Search

We consider Bayesian methodology for comparing two or more unlabeled point sets. Application of the technique to a set of steroid molecules illustrates its potential utility involving the comparison of molecules in chemoinformatics and bioinformatics. We initially match a pair of molecules, where one molecule is regarded as random and the other fixed. A type of mixture model is proposed for the point set coordinates, and the parameters of the distribution are a labeling matrix (indicating which pairs of points match) and a concentration parameter. An important property of the likelihood is that it is invariant under rotations and translations of the data. Bayesian inference for the parameters is carried out using Markov chain Monte Carlo simulation, and it is demonstrated that the procedure works well on the steroid data. The posterior distribution is difficult to simulate from, due to multiple local modes, and we also use additional data (partial charges on atoms) to help with this task. An approximation is considered for speeding up the simulation algorithm, and the approximating fast algorithm leads to essentially identical inference to that under the exact method for our data. Extensions to multiple molecule alignment are also introduced, and an algorithm is described which also works well on the steroid data set. After all the steroid molecules have been matched, exploratory data analysis is carried out to examine which molecules are similar. Also, further Bayesian inference for the multiple alignment problem is considered.

Several marine macrolide toxins act as potent and specific actin-severing molecules. Recent elucidation of their stereochemistries and modes of interaction with actin has allowed the syntheses of bioactive analogues. Here we used synthetic analogues in a structure-function analysis of ulapualide A, a trisoxazole-based macrolide. Ulapualide A harboured potent actin-depolymerising activity both in cells and in vitro. Its synthetic diastereoisomer was three orders of magnitude less active than the natural toxin and synthetic macrolide fragments lacked actin-capping/ severing activity altogether. Modulation of serum response factor (SRF)-dependent gene expression, as described for other actin-binding toxins, was also examined. Specific changes in response to ulapualide A were not observed, primarily due to its profound effects on cytoskeletal integrity and cell adhesion. Several synthetic fragments of ulapualide A also had no effect on SRF-dependent gene expression. However, inhibition was observed with a molecule corresponding to the extended aliphatic side chain of halichondramide, a structurally related macrolide. These findings indicate that side-chain derivatives of trisoxazole-based macrolides may serve to uncouple gene-regulatory events from actin dynamics.

We present a simple and effective method for similarity searching in virtual high-throughput screening, requiring only a string-based representation of the molecules (e.g., SMILES) and standard compression software, available on all modern desktop computers. This method utilizes the normalized compression distance, an approximation of the normalized information distance, based on the concept of Kolmogorov complexity. On representative data sets, we demonstrate that compression-based similarity searching can outperform standard similarity searching protocols, exemplified by the Tanimoto coefficient combined with a binary fingerprint representation and data fusion. Software to carry out compression-based similarity is available from our Web site at http://comp.chem.nottingham.ac.uk/download/zippity.

We present a comparative assessment of several state-of-the-art machine learning tools for mining drug data, including support vector machines (SVMs) and the ensemble decision tree methods boosting, bagging, and random forest, using eight data sets and two sets of descriptors. We demonstrate, by rigorous multiple comparison statistical tests, that these techniques can provide consistent improvements in predictive performance over single decision trees. However, within these methods, there is no clearly best-performing algorithm. This motivates a more in-depth investigation into the properties of random forests. We identify a set of parameters for the random forest that provide optimal performance across all the studied data sets. Additionally, the tree ensemble structure of the forest may provide an interpretable model, a considerable advantage over SVMs. We test this possibility and compare it with standard decision tree models.

Charge-transfer transitions in proteins play a key role in many biophysical processes, from the behavior of redox proteins to photochemical reactions. We present ab initio calculations on a model dipeptide and more approximate calculations of the electronic excited states of proteins which, taken together, provide the most definitive assignment and characterization of charge-transfer transitions in proteins to date. We have calculated from first principles the electronic circular dichroism (CD) spectra of 31 proteins on the basis of their structures. Compared to previous studies, we achieve more accurate calculated CD spectra between 170 and 190 nm, owing mainly to the importance in alpha-helices of a charge-transfer transition from the lone pair on one peptide group to the pi* orbital on the next peptide group.

