engleski

A prototype structure-activity relationship model based on National Cancer Institute cell line screening data

BACKGROUND AND PURPOSE: The National Cancer Institute maintains a database of over 45000 chemicals screened for in vitro activity on many human cancer cell lines. Such abundance of data might be exploited to computationally predict the utility of antitumor drug candidates, starting from a numerical description of molecular structures. MATERIALS AND METHODS: Each of the 10000 examined chemicals is described with 450 topological descriptors in attempt to capture the composition of the molecule and the way its atoms are interconnected. The descriptors are compressed using PCA down to 50 components, retaining 95% of the original information. Self-organizing maps were built (using in-house I2SOM software) and support vector machines trained (using Weka software) to demonstrate and quantify the utility of predicting anticancer activity of chemicals. RESULTS: The self-organizing maps demonstrated a clear improvement over negative controls in visualizations of compound activity on two selected cell lines, although the patterns were not clear enough to reliably perform clustering. An analogous conclusion can be drawn for the self-organizing map trained to show cellular mechanism of action. An SVM has proven moderately able to perform both regression (activity level) and classification (active vs. inactive) tasks on the selected chemicals. CONCLUSIONS: We have shown that the in vitro antitumor activity of a compound can be predicted, although higher accuracy would be desired in most situations. The characteristics of the trained SVM show that performance of the models could be greatly improved by expanding the dataset with additional descriptors and additional chemicals.

antitumor drugs; topological descriptors; self-organizing map; support vector machine

nije evidentirano