Docking scores and quantum descriptors tested for blood-brain ratio prediction
- Preprint builds regression models of brain-to-blood distribution (logBB) for a curated set of several hundred compounds.
- It asks whether Vina docking scores against the efflux transporters P-gp and BCRP, plus quantum-chemical descriptors, add predictive information.
- Among many algorithms compared, a support vector regressor with an RBF kernel was chosen and gave moderate accuracy on a held-out set.
- The authors also weigh the computing cost of generating the descriptors.
💡 Points Based on the abstract
Moderate accuracy on one held-out set is the bar to beat, and the bar is a plain logP/TPSA/HBD model, which already gets you most of the way for BBB. I'd want to see the SVR with and without the Vina and QM columns on identical splits. If the lift is small, the QM compute isn't worth it on a library-scale screen. Abstract-only read, so check the paper.