Boosting in block variable subspaces: An approach of additive modeling for structure–activity relationship

• BBVS combines PLS with a type of gradient boosting in a stepwise way.
• Boosting in block subspaces can extract information more effectively.
• BBVS resists overfitting and it is easy to select the times of boosting iterations.

Quantitative structure activity relationships (QSAR) and quantitative structure property relationships (QSPR) are established by a novel approach of additive modeling: boosting in block variable subspaces (BBVS). Different families of 2D and/or 3D molecular descriptors explain the molecular structure from different points of view. Hence, descriptors from different families could be regarded as variables in different variable subspaces. We define these subspaces as block variable subspaces. Boosting in these subspaces can extract information more effectively and hence build a model of high quality. BBVS combines partial least squares (PLS) regression with a type of gradient boosting in a stepwise way. It is capable of resisting overfitting, making it easier to select the number of boosting iterations than to select the number of components of PLS.