AcepKa

At pH 7.4 histamine is not one molecule. A dication, a monocation and a neutral form coexist, with the monocation accounting for 95.7% of the population. A docking run or simulation setup that silently commits to one drawn structure is guessing which member of that mixture matters, and getting it wrong changes the ligand's formal charge, its hydrogen-bond donors and its electrostatic complementarity with the pocket.

Most machine-learning pKa predictors treat pKa as a site-wise regression target: each ionizable atom gets its own predicted number. When a molecule is polyprotic those numbers are coupled through inductive and resonance effects, and independently regressed constants can violate the thermodynamic cycles that relate them. The Uni-pKa framework from DP Technology reframed the problem — enumerate the complete protonation ensemble, predict a standard Gibbs free energy for every microstate in it, and derive macro-pKa values and pH-dependent populations analytically from the resulting partition function. Consistency stops being something the model learns and becomes a property of the arithmetic.

AcepKa is Acellera's implementation of that framework, described by Francesco Pesce, Stephen Farr and Gianni De Fabritiis in 2026 and deployed inside the PlayMolecule AI platform. The motivation is distribution rather than method: the Uni-pKa reference code is open source, but the trained ensemble behind its published benchmarks is not, so reproducing that accuracy required curated data, GPUs and expertise. Acellera retrained the Uni-Mol backbone independently under the Uni-pKa protocol and put the result behind a web interface that needs no local installation.

#Key Features

• Ensemble output rather than a scalar: single-molecule mode returns macro-pKa values, the relative population of every enumerated microstate across the pH range, and the dominant species at the queried pH.
• Thermodynamic consistency by construction: because the network predicts free energies and the pKa values are derived from Boltzmann weights, predictions for coupled ionization sites cannot contradict each other.
• 3D bound-pose modality: an SDF ligand from a crystal structure or docking pose comes back with protonation and hydrogen placement applied in place, preserving the original coordinates. The state is predicted from the ligand alone — the protein environment is not an input, and pocket-induced pKa shifts are not modeled.
• GPU-native conformer generation: AceConfgen reimplements RDKit's distance-geometry embedding and MMFF94 minimization in fused FP32 kernels, removing what is otherwise the dominant cost of ensemble-based prediction.
• Library mode: the pipeline chains enumeration, batched conformer generation and batched free-energy evaluation end to end, returning the dominant protonated state for every molecule submitted.

#Technical Details

Three components sit in series. A rule-based enumerator applies SMARTS templates to identify ionizable sites and combinatorially builds all valid microstates with net charge between −2 and +2. A Uni-Mol encoder — an SE(3)-invariant transformer over atom types and coordinates — predicts the standard Gibbs free energy of each microstate. An FE2pKa module turns those free energies into the reported values. Retraining followed the Uni-pKa protocol: a weakly supervised pKa task over roughly 1 million ChEMBL molecules with empirical values, expanded to more than 3 million unique protonation states by enumeration, alongside masked atom prediction, masked charge prediction and 3D coordinate recovery; fine-tuning then used DataWarrior and curated i-BonD entries to produce a 5-fold ensemble. On macro-pKa benchmarks the RMSE is 0.97 (Novartis acid), 0.77 (Novartis base), 0.90 (SAMPL6), 0.70 (SAMPL7) and 0.91 (SAMPL8), against 1.06, 0.65, 0.72, 0.73 and 0.88 for the original Uni-pKa ensemble and 1.14, 1.15, 1.25, 0.71 and 1.51 for ChemAxon Marvin. On Platinum 2017 — 50 conformers for each of 4,548 protein-bound ligands, 227,400 in total — AceConfgen finished in 2 min 24 s on a single RTX 4090 against 17 min 40 s for nvMolKit and 28 min 18 s for 16-core RDKit, at matching RMSD-to-reference distributions.

#Applications

The immediate use is ligand preparation. Structure-based campaigns need a defensible protonation state before docking, binding free-energy calculation or simulation, and the 3D modality supplies one without re-embedding a pose that was expensive to obtain. Library mode extends the same treatment to screening sets. The population curves also give medicinal chemists something a single number cannot: when a predicted pKa sits near 7.4, both species are present in appreciable amounts and both may be worth modeling.

#Impact

AcepKa's contribution is availability rather than a new mechanism, and the paper is explicit that close agreement with Uni-pKa is the expected outcome of a faithful retraining. That framing is the point: a thermodynamically rigorous method that existed only as code plus undistributed weights becomes something a chemist can run on a library from a browser. The cost is openness — neither the retrained weights nor the AcepKa and AceConfgen source are released, and access runs through the PlayMolecule web application with a free account and no programmatic API. The work is a preprint awaiting peer review.