Evaluation of boosted regression trees (BRTs) and two-step BRT procedures to model and predict blood-brain barrier passage

Abstract

Two new approaches, boosted regression trees (BRTs) and two-step BRT, were evaluated for modelling and predicting the blood-brain barrier (BBB) passage of drugs. Classification and regression trees (CART) were used as a base learner in BRT. In two-step BRT, a linear model (stepwise multiple linear regression (MLR) or partial least squares (PLS)) was built first, then BRT was applied to model the residuals of the linear model and both models were added. Both approaches were compared with the CART, MLR and PLS models. It was observed that BRT could improve the descriptive and predictive abilities compared to a single CART and that the stepwise MLR-BRT results in slightly improved descriptive and predictive properties compared to the MLR model. The combination of PLS and BRT did not result in an improvement, compared to the individual PLS model. The best models were obtained with stepwise MLR-BRT and PLS. It was shown that the combination of linear models with BRT is an approach that has potential and can be considered for future QSAR modelling.