A recursive-partitioning model for blood-brain barrier permeation

A series of bagged recursive partitioning models for log(BB) is presented. Using a LGO-CV, three sets of physical property descriptors are evaluated and found to have Q(2) values of 0.51 (CPSA), 0.53 (Ro5x) and 0.53 (MOE). Extrapolating these models to Pfizer chemical space is difficult due to P-glycoprotein (P-gp) mediated efflux. Low correlation coefficients for this test set are improved (R-2 stop=0.39) when compounds known to be P-gp substrates or statistical extrapolations are removed. The use of simple linear models for specific chemical series is also found to improve the correlation over a limited chemical space.