Macroscopic pK_a Prediction

Most molecules of practical interest have several places that can gain or lose a proton: in water, these molecules do not exist as a single structure but as an ensemble of protonation microstates. While experimental measurements and phenomena depend on macroscopic behavior—how the average population responds to changing pH—most pK_a-prediction solutions look only at a single microstate, resulting in much lower accuracy. Rowan's macro-pK_a prediction workflow is built to reproduce the complexity of full macroscopic ensembles, while using machine learning to accelerate these complex calculations and make them fast enough for routine work.

(For more on the difference between the macroscopic and microscopic pK_a paradigms, see our blog post.)

How Rowan's Macroscopic pKa Workflow Works

Rowan's macro-pK_a workflow begins by enumerating chemically reasonable protonation and tautomeric microstates within a user-specified charge window. Chemistry-aware rules avoid implausible structures, and a beam-search strategy prevents this from becoming exponentially slower for large molecules. A compact, physics-informed neural network ("Starling") then predicts a dimensionless free energy for each microstate.

Those energies generate a partition function that produces macroscopic pK_a values and, in the same pass, the population of every microstate as a function of pH. Because all quantities are derived from a single energy landscape, the output respects cycle constraints for coupled protonation and tautomerism—constraints that site-by-site methods often violate (as documented by Zheng and co-workers).

For a more detailed explanation of our methodology, see our publication.

Why It Matters

Where multiple sites or tautomers interact, the advantages of macroscopic pK_a prediction become obvious. Glycine is a simple example: the canonical "neutral" microstate is not the dominant species in water, and predicting pKa from this microstate yields microscopic pKa numbers that disagree with experiment and even swap the roles of the acidic and basic groups. In contrast, Rowan's holistic ensemble-based approach reproduces the observed macroscopic values correctly.

Piperazine, a symmetric diamine, offers a second illustration. A site-by-site approach tends to predict two identical pKa values near 9, while experiment shows one near 9 and another around 5. The macro-pK_a calculation recovers this separation: the first protonation alters the environment for the second and the model accounts for that shift automatically. (See the corresponding micro- and macro-pK_a calculations on Rowan.)

Where Macroscopic pK_a Prediction Can Be Used

Rowan's macroscopic pK_a-prediction tool can be used anywhere that pK_a prediction for complex molecules is needed: when predicting charge state before a molecular-dynamics run or FEP calculation, when assessing membrane permeability, or simply in trying to understand how different substitution might affect stability or intermolecular interactions. On a variety of popular pK_a-prediction benchmarks, Rowan's tool gives state-of-the-art accuracy, particularly for compounds in the biologically crucial 5–9 pK_a regime.

Additionally, many properties used in pharmacokinetics and formulation depend on both intrinsic molecular traits and pH-dependent speciation. With pH-dependent microstate populations in hand, Rowan's tool can generate logD–pH profiles by combining speciation with a logP model, clarifying why a scaffold is hydrophilic in the stomach but more lipophilic in blood. For brain penetration, it provides a principled state penalty: the free-energy cost of reaching a low-charge microstate near pH 7.4. Combining that penalty with a descriptor of neutral solvation yields a simple, effective classifier for blood–brain barrier triage. The same ensemble framework can inform salt selection, crystallization conditions, permeability modeling, and protein binding—any task where the balance of neutral and charged forms matters.

How to Run Macro-pK_a Calculations Through Rowan

Rowan's streamlined web-based interface makes running macroscopic-pK_a predictions simple. Input the molecule or molecules you want via any of our supported file formats, select the charge states you want to study, and submit the jobs—Rowan will automatically allocate computational resources, run the calculations, and return clear publication-quality results.

For high-throughput usage, calculations can also be run through Rowan's Python API. Rowan's API is free for all users and allows hundreds or thousands of macro-pK_a calculations to be submitted and analyzed programmatically. Here's what submitting a macro-pK_a calculation through Rowan's API looks like:

import rowan

rowan.api_key = "rowan-sk-your-key-here"

smiles = "c1ncncc1F"

workflow = rowan.submit_macropka_workflow(
    smiles,
    min_pH=0,
    max_pH=14,
    min_charge=-2,
    max_charge=2,
    name=f"Macropka Workflow for {smiles}",
)

workflow.wait_for_result()
workflow.fetch_latest(in_place=True)

print(workflow.data)
print(f"View the 3D structure on the web at labs.rowansci.com/macropka/{workflow.uuid}")

If you're interested in trying out Rowan's macrocopic pK_a predictions, subscribe to Rowan and start running calculations in minutes!