Quantitative structure-activity relationship modeling of mosquito repellents using calculated descriptors

Quantitative structure-activity relationship (QSAR) modeling of 40 DEET-related mosquito repellents was carried out using calculated molecular descriptors. When the four different classes of descriptors (topochemical, topostructural, geometrical and quantum chemical) were used in a hierarchical fashion, topochemical descriptors were found to explain much of the variance in the data and this indicated the importance of the chemical nature of the atoms and groups towards repellency of these compounds. Ridge regression (RR) outperformed partial least square regression (PLS) and principal component regression (PCR). We also used descriptor thinning via a modified Gram-Schmidt algorithm prior to ridge regression. This resulted in a four-parameter model with a 20-fold cross-validated R² of 0.734. The q² (R ²_cv) reported here is the "true-q²" because the descriptor thinning was embedded within the cross-validation step. Inclusion of any calculated physicochemical property (secondary descriptor) did not result in improvement of the models built with the calculated molecular descriptors (primary descriptors). This result has great implications in the QSAR assisted design of new mosquito repellents because calculation of the primary descriptors does not require any input other than the molecular structure.