LogP Prediction

The octanol–water partition coefficient, logP, measures how a neutral compound distributes between 1-octanol and water at equilibrium:

Larger values indicate greater lipophilicity. Because logP describes the neutral compound, it is distinct from pH-dependent distribution coefficients such as logD.

Rowan provides three methods that trade off speed, domain dependence, and physical detail. We recommend:

• Chemprop Sangster2026 (chemprop_sangster2026) for a reasonable balance of throughput and accuracy,
• Crippen (crippen) for an immediate, reproducible 2D baseline or very large screens, and
• COSMO-RS (cosmors) when conformational and solvent physics are important enough to justify the additional cost, especially for chemistry that may be outside the Chemprop model's training domain.

All three methods accept one molecular structure and return a scalar logP prediction. An example API script is available on GitHub.

Chemprop Sangster2026

Chemprop Sangster2026 is the default method. It uses a Chemprop v2 directed-message-passing neural network trained on the SangsterLogP dataset, which contains approximately 23,000 unique experimental logP values. It is fast enough for screening and generally offers the best balance of speed and predictive accuracy for drug-like molecules represented in its training domain.

Crippen

The Crippen method evaluates the RDKit implementation of the Wildman–Crippen atom-fragment model directly from molecular connectivity. It is effectively instantaneous and useful for rapid baselines or very large libraries, but it does not explicitly model three-dimensional conformations or solvent thermodynamics.

COSMO-RS

COSMO-RS is a physics-based, conformer-aware method. It is much slower because it requires conformer sampling and quantum-chemical COSMO surface calculations, but it can be useful for molecules outside the domain of a learned model.

The calculation proceeds as follows:

1. An initial conformer ensemble is generated with openconf.
2. Conformers are optimized with GFN2-xTB/ALPB in water and scored with g-xTB/CPCM-X in water.
3. Solvent-dependent clustering selects conformers that represent the ensemble across hexane, octanol, chloroform, DMSO, and water.
4. A COSMO surface is calculated once for each retained conformer using the workflow's BP86/def2-TZVPD surface settings.
5. COSMO-RS calculates water and octanol solvation free energies from the same surface.

For conformer $i$, the partition coefficient is obtained from the difference between its solvation free energies:

Rowan combines the conformer partition coefficients using water-phase Boltzmann weights at 298.15 K.