Bruce L. Bush

Fast, efficient generation of high‐quality atomic charges. AM1‐BCC model: I. Method

The AM1‐BCC method quickly and efficiently generates high‐quality atomic charges for use in condensed‐phase simulations. The underlying features of the electron distribution including formal charge and delocalization are first captured by AM1 atomic charges for the individual molecule. Bond charge corrections (BCCs), which have been parameterized against the HF/6‐31G* electrostatic potential (ESP) of a training set of compounds containing relevant functional groups, are then added using a formalism identical to the consensus BCI (bond charge increment) approach. As a proof of the concept, we fit BCCs simultaneously to 45 compounds including O‐, N‐, and S‐containing functionalities, aromatics, and heteroaromatics, using only 41 BCC parameters. AM1‐BCC yields charge sets of comparable quality to HF/6‐31G* ESP‐derived charges in a fraction of the time while reducing instabilities in the atomic charges compared to direct ESP‐fit methods. We then apply the BCC parameters to a small “test set” consisting of aspirin, d‐glucose, and eryodictyol; the AM1‐BCC model again provides atomic charges of quality comparable with HF/6‐31G* RESP charges, as judged by an increase of only 0.01 to 0.02 atomic units in the root‐mean‐square (RMS) error in ESP. Based on these encouraging results, we intend to parameterize the AM1‐BCC model to provide a consistent charge model for any organic or biological molecule.

Sample-distance partial least squares: PLS optimized for many variables, with application to CoMFA

SummaryThree-dimensional molecular modeling can provide an unlimited number m of structural properties. Comparative Molecular Field Analysis (CoMFA), for example, may calculate thousands of field values for each model structure. When m is large, partial least squares (PLS) is the statistical method of choice for fitting and predicting biological responses. Yet PLS is usually implemented in a property-based fashion which is optimal only for small m. We describe here a sample-based formulation of PLS which can be used to fit any single response (bioactivity). SAMPLS reduces all explanatory data to the pairwise ‘distances’ among n sample (molecules), or equivalently to an n-by-n covariance matrix C. This matrix, unmodified, can be used to fit all PLS components. Furthermore, SAMPLS will validate the model by modern resampling techniques, at a cost independent of m. We have implemented SAMPLS as a Fortran program and have reproduced conventional and cross-validated PLS analyses of data from two published studies. Full (leaveach-out) cross-validation of a typical CoMFA takes 0.2 CPU s. SAMPLS is thus ideally suited to structure-activity analysis based on CoMFA fields or bonded topology. The sample-distance formulation also relates PLS to methods like cluster analysis and nonlinear mapping, and shows how drastically PLS simplifies the information in CoMFA fields.

PATTY: A programmable atom type and language for automatic classification of atoms in molecular databases

PATTY (Programmable ATom TYper) is an algorithm for assigning “atom types” to a molecule based on its connection table of atoms and bonds. Its operation is controlled entirely by a rules file created by the user. Each rule contains a “pattern”, a description in linear notation of a class of chemical substructures which may contain branches or rings. Rules are of two sorts: those which define properties to be used in subsequent rules and those which assign the final atom types. As an example, we present rules for classifying atoms into seven broad types based on their physical properties (e.g. cation, H-bond donor, hydrophobe). This classification has been applied to large databases of three-dimensional modes of druglike compounds. PATTY is very rapid and operates on a variety of hardware platforms.

Extending the trend vector: the trend matrix and sample-based partial least squares.

SummaryTrend vector analysis [Carhart, R.E. et al., J. Chem. Inf. Comput. Sci., 25 (1985) 64], in combination with topological descriptors such as atom pairs, has proved useful in drug discovery for ranking large collections of chemical compounds in order of predicted biological activity. The compounds with the highest predicted activities, upon being tested, often show a several-fold increase in the fraction of active compounds relative to a randomly selected set. A trend vector is simply the one-dimensional array of correlations between the biological activity of interest and a set of properties or ‘descriptors’ of compounds in a training set. This paper examines two methods for generalizing the trend vector to improve the predicted rank order. The trend matrix method finds the correlations between the residuals and the simultaneous occurrence of descriptors, which are stored in a two-dimensional analog of the trend vector. The SAMPLS method derives a linear model by partial least squares (PLS), using the ‘sample-based’ formulation of PLS [Bush, B.L. and Nachbar, R.B., J. Comput.-Aided Mol. Design, 7 (1993) 587] for efficiency in treating the large number of descriptors. PLS accumulates a predictive model as a sum of linear components. Expressed as a vector of prediction coefficients on properties, the first PLS component is proportional to the trend vector. Subsequent components adjust the model toward full least squares. For both methods the residuals decrease, while the risk of overfitting the training set increases. We therefore also describe statistical checks to prevent overfitting. These methods are applied to two data sets, a small homologous series of disubstituted piperidines, tested on the dopamine receptor, and a large set of diverse chemical structures, some of which are active at the muscarinic receptor. Each data set is split into a training set and a test set, and the activities in the test set are predicted from a fit on the training set. Both the trend matrix and the SAMPLS approach improve the predictions over the simple trend vector. The SAMPLS approach is superior to the trend matrix in that it requires much less storage and CPU time. It also provides a useful set of axes for visualizing properties of the compounds. We describe a randomization method to determine the optimum number of PLS components that is very much faster for large training sets than leave-one-out cross-validation.

