engleski

Removal of Correlated Descriptors Does not Necessarily Result in Better Prediction: QSAR Study of Citostatic Activity of Selected Hydroxyurea Derivatives

Purpose of the study: Hydroxyurea is citostatic agent used in treatment of leukemia and other malignant tumors, as well sickle cell anemia and virus infections. Thus amino acid derivates of hydroxyureas were synthesized.1 The ratio between corresponding therapeutic and toxic effect was used as a dependant variable for QSAR analysis. This data set was used to asses how reduction of dimensionality of well correlated, redundant descriptors (Pearson coefficient of correlation greater than 0.9) influences the prediction based on random forest approach. Methods: 39 descriptors were generated using MOPAC 7 and ACD/ChemSketch freeware v10.02 on a PC computer. Correlation analysis was done in Statistica v7.1 while modeling based on random forests and leave-one-out external validation have been made within R environment v2.8.1. Results obtained: Significant inter-correlation was established for 11 descriptors. Descriptors that were highly correlated with more than one parameter and correlated descriptors which interpretation is ambiguous were removed from the complete set f descriptors. Final set of descriptors was reduced to 28 elements. Due to this reduction prediction accuracy was decreased from 82% to 81%, but what is more important scattering of results was increased from 5% to 9% for non reduced and reduced dimensionality, respectively. Conclusion: Removal of highly-correlated descriptors decreased the quality of prediction and thus lowered the value of QSAR model of citostatic activity of amino acid derivates of hydroxyurea. Reference: 1. Opačić N, Barbarić M, Zorc B, Cetina M, Nagl A, Frković D, Kralj M, Pavelić K, Balzarini J, Andrei G, Snoeck R, De Clercq E, Raić-Malić S, Mintas M: The novel L- and D-amino acid derivatives of hydroxyurea and hydantoins: synthesis, X-ray crystal structure study, and cytostatic and antiviral activity evaluations. J Med Chem, 2005 ; 48(2):475-82.

correlation; prediction; QSAR

nije evidentirano