In this paper different floating oscillator models for describing the amide I band of peptides and proteins are compared with density functional theory (DFT) calculations. Models for the variation of the frequency shifts of the oscillators and the nearest-neighbor coupling between them with respect to conformation are constructed from DFT normal mode calculations on N-acetyl-glycine-N(')-methylamide. The calculated frequencies are compared with those obtained from existing electrostatic models. Furthermore, a new transition charge coupling model is presented. We suggest a model which combines the nearest-neighbor maps with long-range interactions accounted for using the new transition charge model and an existing electrostatic map for long-range interaction frequency shifts. This model and others, which account for the frequency shifts by electrostatic maps exclusively, are tested by comparing the predicted IR spectra with those from DFT calculations on the pentapeptide [Leu]-enkephalin. The new model described above gives the best agreement and, after a systematic blueshift is accounted for, reproduces the DFT frequencies to within 3.5 cm(-1). The correlation of the intensities for this model with intensities from DFT calculations is 0.94.

Understanding the relationship between the amino acid sequence of a protein and its unique, compact three-dimensional structure is one of the grand challenges in molecular biophysics. One exciting approach to the protein-folding problem is fast time-resolved spectroscopy in the ultra-violet (UV). Time-resolved electronic circular dichroism (CD) spectroscopy offers resolution on a nanosecond (or faster) timescale, but does not provide the spatial resolution of techniques like X-ray crystallography or NMR. There is a need to underpin fast timescale spectroscopic studies of protein folding with a stronger theoretical foundation. We review some recent studies in this regard and briefly highlight how modern quantum chemical models of aromatic groups have improved the accuracy of calculations of protein CD spectra near-UV. On the other side of the far-UV, we describe calculations indicating that charge-transfer transitions are likely to be responsible for bands observed in the vacuum UV in protein CD.

We have applied time-dependent density functional theory (TDDFT) to study the valence pi-pi* excited states of the tryptophan chromophore in the environment of the proteins barnase and human serum albumin. The chromophore is represented by indole. Due to the approximate nature of TDDFT, in the gas phase the calculated vertical transition energies to the 1L valence states are reordered with respect to experiment. The 1L(a) state responds more than the 1L(b) state to the local environment, described fully at the TDDFT level, and to bulk environment, described by a set of point charges. Nevertheless, the vertical transitions are readily identified. For human serum albumin, our calculations predict distinct spectral characteristics between structures with different tryptophan side chain torsion angles. The computational tractability of TDDFT relative to more accurate ab initio methods allows a large part of the surrounding protein environment (up to 100 atoms) to be explicitly included in the TDDFT calculations.

We present a series of conformational search calculations on the aggregation of short peptide fragments that form fibrils similar to those seen in many protein mis-folding diseases. The proteins were represented by a face-centered cubic lattice model with the conformational energies calculated using the Miyazawa-Jernigan potential. The searches were performed using algorithms based on the Metropolis Monte Carlo method, including simulated annealing and replica exchange. We also present the results of searches using the tabu search method, an algorithm that has been used for many optimization problems, but has rarely been used in protein conformational searches. The replica exchange algorithm consistently found more stable structures then the other algorithms, and was particularly effective for the octamers and larger systems., (Copyright 2005 Wiley Periodicals, Inc.)

In order to probe the fundamental principles that govern protein evolution, we use a minimalist model of proteins to provide a mapping from genotype to phenotype. The model is based on physically realistic forces of protein folding and includes an explicit definition of protein function. Thus, we can find the fitness of a sequence from its ability to fold to a stable structure and perform a function. We study the fitness landscapes of these functional model proteins, that is, the set of all sequences mapped on to their corresponding fitnesses and connected to their one mutant neighbors. Through population dynamics simulations we directly study the influence of the nature of the fitness landscape on evolution. Populations are observed to move to a steady state, the distribution of which can often be predicted prior to the population dynamics simulations from the nature of the fitness landscape and a quantity analogous to a partition function. In this paper, we develop a scheme for predicting the steady-state population on a fitness landscape, based on the nature of the fitness landscape, thereby obviating the need for explicit population dynamics simulations and providing some insight into the impact on molecular evolution of the nature of fitness landscapes. Poor predictions are indicative of fitness landscapes that consist of a series of weakly connected sublandscapes.

