Accurate pKa predictions matter more than most other molecular descriptors. They determine protonation state at physiological pH, which affects solubility, membrane permeability, protein binding, and metabolic clearance. Getting pKa wrong propagates through every downstream model. In 2026, you have three main options for online pKa prediction: ChemAxon's Marvin, SwissADME, and Egret-1. This post compares them fairly so you can pick the right tool.
The contenders
ChemAxon Marvin
The industry standard empirical pKa predictor for decades. Marvin uses a substructure-based model fitted to a large experimental pKa dataset. It is fast, it is available through a friendly web interface and a REST API, and it is very accurate on molecules similar to its training set.
SwissADME
A free academic web service from the Swiss Institute of Bioinformatics. Pastes in a SMILES, gives you back pKa, logP, TPSA, and other drug-like descriptors. Uses a mix of published empirical methods. Not designed for high-throughput API use, but an excellent starting point for academic projects and students.
Egret-1
A neural network potential specialized for small-molecule organic chemistry. Computes pKa from the underlying energy model rather than from fragment matching. Grounded in quantum chemistry, which helps with novel scaffolds. For a deeper introduction, see our Egret-1 overview.
Accuracy
On standard drug-like chemistry benchmarks, the three tools are closer than you might expect. Typical mean absolute errors are:
- Marvin: 0.4 to 0.7 pKa units, with excellent performance on chemistries well represented in its training set.
- SwissADME: 0.5 to 0.9 pKa units, depending on which of its underlying models dominates for your chemistry.
- Egret-1: 0.4 to 0.6 pKa units on drug-like molecules, with more robust generalization to novel scaffolds because it learns from quantum chemistry rather than fragment matching.
For reference, experimental pKa values for the same compound measured by different techniques scatter by 0.3 to 0.5 units. All three tools are therefore often within the noise floor of the underlying experiments.
Coverage
Coverage is where empirical and neural tools start to diverge:
- Marvin is excellent for common drug-like scaffolds and less reliable for natural products, macrocycles, and organometallic fragments.
- SwissADME has similar coverage to Marvin for small organic molecules and does not claim to handle unusual chemistry.
- Egret-1 has the broadest scaffold coverage in the organic space because it learns from quantum energies rather than substructure patterns. It still fails on transition metals and very large systems.
Speed
All three tools are fast at the per-molecule level:
- Marvin: tens of milliseconds per molecule through the REST API.
- SwissADME: a few seconds per molecule through the web interface, not designed for batch API calls.
- Egret-1: sub-second per molecule through the SciRouter API, with batching for large screens.
For a library of 10,000 molecules, Marvin and Egret-1 are both in the under-an-hour range. SwissADME is not the right tool for that scale.
Licensing and cost
This is often the decisive factor:
- Marvin: free for academic use, commercial license required for industry. Pricing varies by seat count and feature scope.
- SwissADME: free for everyone, with the usual citation request for academic work. No API.
- Egret-1: available through SciRouter on a usage-based plan with a free tier for early exploration, and the underlying model is open source.
Which one to pick
Academic research
SwissADME is free and sufficient for most student projects and exploratory work. If you have a ChemAxon academic license, Marvin is the gold standard. Egret-1 is useful as a sanity check, especially for novel scaffolds.
Industrial medicinal chemistry
Compare Marvin and Egret-1 on your specific chemistry. Where they agree, trust the answer. Where they disagree, that is where wet-lab measurements are most valuable. Egret-1 is a good choice for chemistries that are new to your organization.
High-throughput screening
Either Marvin via API or Egret-1 via SciRouter. SwissADME is not appropriate at this scale. For very large libraries (hundreds of thousands of molecules), Egret-1 tends to win on speed per dollar.
What this comparison does not tell you
One thing worth calling out: pKa accuracy numbers from published benchmarks reflect the benchmark molecules, not your molecules. The single biggest thing you can do to pick a tool is to run all three on a handful of compounds you already have experimental pKa values for. Whichever tool is most accurate on your chemistry is the right choice for your project, regardless of what general benchmarks say.
Bottom line
In 2026, you do not have to pick a single pKa predictor and live with it. Academic users start with SwissADME. Industry users default to Marvin for in-distribution chemistry and add Egret-1 for novel scaffolds. High-throughput users go straight to Egret-1 or the Marvin API. All three are good tools, and the right one for your project depends on the specific molecules you care about.