Computational QSAR models with high-dimensional descriptor selection improve antitumor activity design of ARC-111 analogues

Abstract

ARC-111 has potent topoisomerase I-targeting activity and pronounced antitumor activity. To design ARC-111 analogues with improved efficiency, we performed analyses on the quantitative structure–activity relationship of 22 ARC-111 analogues assessed in P388 tumor cells. First, the support vector regression (SVR) models were constructed and optimized based on literature descriptors (the low-dimensional descriptor space) and the worst descriptor elimination multi-round (WDEM) method. The optimized SVR model had greater generalization ability than multiple linear regression (MLR) and stepwise linear regression (SLR) in the independence test, which indicated that our nonlinear WDEM method could remove redundant descriptors more effectively, and our optimized SVR was a more powerful modeling technique. Second, to identify more accessible and effective descriptors, our modeling descriptors with clear meanings were selected from a large number of descriptors calculated by the software PCLIENT. Through the high-dimensional descriptor selection nonlinear method and the WDEM method, seven independent variable combinations with tens of descriptors were selected out of 2,923 descriptors. The seven corresponding SVR models performed better in the independent test, compared to MLR and SLR. The evaluation measures supported the excellent predictive power of the new models. According to the interpretability analysis of the SVR model, the regression significance of the model and the importance of single indicator were evaluated based on F tests. Our study offers some useful theories for understanding the function mechanism and finds parameters for designing ARC-111 analogues with enhanced antitumor activity.