Consensus bond-charge increments fitted to electrostatic potential or field of many compounds: Application to MMFF94 training set

Bond‐charge increments (BCIs) are additive parameters used to assign atomic charges for the MMFF force field. BCI parameters are classified parsimoniously according to two atom types and the bond order. We show how BCIs may be fitted rapidly by linear least squares to the calculated ab initio electrostatic potential (ESP) or to the electrostatic field. When applied simultaneously to a set of compounds or conformations, the method yields consensus values of the BCIs. The method can also derive conventional “ESP‐fit” atomic charges with improved numerical stability. The method may be generalized to determine atom multipoles, multicenter charge templates, or electronegativities, but not polarizability or hardness. We determine 65 potential‐derived (PD) BCI parameters, which are classified as in MMFF, by fitting the 6‐31G* ESP or the electrostatic field of the 45 compounds in the original MMFF94 training set. We compare the consensus BCIs with classified BCIs that were fit to each molecule individually and with “unique‐bond” BCIs (ESP‐derived atom charges). Consensus BCIs give a satisfactory representation for about half of the structures and are robust to the adjustment of the alkyl CH bond increment to the zero value employed in MMFF94. We highlight problems at three levels: Point approximation: the potential near lone pairs on sulfur and to some extent nitrogen cannot be represented just by atom charges. Bond classification: BCIs classified according to MMFF atom types cannot represent all delocalized electronic effects. The problem is especially severe for bonds between atoms of equivalent MMFF type, whose BCI must be taken as zero. Consensus: discrepancies that occur in forming the consensus across the training set indicate the need for a more detailed classification of BCIs. Contradictions are seen (e.g., between acetic acid and acetone and between guanidine and formaldehydeimine). We then test the three sets of PD‐BCIs in energy minimizations of hydrogen‐bonded dimers. Unique‐bond BCIs used with the MMFF buffered 14–7 potential reproduce unscaled quantum chemical dimer interaction energies within 0.9 kcal/mol root mean square (or 0.5, omitting two N‐oxides). These energies are on average 0.7 (or 0.5) kcal/mol too weak to reproduce the scaled quantum mechanical (SQM) results that are a benchmark for MMFF parameterization. Consensus BCIs tend to weaken the dimer energy by a further 0.4–0.6 kcal/mol. Thus, consensus PD‐BCIs can serve as a starting point for MMFF parameterization, but they require both systematic and individual adjustments. Used with a “harder” AMBER‐like Lennard–Jones potential, unique‐bond PD‐BCIs without systematic adjustment give dimer energies in fairly good agreement with SQM. ©1999 John Wiley & Sons, Inc. J Comput Chem 20: 1495–1516, 1999

Recognition by HLA-A2-restricted cytotoxic T lymphocytes of endogenously generated and exogenously provided synthetic peptide analogues of the influenza a virus matrix protein

Experiments were carried out to determine whether complexes between MHC class I molecules and synthetic peptides are representative of those formed under more physiologically relevant conditions, with peptides derived intracellularly from processed antigens. Lysis of cells sensitized with exogenously provided and endogenously generated peptide analogues of the optimal nonameric peptide 58-66 (GILGFVFTL; derived from the influenza virus matrix protein) was compared. Endogenous loading was accomplished by expressing minigene DNA coding for alanine-substituted analogues of peptide 58-66 in HLA-A2-positive cells. Susceptibility to lysis by HLA-A2-restricted, peptide-specific cytotoxic lymphocytes was compared with lysis of cells sensitized with the same synthetic peptides. Although results were quite comparable, differences were observed. The endogenously presented analogues 58-66L60A, G61A, T65A, and L66A were recognized more efficiently than the corresponding exogenously presented analogues. This difference in recognition was most striking for peptide 58-66G61A. These results indicate the need for caution in using synthetic peptides in defining peptide binding motifs. Additional experiments with endogenously expressed analogues of 58-66 with substitutions other than alanine were carried out to define the interaction between this peptide and HLA-A2. Results are compatible with the interpretation that residues 58, 59, and 60 interact with pockets A, B, and D, respectively, in the HLA-A2 binding groove and that these interactions contribute to peptide binding.

