Small-molecule pKa predictor that scores the free energy of every protonation microstate, so macro-pKa and pH populations stay consistent.
No providers recorded yet. Browse all providers
Glycine carries two ionizable groups, and at neutral pH it exists mostly as a zwitterion. Whether a predictor gets that right depends on four micro-pKa values that are tied together by a thermodynamic cycle: pK1 + pK2 must equal pK3 + pK4. A model that regresses each site's pKa independently has no reason to respect that identity, and when it does not, the same population ratio computed along two routes can disagree by orders of magnitude. Experimental data carry the mirror-image problem: most measurements are macro-pKa values that reflect several coupled equilibria at once, yet datasets usually assign each one to a single presumed site.
Uni-pKa, from DP Technology with collaborators at Renmin University of China and MIT, answers both problems by changing what the network predicts. Instead of a pKa per site, it predicts the standard Gibbs free energy of every protonation microstate a molecule can adopt. Macro-pKa values, micro-pKa values and pH-dependent populations are then derived analytically from Boltzmann-weighted sums over those free energies, so thermodynamic consistency is a property of the arithmetic rather than something the model has to learn. The work first appeared on chemRxiv in August 2023 and was published in JACS Au in 2024 under the title "Bridging Machine Learning and Thermodynamics for Accurate pKa Prediction".
The encoder is a modified Uni-Mol, DP Technology's 3D molecular transformer. Uni-pKa's protocol was later retrained independently by Acellera as AcepKa.
Pretraining uses about 1 million ChEMBL molecules with computationally predicted pKa values, which enumerate to more than 3 million microstate structures. It combines a weakly supervised pKa task, where the loss compares the log ratio of the acid and base partition functions to the label, with three self-supervised tasks: masked atom prediction (15% of atoms), masked charge prediction and 3D coordinate recovery. Fine-tuning uses the Dwar-iBonD set, built from DataWarrior and selected i-BonD entries, which grows from 8,232 data points to 27,138 microstates after enumeration; conformer augmentation is applied, and the released workflow averages a 5-fold ensemble. The default model uses formal charges, with Gasteiger and GFN2-xTB partial charges evaluated as alternatives.
On external macro-pKa benchmarks the paper reports MAE/RMSE of 0.810/1.061 on Novartis acids, 0.493/0.653 on Novartis bases, 0.554/0.716 on SAMPL6, 0.570/0.735 on SAMPL7 and 0.631/0.878 on SAMPL8, against 0.835/1.150 for MolGpKa on SAMPL8. On 10 NMR-measured SAMPL6 micro-pKas it reaches an MAE of 0.592. On 27 tertiary amines it records an MAE of 0.52 against 0.69 for Schrödinger Jaguar without conformational search and 0.51 with a comprehensive one, while predicting about 28 macro-pKas per second. Tautomer equilibria are outside its training scope, and on Tautobase it reaches an MAE of 3.435 log K units.
Uni-pKa is aimed at the ligand-preparation step of computational drug discovery: choosing the protonation state and formal charge to carry into docking, free-energy perturbation or molecular dynamics. Because it returns populations across the pH range, a chemist can see when two species coexist near physiological pH. Its pKa estimates also feed solubility, permeability and distribution-coefficient reasoning in lead optimization. DP Technology serves it as a hosted protonation-state ranking app on its Bohrium platform.
Uni-pKa showed that a physical constraint can be built into a pretrained molecular representation model at the output layer rather than hoped for in the training data, and it made the protonation-ensemble view of pKa data concrete by releasing reconstructed datasets in that format. The framework has been reused directly: Acellera's AcepKa retrains the same Uni-Mol backbone under the Uni-pKa protocol. The training and inference code and the enumerator are released under Apache-2.0, and the datasets are on AIS Square under Apache-2.0; the GitHub repository contains no trained checkpoints, and its README directs users to a Bohrium notebook for the model weights.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.