A new redundant variable pruning approach—minor latent variable perturbation–PLS used for QSAR studies on anti-HIV drugs

Abstract

A new approach for eliminating the redundant variables in the multivariable data matrix encountered in QSAR studies, minor latent variable perturbation (MLVP)-PLS method has been proposed. In the latent variable (LV) space, the minor latent variables (LVs) with small covariances are mainly formulated by linear combinations of the redundant variables including information-deficient and highly correlative ones, while the major LVs with large covariances are mainly contributed by the informative variables. Deleting a minor LV, which is equivalent to a perturbation for LV space, could make the redundant variables not well be represented in LV subspace, leading to strong variation of their PLS regression coefficients. The informative variables could still be normally represented in LV subspace with the PLS regression coefficients remaining relatively stable. MLVP-PLS utilizes this fact to discriminate the informative and redundant variables. It gradually identifies and eliminates the redundant variables according to the relative variation of PLS regression coefficients after perturbations are given. The elimination process is terminated according to some proposed criteria. Applying the method to the quantitative structure–activity relationship (QSAR) studies on TIBO derivatives as potential anti-HIV drugs has demonstrated the feasibility and robustness of the proposed approach. A deeper insight into the effect of different structural parameters on the bio-activity of TIBO derivatives has been reached.