Quantitative Structure-Selectivity Relationships (QSSR) are developed for a library of 40 phase-transfer asymmetric catalysts, based around quaternary ammonium salts, using Comparative Molecular Field Analysis (CoMFA) and closely related variants. Due to the flexibility of these catalysts, we use molecular dynamics (MD) with an implicit Generalized Born solvent model to explore their conformational space. Comparison with crystal data indicates that relevant conformations are obtained and that, furthermore, the correct biphenyl twist conformation is predicted, as illustrated by the superiority of the resulting model (leave-one-out q(2) = 0.78) compared to a random choice of low-energy conformations for each catalyst (average q(2) = 0.22). We extend this model by incorporating the MD trajectory directly into a 4D QSSR and by Boltzmann-weighting the contribution of selected minimized conformations, which we refer to as '3.5D' QSSR. The latter method improves on the predictive ability of the 3D QSSR (leave-one-out q(2) = 0.83), as confirmed by repeated training/test splits.

We present DESTRUCT, a new method of protein secondary structure prediction, which achieves a three-state accuracy (Q3) of 79.4% in a cross-validated trial on a nonredundant set of 513 proteins. An iterative set of cascade-correlation neural networks is used to predict both secondary structure and psi dihedral angles, with predicted values enhancing the subsequent iteration. Predictive accuracies of 80.7% and 81.7% are achieved on the CASP4 and CASP5 targets, respectively. Our approach is significantly more accurate than other contemporary methods, due to feedback and a novel combination of structural representations., (Copyright 2005 Wiley-Liss, Inc.)

A set of 538 inhibitors of the tyrosine kinase, Syk, including purines, pyrimidines, indoles, imidazoles, pyrazoles, and quinazolines, has been analyzed using a stepwise nonparametric regression (SNPR) algorithm, which has been developed for QSAR studies of pharmacological data. The algorithm couples stepwise descriptor selection with flexible, nonparametric, kernel regression, to generate structure-activity relationships. A further 371 molecules have been used as a test set to evaluate the models generated. Descriptors were selected using an internal monitoring set, and models were assessed using 10% of the principal (538-compound) data set, selected randomly, as an external validation set. The best model had a Q(2) of 0.46 for the external validation set. Test set predictions were significantly less accurate, partly due to the higher mean activity of the test molecules. However at a more coarse-grain level the SNPR models classified active molecules accurately, giving good enrichments. The data sets are difficult to model accurately and SNPR performs better than multilinear regression and a neural network analysis. In the additive implementation of SNPR multidimensional models are considered as a sum of single dimensional regressions. This makes the resultant models easily interpretable. For example, in the most predictive SNPR models, there is a clear nonlinear relationship between hydrophobicity (AlogP98) and inhibitory activity.

We report gas-phase electronic spectra of formamide, N-methyformamide, acetamide, and N-methylacetamide at 300 K calculated using a combination of classical molecular dynamics and time-dependent density functional theory (TDDFT). In comparison to excitation energies computed using the global minima structures, the valence npi* and pi(nb)pi* states show a significant red-shift of 0.1-0.35 eV, while smaller shifts are found for the n3s and pi(nb)3s Rydberg states. In this work, we have identified the physical origin of these shifts arising from variations of the molecular structure. We present simple relationships between key geometrical parameters and spectral shifts. Consequently, electronic spectra can be generated directly from ground-state structures, without additional quantum chemical calculations. The electronic spectrum of formamide in aqueous solution is computed using TDDFT using an explicit solvent model. This provides a quantitative determination of the condensed-phase spectrum. In general, this study shows that temperature effects can change the predicted excitation energies significantly and demonstrates how electronic spectra at elevated temperatures can be computed in a computationally efficient way.

