Type (I, II) errors variable selection in quantitative structure activity relationships

• For the first time type (I) and type (II) errors based variable selection is proposed in this study.
• Chance correlation and sign change are considered as criteria for selecting descriptors.
• Valid selected descriptors resulted in valid models.
• Properness of pre-selection procedure depends on the nature of data set.
• Ordered variable selection resulted in almost smallest set of informative variables.

Selection of valid and informative descriptors is one of the most crucial steps in the QSAR studies. In this contribution, type (I, II) errors variable selection is proposed as a statistically based variable selection method for selecting valid descriptors. Developed selection strategy can be considered as a sophisticated combination of jackknifing and y-randomization. Type (I) error is a measure of descriptors' importance, and it is provided using jackknifing for each descriptor. Type (II) error is a measure of chance correlation and it is calculated using y-randomization. Successive projection algorithm and Gram–Schmidt orthogonalization were utilized as pre-selection methods for the initial reduction of collinear and uninformative variables. Selwood data including 31 molecules and 53 descriptors, anti-HIV data including 107 molecules and 160, and Flour data including 116 molecules and 1268 descriptors were utilized in this research. Model parameters for each data set before and after variable selection confirm the adequacy of the proposed novel variable selection.