Back to News & Insights

Taking Charge: Making Protonation-State Prediction Accessible with pKasso

November 17, 2026

Bind

Bind Research Blog

pKasso is an open-source tool that helps researchers predict how small molecules gain and lose protons as pH changes. Existing machine-learning models can predict protonation behaviour, but turning their predictions into practically usable protonation states can involve substantial effort, particularly for non-experts.

Developed at Bind Research in collaboration with University College London, pKasso brings established machine-learning models into an accessible workflow. It also introduces new methods to predict complete protonation-state distributions and improve prediction accuracy, supporting downstream applications such as virtual screening and molecular simulation in drug discovery. 


Small changes in protonation can change molecular behaviour 

A molecule’s chemical structure tells only part of its story. In solution, many small molecules, including most conventional medicines, can gain or lose protons depending on the pH. These changes affect their charge and hydrogen-bond pattern, which can influence properties such as solubility, membrane permeability, and interactions with proteins.

Figure 1: Protonation states can affect the interactions between small molecules and biomolecules including intrinsically disordered proteins. 

For scientists modelling a potential drug, getting the protonation state right matters. A docking calculation or molecular simulation uses a particular molecular representation as its starting point. If the protonation state is inappropriate for the conditions being studied, the calculation may give a misleading picture of how that molecule behaves. 

The challenge becomes more complicated when a molecule has several sites that can gain or lose protons. There may be multiple populated forms at the same pH, including forms with the same overall charge but protons attached at different positions. Scientists therefore need to consider an ensemble of protonation states, rather than simply assign one charge to a molecule. 


From pKa predictions to usable molecular inputs

Experimental measurements provide valuable information about protonation equilibria, often expressed as pKa values. These describe where the balance between protonated and deprotonated forms shifts as pH changes. 

Computational prediction helps researchers explore these properties across much larger collections of molecules. However, turning pKa predictions into the molecular forms needed for downstream calculations can involve several additional steps. Researchers must identify ionisable sites, generate candidate protonation states, and estimate which states are likely to be populated under their chosen conditions. Installing models and connecting these steps can introduce substantial workflow overhead. Commercial software offers convenient alternatives, but licensing costs can limit access. 

We built pKasso to make this process easier and faster. 


pKasso brings protonation-state prediction into one workflow 

pKasso takes a molecular structure, identifies its ionisable sites, and generates possible protonation states. It then predicts how likely each state is to occur at a given pH or across a pH range. 

pKasso builds on existing machine-learning models in a flexible framework, processing and adjusting their predictions where needed, and turns them into practically useable distributions of protonation states. The initial release includes two prediction models, MolGpKa (Pan et al., J. Chem. Inf. Model., 2021) and Uni-pKa (Luo et al., JACS Au, 2024), and additional models can be added in the future. 

Users can run either predictor or combine their predictions in a mixed model. The tool also reports disagreement between models as an indicator of uncertainty, helping researchers identify cases that warrant closer inspection.

Figure 2: A simplified pKasso workflow showing molecular input, candidate protonation states (molecular charges in parentheses), and their predicted populations across pH. Shaded areas in the plot indicate the uncertainty between prediction models.

The results include molecular structures, protonation state populations across different pH values, and macroscopic pKa estimates. This allows researchers to see how the predicted protonation states change as the solution becomes more acidic or basic, which can be compared against experimental data. 


Competitive accuracy, with options for different workloads

We evaluated pKasso against experimental pKa values across six benchmark datasets. Its predictions achieved excellent accuracy competitive with published predictors. The benchmarks primarily assess macroscopic pKa accuracy. Selected molecular examples also illustrate protonation-site assignments, but quantitative validation of complete microstate populations remains an area for further work. 

Speed also matters when preparing large compound collections. pKasso speeds up the calculation by adaptively partitioning different protonation sites in a larger molecule, allowing researchers to determine protonation states for molecules with many sites. As a result, many predictions with the fast MolGpKa-based workflow are completed in fractions of seconds. Geometry-based predictions with the Uni-pKa-based workflow are even more accurate but take longer to complete, giving users a choice according to their available resources and workload.


Bringing pKasso into research workflows

For a quick look at one molecule, researchers can use the web interface to explore predicted protonation states and download structures. For larger projects, the command-line interface and Python library support local use and batch processing. 

A medicinal chemist might use the web tool to inspect how a compound’s predicted charge changes across pH. A computational chemist might incorporate pKasso into compound preparation, generating candidate protonation states before docking, molecular simulation, or downstream AI pipelines.

Figure 3: An example workflow linking compound input, pKasso predictions of protonation state populations, and downstream modelling.

Predictions with pKasso describe the molecule in solution; they do not explicitly account for changes caused by binding to a protein. Researchers should also inspect cases where the two built-in models disagree strongly. 

For teams in the pharmaceutical industry, pKasso offers a way to integrate protonation-state prediction into existing computational workflows. For academic researchers, its open-source code and accessible interfaces reduce the practical barriers to using and extending these methods. 

At Bind Research, we want useful computational methods to be available to the scientists who need them. pKasso supports that aim by combining established prediction models, openly available code and model weights, and a straightforward route from a molecular structure to predicted protonation states.


Tell us how you use pKasso in your research: bindresearch.org/contact

Learn more about pKasso in our preprint.

Use the tool: tools.bindresearch.org/pkasso

Read the source code and installation instructions: github.com/bindresearch/pkasso

More News & Insights