Electronic transitions in aromatic side chains are responsible for the characteristics of proteins in the near UV. We present the first systematic study of a large number of proteins focused on the accurate calculation of near-UV circular dichroism (CD) spectra. We report new parameter sets derived from ab initio calculations for benzene, phenol, and indole that describe the valence electronic transitions to the (1)L(b), (1)L(a), (1)B(b), and (1)B(a) states in the side chains of amino acids phenylalanine, tyrosine, and tryptophan. CD spectra were calculated, using the matrix method with the new side-chain parameters, for 30 proteins whose CD spectra and crystal structures have been made publicly available. The new parameter sets are fully self-consistent and yield near-UV spectra better than those obtained using previous parameter sets. The mean absolute errors for computed wild-type spectra in the near UV are reduced by a factor of approximately 2. A similiar reduction is found for the near-UV spectra (and difference spectra) of mutants involving aromatic amino acids. Empirical modifications to model vibronic coupling in the side-chain chromophore of phenylalanine offer no overall improvement. Protein CD calculations from first principles coupled with atomic-level modeling enhance the utility and interpretability of CD measurements in the near UV.

Abrupt, smooth, and box methods for the calculation of electrostatic and steric field values in the comparative molecular field analysis (CoMFA) 3D QSAR technique are assessed on three diverse data sets of medicinal chemistry interest. While the standard CoMFA settings are robust to small changes in the position of the lattice, superior results may sometimes be obtained by use of only one field. However, if only the electrostatic field is used, then sometimes large differences between models are apparent. This appears to be due to a lack of column dropping, and these difficulties can be remedied by use of the box method.

A catalyst design methodology, utilizing combinatorial synthesis in parallel with chemometric analysis, is presented, which considers the 3D steric and electrostatic properties of substituents about a constant core structure.

Understanding the relationship between the amino acid sequence of a protein and its unique, compact 3D structure is one of the grand challenges in molecular biophysics. One particularly exciting approach is time-resolved electronic circular dichroism (CD) spectroscopy, which offers resolution on a nanosecond (or faster) time scale, although it does not provide the spatial resolution of techniques like X-ray crystallography or NMR. The thrust of our work is to underpin fast time scale spectroscopic studies of protein folding with a stronger theoretical foundation. Ultimately, we seek to use molecular dynamics simulations to study the influence of conformational dynamics and conformational transitions on the electronic CD spectra of proteins. We discuss how improved quantum chemical models of individual chromophores, including aromatic sidechains, can be incorporated into calculations of the electronic structure of proteins and their CD., (Copyright 2004 Wiley-Liss, Inc.)

The back-propagation neural network algorithm is a commonly used method for predicting the secondary structure of proteins. Whilst popular, this method can be slow to learn and here we compare it with an alternative: the cascade-correlation architecture. Using a constructive algorithm, cascade-correlation achieves predictive accuracies comparable to those obtained by back-propagation, in shorter time.

Helicases are ubiquitous enzymes involved in nucleic acid metabolism. The PcrA DNA helicase is an essential bacterial protein involved in rolling circle plasmid replication and DNA repair. Recent crystal structures of PcrA bound to DNA indicate that a flexible loop mediates a functionally important rigid-body-domain rotation. In this study, we report stochastic boundary molecular dynamics simulations focused on this region for wild-type and mutants designed to increase the rigidity of the region. Residues in the region that were helix-disfavoring, such as glycine, threonine, and others, were mutated to alanine. The simulated dynamics, analyzed with a variety of measures of structure and mobility, indicate that a few point mutations will substantially increase helix formation in this region. Subnanosecond stochastic boundary molecular dynamics simulations at several temperatures offer a rapid protocol for assessing large numbers of mutants and provides a novel strategy for the design of experiments to test the role of this flexible loop region in the function of PcrA., (Copyright 2003 Wiley-Liss, Inc.)

