Predictive Array Design. A method for sampling combinatorial chemistry library space.

A method, Predictive Array Design, is presented for sampling combinatorial chemistry space and selecting a subarray for synthesis based on the experimental design method of Latin Squares. The method is appropriate for libraries with three sites of variation. Libraries with four sites of variation can be designed using the Graeco-Latin Square. Simulated annealing is used to optimise the physicochemical property profile of the sub-array. The sub-array can be used to make predictions of the activity of compounds in the all combinations array if we assume each monomer has a relatively constant contribution to activity and that the activity of a compound is composed of the sum of the activities of its constitutive monomers.

Classification of toxin-induced changes in 1H NMR spectra of urine using an artificial neural network.

NMR spectra of urine from rats treated with a range of liver, kidney and testicular toxins at various doses were measured and classified using neural network methods. Toxin-induced changes in the levels of 18 low molecular weight endogenous urinary metabolites were assessed using a simple semi-quantitative scoring system. These scores were used as input to an artificial neural network, the use of which has been explored as a means of predicting the class of toxin. With this limited data set, based only the level of the maximal changes of these 18 metabolites, the network was able to predict the class and hence target organ of the toxins. Renal cortical toxicity was well predicted as was liver toxicity. The few examples of renal medullary toxins in the data set resulted in relatively poor training of the network although correct classification was still possible.

Structure-activity relationships of antifilarial antimycin analogues: a multivariate pattern recognition study.

The structure-activity relationships of a series of novel antifilarial antimycin A1 analogues have been investigated by using computational chemistry and multivariate statistical techniques. The physiochemical descriptors calculated in this way contained information which was useful in the classification of compounds according to their in vitro antifilarial activity. This approach generated a 53 parameter descriptor set, which was reduced with a multivariate pattern recognition package, ARTHUR. Regression analysis of the reduced set yielded several statistically significant regression equations; e.g.-log in vitro activity = 0.017 mp + 0.65 log P - 0.81ESDL10-7.33 (R = 0.9). With use of this equation, it was possible to make predictions for further untested analogues. The analysis indicated that membrane or lipid solubility is an important determinant in biological activity agreeing with the proposed primary mode of action of the compounds as disrupters of cuticular glucose uptake.