Tetrahedron Computer Methodology最新文献

We describe a large scale application of a back-propagation neural network to the analysis, classification and prediction of protein secondary and tertiary structure from sequence information alone. A back-propagation network called BigNet has been implemented along with a Network Description Language (NDL) on the 512 MWord Cray 2 at the Minnesota Supercomputer Center. The proof-of-concept experiments described here used a small, heterologous training set of small protein structures (15 proteins each with less than 133 residues) from the Brookhaven Protein Data Bank (PDB). Simulations with one hidden layer and one half to ten million connections execute at three to five million connection updates per second in full back-propagation learning mode and routinely converge to solutions where input of hydrophobicity-coded sequence yields output distance matrices with 0.3 to 1.5% RMS deviation from actual distance matrices. Although the training set used is too small to expect useful generalization, some evidence of generalization was evident in similarity of learning progress of homologous pairs within the training set and in production of novel distance matrix outputs upon presentation with novel input sequences. The discussion addresses limitations in the current implementation, plans for software improvements, and characteristics of future training sets.

Organic chemists, when creating and planning the synthesis of new molecules, use a cognitive process known as “retrosynthetic analysis”, along with an ordered set of interrelations between various chemical elements, compounds, and properties embodied in what is termed as “chemical intuition”. Retrosynthetic analysis can be represented as a pattern recognition problem, and as such, can be modelled using neural networks, in which ordered representations (which model “chemical intuition”) of molecular subunits and features are used as input. Actual results from one neural network which performs tactical disconnections α and β to carbonyl groups are presented and discussed, as well as the structure for the hybrid expert system, incorporating a collection of neural nets. The program is included, along with all files needed to run ACME.

Chemical Abstracts Service (CAS) is exploring approaches for searching on 3D and related molecular property data for CAS Registry substances. This searching includes “fuzzy-match” similarity searching. As the first part of this effort, the 3D coordinates have been generated by the CONCORD program for sample files of Registry substances, including a file of ring system, “framework” substances. In addition, molecular property data such as partial atom charges and ionization potentials have been derived from the corresponding 2D and/or 3D data via computational chemistry programs. Experimental software is being developed to identify, analyze and search various characteristics of 3D and molecular property data. These characteristics include 3D structural, flexibility, shape and molecular property features for portions of a substance and/or the entire substance. This paper will discuss some preliminary results of the analysis and searching of these data characteristics.

A system has been developed that can automatically generate three-dimensional atomic coordinates from the constitution of a molecule as expressed by a connection table. The program, CORINA, is applicable to the entire range of organic chemistry. It can also handle structures that are beyond the scope of some other programs, $\text{e.g.}$, macrocyclic and polymacrocyclic molecules. Computation times are short and the results compare favorably with data from X-ray crystallography and with those of molecular mechanics calculations.

The MINDO/3, MNDO, and AM1 geometries for five and six membered heteroaromatics have been compared with available experimental data and with some ab initio geometries. Geometry optimizations using the AM1 and MNDO methods gave the best results of the semi-empirical methods examined and yielded molecular geometries in good agreement with available experimental bond angles of all types and for C-C, C-N, C-O and C-S bond distances. The AM1 and MNDO calculated that N-N, N-O and C=S bond distances are significantly shorter than experimental values due to a systematic error. In general, AM1 ring geometries provide a reliable estimate for the majority of heteroaromatic compounds.

The Bird I₆ and I₅ aromaticity indices calculated from semiempirical and ab initio geometries are compared with those calculated from experimental bond lengths. None of these semiempirical theoretical methods are successful for rings when the number of heteroatoms exceed the number of carbon atoms. For other heterocycles, AM1 and ab initio 3–21G basis set give the best results, followed by MNDO and then by MINDO/3. Rings containing carbonyl groups are an exception in that MINDO/3 provides the best 16 estimates.

A molecular input/output data file conversion utility for the IBM PC is presented. The program included on disk in this issue reads and writes the file format of CPSS, Alchemy, Sybyl, Molidea, Desktop MM, PCMODEL, MGP+ modeling software as well as Z-matrix, fractional X-ray and Cartesian coordinates. In addition to conversion, the molecules can be inspected and some geometric data can be calculated. Other utilities are also included.

A class of associative memory Hamiltonians for protein tertiary recognition was recently introduced by us. By using a minimization scheme based on molecular dynamics with simulated annealing, we are able to improve and expand upon those initial results. For small proteins lower bound estimates of the Hamiltonians' capacity (the maximum size database for which the Hamiltonian has the ability to reproduce structures) are given; in addition, studies of the dependence of this capacity on various global parameters, such as the choice of sequence encodings, the rate of tolerable mutations in the sequence, and the range of active interactions, are reported. The introduction of the molecular dynamics procedure also permits estimates of the capacity for medium-sized proteins (125–200 residues) to be made. These results demonstrate that the capacity for the simplest realizations of the associative memory Hamiltonians grows as 0.5-0.7N, where N is the number of amino acid residues of the protein to be recalled.

Statistical and numerical methods for 3D pattern recognition and classification are now commonly used to acquire structural knowledge from the rapidly growing crystallographic databases. These methods are reviewed, and a semantic network model is described for the storage of crystal and molecular knowledge. The knowledge base is central to an integrated computational strategy for molecular scene analysis: the reconstruction and interpretation of 3D molecular structures and molecular interactions.