Recently, we have made significant improvements in the accuracy of calculations of the circular dichroism of proteins from first principles. The quality of these calculations (especially at 220 nm, a key wavelength, where the intensity of the band correlates well with the helical content of polypeptides) has given us confidence to use such calculations to analyse nanosecond molecular dynamics simulations of the folding of polypeptides. We use this combined approach to explore the influence of dynamics on the circular dichroism spectroscopy of polypeptides. We apply it to equilibrium molecular dynamics simulations of two beta-sheet proteins with similar structures, but differing circular dichroism spectra. We analyse a molecular dynamics simulation of the acid-unfolding of myoglobin. For both alpha-helical and beta-sheet conformations, we find that changes in dihedral angles of 30 degrees can change intensities of bands in circular dichroism spectra by up to 5000 degree cm2 dmol(-1). Thus, in isolation, moderate differences in circular dichroism spectra cannot be interpreted uniquely in terms of conformational changes. Examination of individual structures allows us to dissect the influence of conformation on the calculated circular dichroism spectra. Our results are aimed at providing a deeper understanding of the optical properties of proteins. An atomic level connection between molecular dynamics simulations and optical spectroscopy is increasingly desirable as theoretical and experimental studies begin to probe protein folding events reliably on the nanosecond timescale.

Here we investigate the structures and energetics of interactions between aromatic (Phe or Tyr) and basic (Lys or Arg) amino acids in alpha-helices. Side chain interaction energies are measured using helical peptides, by quantifying their helicities with circular dichroism at 222 nm and interpreting the results with Lifson-Roig-based helix/coil theory. A difficulty in working with Tyr is that the aromatic ring perturbs the CD spectrum, giving an incorrect helicity. We calculated the effect of Tyr on the CD at 222 nm by deriving the intensities of the bands directly from the electronic and magnetic transition dipole moments through the rotational strengths corresponding to each excited state of the polypeptide. This gives an improved value of the helix preference of Tyr (from 0.48 to 0.35) and a correction to the helicity for the peptides containing Tyr. We find that Phe-Lys, Lys-Phe, Phe-Arg, Arg-Phe, and Tyr-Lys are all stabilizing by -0.10 to -0.18 kcal.mol-1 when placed i, i + 4 on the surface of a helix in aqueous solution, despite the great difference in polarity between these residues. Interactions between these side chains have previously been attributed to cation-pi bonds. A survey of protein structures shows that they are in fact predominantly hydrophobic interactions between the CH2 groups of Lys or Arg and the aromatic rings.

Several non-parametric regressors have been applied to modelling quantitative structure-activity relationship (QSAR) data. Performances were benchmarked against multilinear regression and the nonlinear method of smoothing splines. Variable selection was explored through systematic combinations of different variables and combinations of principal components. For the training set examined--539 inhibitors of the tyrosine kinase, Syk--the best two-descriptor model had a 5-fold cross-validated q2 of 0.43. This was generated by a multi-variate Nadaraya-Watson kernel estimator. A subsequent, independent, test set of 371 similar chemical entities showed the model had some predictive power. Other approaches did not perform as well. A modest increase in predictive ability can be achieved with three descriptors, but the resulting model is less easy to visualise. We conclude that non-parametric regression offers a potentially powerful approach to identifying predictive, low-dimensional QSARs.

Several nonparametric regressors have been applied to modeling quantitative structure-activity relationship (QSAR) data. The simplest regressor, the Nadaraya-Watson, was assessed in a genuine multivariate setting. Other regressors, the local linear and the shifted Nadaraya-Watson, were implemented within additive models--a computationally more expedient approach, better suited for low-density designs. Performances were benchmarked against the nonlinear method of smoothing splines. A linear reference point was provided by multilinear regression (MLR). Variable selection was explored using systematic combinations of different variables and combinations of principal components. For the data set examined, 47 inhibitors of dopamine beta-hydroxylase, the additive nonparametric regressors have greater predictive accuracy (as measured by the mean absolute error of the predictions or the Pearson correlation in cross-validation trails) than MLR. The use of principal components did not improve the performance of the nonparametric regressors over use of the original descriptors, since the original descriptors are not strongly correlated. It remains to be seen if the nonparametric regressors can be successfully coupled with better variable selection and dimensionality reduction in the context of high-dimensional QSARs.

