Thermodynamics-Informed Accurate pKa Prediction and Protonation State Generation in PlayMolecule AI

Abstract

Accurate prediction of acid dissociation constants (p$K_{\rm a}$) and the determination of dominant protonation states is critical in drug discovery, influencing molecular properties such as solubility, permeability, and protein-ligand binding. We present Acep$K_{\rm a}$, an advanced application integrated into the PlayMolecule AI platform. Acep$K_{\rm a}$ is built upon the theoretically rigorous Uni-p$K_{\rm a}$ framework, which unifies statistical mechanics with representation learning. By modeling the complete protonation ensemble rather than treating p$K_a$ as a scalar regression target, Acep$K_{\rm a}$ ensures thermodynamic consistency across coupled ionization sites. We describe the application's enhanced architecture, which features a retrained Uni-Mol backbone achieving state-of-the-art performance on standard benchmarks. Furthermore, we detail critical engineering advancements. These include AceConfgen, a proprietary GPU-accelerated conformer generator that achieves a ~40x speed-up compared to NVIDIA's nvmolkit, a streamlined inference engine to directly protonate molecules, and a 3D-aware modality for applying protonation states to bound ligand poses. Finally, we discuss the integration of Acep$K_{\rm a}$ into the PlayMolecule AI ecosystem, a modern AI-assisted environment for molecular modelling and drug discovery.

Figures (5)

• Figure 1: Benchmark of models on public p$K_{\rm a}$ datasets. All data except for AcepKa are taken from luo2024bridging and wagen2025physics.
• Figure 2: Acep$K_{\rm a}$ app workflow. (1) App takes in input one or multiple SMILEs strings (or 3D conformers) and a pH value. (2) Then, it generates the protonation ensemble for each of the provided SMILES. (3) A conformational ensemble for each of the microstates is generated using nvmokit. (4) The 3D conformations are fed to the Uni-Mol model that predicts their free energies. (4) Free-energies are used in Eq. \ref{['eq:boltz_weights']} to derive the relative pH-dependent population of each microstate.
• Figure 3: Output example from Acep$K_{\rm a}$ in single-molecule mode using histamine as query molecule. The figure shows the enumerated microstates on the left and their pH-dependent relative population on the right.
• Figure 4: Comparison of AceConfgen and nvMolKit performance on the Platinum Benchmark. (a) Distributions of minimum RMSD values. (b) Total runtime required to complete the benchmark.
• Figure 5: PlayMolecule AI GUI and Acep$K_{\rm a}$ usage orchestrated by the LLM agent, showing the prediction of the protonated state of a ligand sitting in a protein pocket. (a) The agent is asked to fetch a structure with PDB ID 3MJ2. (b) The agent is asked to predict the protonation state of the ligand present in the loaded structure at a target pH, and to visualize the results.