBP is predicted to have a similar secondary structure as expected because of the high degree of homology between the FBP's. Discriminant analysis predicts two transmembrane segments for the human FBP sequence, one at the amino terminus (a leader sequence) and the other at the carboxy terminus. These predicted transmembrane domains are absent in the bovine milk soluble FBP, further supporting these predictions. The present set of secondary structural predictions for human FBP is obtained by 'consensus' to aid in modeling the super-secondary structure of the protein.

Structural mimicry of adenosine by the antitumor agents 4-methoxy- and 4-amino-8-(beta-D-ribofuranosylamino)pyrimido[5,4-d]pyrimidine as viewed by a molecular modeling method.

A rationale for the antitumor activity of 4-methoxy- and 4-amino-8-(beta-D-ribofuranosylamino)pyrimido-[5,4-d]pyrimidine (beta-MRPP and beta-ARPP, respectively) was studied by a molecular modeling method. Although these nucleoside analogues are structurally different from adenosine, they act as substrates for adenosine kinase. The molecular modeling method, which considered the three-dimensional structure and atom-based physicochemical properties of the nucleosides to quantify the molecular similarities, showed that certain low-energy conformations of the beta anomers of a series of nucleosides including beta-MRPP, beta-ARPP, and their 4-hydroxy, 4-amino-6-chloro, 4-methylthio-2,6-dichloro, 4,6-diamino, 4-dimethylamino, 4-methylamino, and 4-hydroxy-2,6-dichloro analogues have remarkable structural similarity to adenosine. The method also suggested that the selection of the reference compound adenosine in the structural comparison is of primary importance to gain insight into the observed antitumor activity. The success of the present method led to AM1 (Austin model 1) molecular orbital calculations and experimental studies indicating that the antitumor activity of the alpha anomer of ARPP is probably due to equilibration to the beta anomer. The AM1 calculation of the protonation energy of N5 of pyrimido[5,4-d]pyrimidines, which occupies the same position in space as the N1 of adenosine, gave a direct correlation between the basicity of the nitrogen with a lone pair of electrons and the observed antitumor activity.

Sidechain and backbone potential function for conformational analysis of proteins.

An improved potential function has been devised for the calculation of protein conformations. Each amino acid residue is represented by two points. The mainchain is traced by the sequence of C alpha atoms, and the details of sidechain structure and interactions are represented by a representative sidechain atom. This potential function has been developed from a data base of 22 high resolution protein crystal structures and includes the components of an earlier potential developed from a similar data base where each amino acid residue is represented by only its C alpha atom. In virtually all aspects of testing, the present potential betters the previous single-point potential, and is shown to be useful in the simulation of protein folding.

A potential function for conformational analysis of proteins.

We have devised a residue-residue potential function for low resolution protein conformational calculations. The interactions between residues near in sequence maintain correct secondary structure, while the long-range terms in the potential govern the larger packing features and overall globularity. The short-range terms were calculated by comparing the observed distributions of distances between C alpha atoms in 35 protein crystal structures to the expected distributions and assigning the discrepancies to a Boltzmann distribution due to an effective potential. Long-range terms were adjusted to ensure that the crystal structure of bovine pancreatic trypsin inhibitor has a lower total energy than perturbed conformations of the same molecule. Thus the empirical potential function implicity contains solvation and conformational entropy effects along with the usual Van der Waals and electrostatic energies. Extensive testing of the potential on trypsin inhibitor and other proteins establishes that it is generally applicable to small proteins, it does not attempt to compress or expand the conformations found by X-ray crystallography, standard secondary structural features are maintained under the potential, and there are so many local minima that local minimization can be trusted to return a perturbed structure to the native conformation only if they differ initially by less than 1 A.

Specificity of spatial clustering of amino acid residues as an evolutionary trend in proteins.

Three-dimensional structures of cytochrome c and myoglobin molecules together with amino acid substitution data on these proteins from a close range of taxa highlight an interesting evolutionary trend. Assuming three-dimensional structures of homologous proteins to be largely similar, the distribution of amino acid residue types around one another was studied. The application of information theory to three dimensional arrangement of amino acid residues revealed that in general, the higher the taxonomic order, the more specific is the clustering of residue types with one another.

Information contained in protein shapes.

The sequence of local conformations at C alpha atoms of a protein has been considered as an informational message string. The total self information contents and self information per letter have been evaluated for 83 globular proteins whose structures are known from X-ray crystallography. The derived information contents provide a method of quantitating structural specificity of proteins. This method of analysis enables repeating, intricate structural features to be recognized. Among the globular proteins whose structures have been solved, high potential iron protein stands out with the largest three-letter dependence.