To study the distinct influences of structure and function on evolution, we propose a minimalist model for proteins with binding pockets, called functional model proteins, based on a shifted-HP model on a two-dimensional square lattice. These model proteins are not maximally compact and contain an empty lattice site surrounded by at least three nearest neighbors, thus providing a binding pocket. Functional model proteins possess a unique native state, cooperative folding and tolerance to mutation. Due to the explicit functionality in these models (by design), we have been able to explore their fitness or evolutionary landscapes, as characterized by the size and distribution of homologous families and by the complexity of the inter-relatedness of the functional model proteins. Mindful that these minimalist models are highly idealized and two-dimensional, functional model proteins should nevertheless provide a useful means for exploring the constraints of maintaining structure and function on the evolution of proteins.

We compare the low free energy structures of ten small, polar ligands in solution to their conformations in their respective receptor active sites. The solution conformations are generated by a systematic search and the free energies of representative structures are computed with a continuum solvation model. Based on the values of torsion angles, we find little similarity between low energy solution structures of small ligands and their active site conformations. However, in nine out of ten cases, the positions of 'anchor points' (key atoms responsible for tight binding) in the lowest energy solution structures are very similar to the positions of these atoms in the active site conformations. A metric that more closely captures the essentials of binding supports the basic premise underlying pharmacophore mapping, namely that active site conformations of small flexible ligands correspond to their low energy structures in solution. This work supports the efforts of building pharmacophore models based on the information present in solution structures of small isolated ligands.

Computer-aided drug design is a burgeoning discipline, in part, because of the fundamental role that molecular recognition plays in chemistry and biology, and partly due to the obvious commercial implications. This review focuses on methods rather than applications and a selection of novel techniques for predicting ligand binding energies that have recently emerged in the field, ranging from statistical analyses of highly reduced descriptions of ligands (with no information about the receptor) to statistical mechanical methods requiring atomic detailed knowledge of both ligand and receptor are presented.

A novel method for quantitative structure-activity relationship (QSAR) analysis is presented. The method, which does not assume any particular functional form for the QSAR, develops nonlinear relationships between parameters describing a set of molecules and the activity of the molecules. For the QSAR of the inhibition of Escherichia coli dihydrofolate reductase by 2,4-diamino-5-(substituted benzyl)pyrimidines, the method compares favorably to other nonlinear methods. Cross-validation trials demonstrate that the predictive ability is as accurate as other methods, and the method is simpler and faster than neural network and machine-learning methods. Consequently, its implementation is much easier, and interpretation of the generated QSAR is more straightforward.

The leucine zipper structure is adopted by one family of the coiled coil proteins. Leucine zippers have a characteristic leucine repeat: Leu-X6-Leu-X6-Leu-X6-Liu (where X may be any residue). However, many sequences have the leucine repeat, but do not adopt the leucine zipper structure (we shall refer to these as non-zippers). We have found and analyzed residue pair patterns that allow one to identify correctly 90% of leucine zippers and 97% of non-zippers. Simpler analyses, based on the frequency of occurrence of residues at certain positions, specify, at most, 65% of zippers and 80-90% of non-zippers. Both short and long patterns contribute to the successful discrimination of leucine zippers from non-zippers. A number of these patterns involve hydrophobic residues that would be placed on the solvent-exposed surface of the helix, were the sequence to adopt a leucine zipper structure. Thus, an analysis of protein sequences has allowed us to improve discrimination between leucine zippers and non-zippers, and has provided some further insight into the physical factors influencing the leucine zipper structure.