Restrained point-charge models for disaccharides

Various methods for deriving atomic partial charges from the quantum chemical electrostatic potential and moments have been tested for the sucrose molecule. We show that if no further information is used, the charges on some carbon atoms become large and charge patterns involving these atoms are badly determined and poorly transferable. Adding lone‐pairs on the ether oxygen atoms or dividing the molecule into smaller fragments did not cure the instabilities. We develop a method, CHELP‐BOW0, that restrains charges toward zero with different weights for different atoms. These harmonic restraints preserve the linear form of the least‐squares equations, which are solved in a single step using singular‐value decomposition. CHELP‐BOW0 improves the chemical transferability of the charges compared to unrestrained methods, and slightly improves their conformational transferability. It introduces a modest degradation of the fit compared to unrestrained CHELP‐BOW (mean average deviation of the potential 0.00016 vs. 0.00010 a.u.). A second new method, CHELP‐BOWC, avoids the need for restraints by including several conformations in the fit, weighting each according to its estimated energy in solution. CHELP‐BOWC charges are more transferable than CHELP‐BOW or CHELP‐BOW0 charges to conformations not included in the training set. Restraints to zero charge do not further improve transferability of the CHELP‐BOWC charges. We, therefore, recommend CHELP‐BOW charges for rigid molecules and CHELP‐BOWC charges for flexible molecules.

Modeling the Mechanism of Peptide Cleavage by Thermolysin

Merck & Co. has for some time utilized a molecular modeling system that we developed for aiding drug design.’.’ More secently, it became clear that more powerful systems were needed to study the relevant larger structures which were becoming increasingly available, and to study steric and electrostatic surfaces. We have obtained such equipment and utilized it for drug design. At the same time, realizing that many of the modeling functions can be performed on less expensive equipment, we have modified our program to run on several types of graphics terminals. Characteristics of our current system for molecular modeling, and some applications to drug design, are discussed below.

Enhanced homology searching through genome reading frame predetermination

MOTIVATION Many bioinformatic approaches exist for finding novel genes within genomic sequence data. Traditionally, homology search-based methods are often the first approach employed in determining whether a novel gene exists that is similar to a known gene. Unfortunately, distantly related genes or motifs often are difficult to find using single query-based homology search algorithms against large sequence datasets such as the human genome. Therefore, the motivation behind this work was to develop an approach to enhance the sensitivity of traditional single query-based homology algorithms against genomic data without losing search selectivity. RESULTS We demonstrate that by searching against a genome fragmented into all possible reading frames, the sensitivity of homology-based searches is enhanced without degrading its selectivity. Using the ETS-domain, bromodomain and acetyl-CoA acetyltransferase gene as queries, we were able to demonstrate that direct protein-protein searches using BLAST2P or FASTA3 against a human genome segmented among all possible reading frames and translated was substantially more sensitive than traditional protein-DNA searches against a raw genomic sequence using an application such as TBLAST2N. Receiver operating characteristic analysis was employed to demonstrate that the algorithms remained selective, while comparisons of the algorithms showed that the protein-protein searches were more sensitive in identifying hits. Therefore, through the overprediction of reading frames by this method and the increased sensitivity of protein-protein based homology search algorithms, a genome can be deeply mined, potentially finding hits overlooked by protein-DNA searches against raw genomic data.

Design of Renin Inhibitors Containing Conformationally Restricted Mimetics of the P1-P1′ and P1 through P2′ Sites

The clinical efficacy of converting enzyme inhibitors1 for reducing blood pressure in a large percentage of hypertensive patients has aroused considerable interest in developing agents that interrupt the renin-angiotensin system at other points, for example by blockade of the angiotensin II receptor2 or by inhibition of the aspartic proteinase, renin. Substrate based design of renin inhibitors in which the scissile P1-P1′ dipeptide is replaced with a non-hydrolyzable group, often a mimetic of a tetrahedral transition state for amide bond hydrolysis, has provided a useful approach for obtaining a variety of inhibitor structure types,3 many with Ki’s of better than 10−9 M. A clinically useful renin inhibitor, however, has not yet emerged due to poor pharmacokinetics (i.e., metabolism or rapid clearance) or poor oral absorption, problems often encountered with peptidic drug targets. An example of this is seen with 1, a “tetrapeptide” inhibitor4 that spans the P4 through P3′ sites of the renin substrate and which utilizes the statine analog, ACHPA5, as a P1-P1′ dipeptide replacement (Figure 1). Although 1 is quite potent in vitro, very low levels of drug are found in the blood after oral administration to the rhesus monkey at 50 mg/kg, and the half life after intravenous administration is short (< 1 h). Rapid biliary excretion of intact drug has been demonstrated for a number of other renin inhibitors.6