Search

In order to find more effective anticancer drugs, the U.S. National Cancer Institute (NCI) screens a large number of compounds in vitro against 60 human cancer cell lines from different organs of origin. About 70,000 compounds have been tested in the program since 1990, and each tested compound can be characterized by a vector (i.e., "fingerprint") of 60 anticancer activity, or -[log(GI50)], values. GI50 is the concentration required to inhibit cell growth by 50% compared with untreated controls. Although cell growth inhibitory activity for a single cell line is not very informative, activity patterns across the 60 cell lines can provide incisive information on the mechanisms of action of screened compounds and also on molecular targets and modulators of activity within the cancer cells. Various statistical and artificial intelligence methods, including principal component analysis, hierarchical cluster analysis, stepwise linear regression, multidimensional scaling, neural network modeling, and genetic function approximation, among others, can be used to analyze this large activity database. Mining the database can provide useful information: (a) for the development of anticancer drugs; (b) for a better understanding of the molecular pharmacology of cancer; and (c) for improvement of the drug discovery process.

The U.S. National Cancer Institute (NCI) conducts a drug discovery program in which approximately 10,000 compounds are screened every year in vitro against a panel of 60 human cancer cell lines from different organs of origin. Since 1990, approximately 63,000 compounds have been tested, and their patterns of activity profiled. Recently, we analyzed the antitumor activity patterns of 112 ellipticine analogues using a hierarchical clustering algorithm. Dramatic coherence between molecular structures and activity patterns was observed qualitatively from the cluster tree. In the present study, we further investigate the quantitative structure-activity relationships (QSAR) of these compounds, in particular with respect to the influence of p53-status and the CNS cell selectivity of the activity patterns. Independent variables (i.e., chemical structural descriptors of the ellipticine analogues) were calculated from the Cerius2 molecular modeling package. Important structural descriptors, including partial atomic charges on the ellipticine ring-forming atoms, were identified by the recently developed genetic function approximation (GFA) method. For our data set, the GFA method gave better correlation and cross-validation results (R2 and CVR2 were usually approximately 0.3 higher) than did classical stepwise linear regression. A procedure for improving the performance of GFA is proposed, and the relative advantages and disadvantages of using GFA for QSAR studies are discussed.