Abstract: 【Aim】 Repellent can protect the users by driving target pests away from them. It is important to establish a nonlinear quantitative structureactivity relationship (QSAR) model with high precision and strong interpretation for designing and synthesizing the new insect repellent with higher bioactivity. 【Methods】 Based on the repellent activities of 37 aromatic carboxylic acid derivatives against the housefly, Musca domestica, the initial descriptors were generated with stoichiometry software PCLIENT, and then the binary matrix shuffling filter (BMSF) and worst descriptor elimination multi round method (WDEM) were successively used to conduct the nonlinear selection for initial descriptors. With the reserved descriptors, a support vector regression (SVR) model was established for the QSAR analysis of these 37 repellent derivatives. The influence of reserved descriptors on repellent activities was further analyzed with SVR interpretation system. 【Results】 The F-score of SVR model with original 1 542 descriptors was 1.2. However, it was 184.6 with the retained six descriptors after feature screening, indicating that feature screening has important effects on the precision of QSAR model. The importance of six molecular descriptors was as follows: p4BCD>GATS7v>T(O..O)>JGI8>SssO> nArCONR2. 【Conclusion】 The nonlinear relationship between reserved descriptors and the repellent activities of aromatic carboxylic acid derivatives against M. domestica was remarkable, and a high performance SVR-QSAR model for repellent derivatives was constructed.

Key words: Repellent, Musca domestica, aromatic carboxylic acid, repellency, nonlinear, quantitative structure activity relationship (QSAR), support vector regression (SVR)