Macroscopic pKa Prediction

Python API · stjames models · API example

How it works

Predict macroscopic pKa values, protonation-state populations, logD, and an isoelectric point with Starling. Macroscopic pKa values describe changes in the whole molecule's protonation ensemble; they do not assign a pKa to a single atom. Rowan enumerates protonation microstates, predicts their free energies, and combines states with the same net charge to obtain macroscopic pKa values. See macroscopic versus microscopic pKa for the distinction.

Settings

Method: the web form offers "Starling (Uni-pKa architecture)."
Min pH / Max pH: set the range for population and property profiles; defaults are 0 and 14. Profiles use 0.1-pH steps and exclude the upper endpoint.
Min charge / Max charge: set the allowed net formal-charge range; defaults are −2 and +2. Widen this range if relevant protonation states fall outside it.
Compute solvation energy and BBB permeability?: add aqueous solvation energy and the predicted probability that the unbound brain-to-plasma ratio, Kp,uu, is at least 0.3. Off by default in the web form.
Predict pH-dependent aqueous solubility?: add Kingfisher solubility predictions, reported as log10 of molar concentration. On by default in the web form.

Notes

Enumeration uses chemical protonation rules and an AIMNet2-guided beam search. It retains promising states rather than every possible microstate, so an incomplete charge range or omitted state can affect predictions. Starling uses the Uni-pKa architecture to predict microstate free energies from molecular conformers.

Populations are calculated from microstate free energies and pH. LogD combines RDKit Wildman–Crippen logP estimates for all microstates, weighted by population in linear partition-coefficient space. Treat this as an approximate profile: ionic partitioning is not modeled explicitly. The isoelectric point is estimated where the average net charge reaches zero within the selected pH range; it may be absent.

Kingfisher predicts aqueous solubility referenced to pH 7.4. Rowan scales that prediction using the ratio of the neutral-state fraction at pH 7.4 to that at each requested pH. Predictions are omitted where the neutral fraction is negligible. This ideal-solution approximation does not model aggregation or salt precipitation, and errors in pKa can strongly affect the solubility profile.

The optional BBB prediction uses AIMNet2 conformers, GFN2-xTB/CPCM-X aqueous solvation energies, and a penalty for the neutral-state population at pH 7.4. The result is a classification probability, not a measured permeability or a predicted numerical Kp,uu value.

Benchmarks and validation

Accuracy

The table reports root-mean-square error (RMSE) in pKa units on the Novartis and SAMPL datasets; lower is better. Starling results come from our publication, while comparison values were collected from prior reports. These are published benchmark results, not a uniform rerun of every tool. Blank entries indicate unavailable results.