The apo form of myoglobin has two non-native stable states that have been experimentally characterized. Investigation of these states has suggested possible folding pathways for myoglobin. We have performed molecular dynamics simulations on solvated isolated helices of myoglobin to investigate the relationship between the intrinsic stabilities of the isolated helices and the structure and folding pathway of apomyoglobin. Analyses of hydrogen bonding and fluctuations from simulations at 298 and 368 K are used to explore the relative stabilities of the helices of myoglobin. The ordering observed is A approximately G approximately H > B > E > F, which mirrors both the experimental equilibrium and kinetic data available for apomyoglobin. The experimental observation that a subdomain comprising helices A, G, and H is an important early intermediate and our result that these helices are the most stable suggest that the intrinsically more stable helices form early in the folding process and that this significantly influences the folding pathway.

In protein unfolding studies, reduction in circular dichroism (CD) at 222 nm has been interpreted as loss of helicity. Estimates of the helicity of a protein from its CD spectrum are calibrated by reference to X-ray crystal structures, based on the assumption that the mean residue ellipticity at 222 nm is directly proportional to the number of residues in a helix. We have examined various influences on the CD at 222 nm, using molecular dynamics simulations to provide the structural detail required. We have found that the fragmentation of long helices, without a reduction in the number of helical residues, significantly reduces the mean residue ellipticity at 222 nm. The dynamical motion of the protein and the precise conformation of helical residues also play an important role. We discuss the implications of these factors to the interpretation of CD for the partial unfolding of apomyoglobin.

Neural networks and inductive logic programming (ILP) have been compared to linear regression for modelling the QSAR of the inhibition of E. coli dihydrofolate reductase (DHFR) by 2,4-diamino-5-(substituted benzyl)pyrimidines, and, in the subsequent paper [Hirst, J.D., King, R.D. and Sternberg, M.J.E. J. Comput.-Aided Mol. Design, 8 (1994) 421], the inhibition of rodent DHFR by 2,4-diamino-6,6-dimethyl-5-phenyl-dihydrotriazines. Cross-validation trials provide a statistically rigorous assessment of the predictive capabilities of the methods, with training and testing data selected randomly and all the methods developed using identical training data. For the ILP analysis, molecules are represented by attributes other than Hansch parameters. Neural networks and ILP perform better than linear regression using the attribute representation, but the difference is not statistically significant. The major benefit from the ILP analysis is the formulation of understandable rules relating the activity of the inhibitors to their chemical structure.

One of the largest available data sets for developing a quantitative structure-activity relationship (QSAR)--the inhibition of dihydrofolate reductase (DHFR) by 2,4-diamino-6,6-dimethyl-5-phenyl-dihydrotriazine derivatives--has been used for a sixfold cross-validation trial of neural networks, inductive logic programming (ILP) and linear regression. No statistically significant difference was found between the predictive capabilities of the methods. However, the representation of molecules by attributes, which is integral to the ILP approach, provides understandable rules about drug-receptor interactions.

The applications of artificial neural networks to the prediction of structural and functional features of protein and nucleic acid sequences are reviewed. A brief introduction to neural networks is given, including a discussion of learning algorithms and sequence encoding. The protein applications mostly involve the prediction of secondary and tertiary structure from sequence. The problems in nucleic acid analysis tackled by neural networks are the prediction of translation initiation sites in Escherichia coli, the recognition of splice junctions in human mRNA, and the prediction of promoter sites in E. coli. The performance of the approach is compared with other current statistical methods.

Neural networks have been applied to a number of protein structure problems. In some applications their success has not been substantiated by a comparison with the performance of a suitable alternative statistical method on the same data. In this paper, a two-layer feed-forward neural network has been trained to recognize ATP/GTP-binding [corrected] local sequence motifs. The neural network correctly classified 78% of the 349 sequences used. This was much better than a simple motif-searching program. A more sophisticated statistical method was developed, however, which performed marginally better (80% correct classification) than the neural network. The neural network and the statistical method performed similarly on sequences of varying degrees of homology. These results do not imply that neural networks, especially those with hidden layers, are not useful tools, but they do suggest that two-layer networks in particular should be carefully tested against other statistical methods.