self-adaptive genetic algorithm-artificial neural network algorithm with leave-one-out cross validation for descriptor selection in QSAR study

• Published in: Journal of computational chemistry Vol. 31; no. 10; pp. 1956 - 1968
• Main Authors: Wu, Jingheng, Mei, Juan, Wen, Sixiang, Liao, Siyan, Chen, Jincan, Shen, Yong
• Format: Journal Article
• Language: English
• Published: Hoboken Wiley Subscription Services, Inc., A Wiley Company 30.07.2010; Wiley Subscription Services, Inc
• Subjects: Algorithms; artificial neural network; Chemistry; Genetic algorithms; leave-multiple-out; leave-one-out; Linear Models; Models, Biological; multiple linear regression; Neural networks; Neural Networks (Computer); Quantitative Structure-Activity Relationship; Validation studies; Validity; Y-randomization
• ISSN: 0192-8651; 1096-987X; 1096-987X
• DOI: 10.1002/jcc.21471

Based on the quantitative structure-activity relationships (QSARs) models developed by artificial neural networks (ANNs), genetic algorithm (GA) was used in the variable-selection approach with molecule descriptors and helped to improve the back-propagation training algorithm as well. The cross validation techniques of leave-one-out investigated the validity of the generated ANN model and preferable variable combinations derived in the GAs. A self-adaptive GA-ANN model was successfully established by using a new estimate function for avoiding over-fitting phenomenon in ANN training. Compared with the variables selected in two recent QSAR studies that were based on stepwise multiple linear regression (MLR) models, the variables selected in self-adaptive GA-ANN model are superior in constructing ANN model, as they revealed a higher cross validation (CV) coefficient (Q²) and a lower root mean square deviation both in the established model and biological activity prediction. The introduced methods for validation, including leave-multiple-out, Y-randomization, and external validation, proved the superiority of the established GA-ANN models over MLR models in both stability and predictive power. Self-adaptive GA-ANN showed us a prospect of improving QSAR model.