I would like to perform a virtual drug screening based on a QSAR model with aim to achieve a set of ligands which can penetrate in BBB. In your opinion which descriptor are key factors for this?
See QSAR Model of Unbound Brain-to-Plasma Partition Coefficient, Kp,uu,brain: Incorporating P-glycoprotein Efflux as a Variable, Dolgikh et al., J. Chem. Inf. Model., 2016, 56 (11), pp 2225–2233.
There are 2 sides to this interesting topic: rate and extent of brain penetration. There are many predictive models for the former, but fewer for the latter. Margareta Hammarlund-Udenaes's group has done a lot of interesting work in both measuring and predicting Kp,uu. I've attached a poster from 2014 presentation that may be of use.
In normal QSAR you would assume that you have a single binding site. Here you actually have a combination of 1) a lipid barrier , 2) efflux transporters (P-gp and BCRP, as well as MRP's) , often with multiple binding sites. On top of this, permeation may not necessarily predict free brain parenchyma concentrations (due to protein binding), as eluded to by Thomas & Alex. If you just want to predict the permeation step, I suppose you ideally could use the Lipinski predictors (MW, oxygen bond donors & acceptors, log P) and combine with QSARs for the binding sites for the efflux transporters, these are however not well characterized.