Rowan uses a physics-based workflow to estimate whether a given molecule will be able to penetrate the blood–brain barrier. The workflow follows precedent from Morgan Lawrenz and co-workers, who showed that pKa values and DFT-computed energy of solvation (Esol) could be used to estimate Kp,uu, the unbound brain to plasma ratio. Rowan's workflow uses the same high-level approach but employs neural network potentials and semiempirical solvent models to dramatically accelerate the calculations. (For more details, read the full paper.)
Rowan's blood–brain-barrier-permeability-prediction workflow begins by running a macroscopic pKa calculation using Starling. The output ensemble is used to calculate the "state penalty," or the free-energy penalty associated with not being in a chemically neutral, non-zwitterionic form at physiological pH. Specifically, we evaluate the Boltzmann-weighted population of microstates that are formally neutral and minimize atom-centered charges; when no fully neutral form exists, we select the microstate with the fewest formal atom-centered charges. The state penalty is then defined as the free energy required to shift the ionization equilibrium toward this reference state:
The remaining change in energy is computed by comparing the electronic energies of conformer ensembles with and without solvent. Conformer ensembles are generated using ETKDG as implemented in RDKit, and conformers were optimized and scored using AIMNet2 and the CPCM-X water model as applicable.
The final Esol values is calculated as the difference between the energy of the water-solvated conformer ensemble and the energy of the gas-phase conformer ensemble minus the aforementioned state penalty.
Log-transformed Kp,uu values were binarized at 0.3 to define a "brain-penetrant" class and logistic-regression models were trained to predict these values from Esol.
On the reported Schrodinger Kp,uu values, Rowan's workflow achieved good predictive accuracy (AUC=0.85 and 75% accuracy), demonstrating that low-cost calculations can yield highly predictive workflows. For more detailed analysis, refer to the full paper.