Partition Coefficient - Prediction

Prediction

QSPR (Quantitative Structure-Property Relationship) algorithms calculate a log P in several different ways:

• Atomic based prediction (atomic contribution; AlogP, XlogP, MlogP, etc.)

A conventional method for predicting log(P) is to parameterize the contributions of various atoms to the over-all molecular partition coefficient, which produces a parametric model. This parametric model can be estimated using constrained least-squares estimation, using a training set of compounds with experimentally measured partition coefficients. In order to get reasonable correlations, the most common elements contained in drugs (hydrogen, carbon, oxygen, sulfur, nitrogen, and halogens) are divided into several different atom types depending on the environment of the atom within the molecule. While this method is generally the least accurate, the advantage is that it is the most general, being able to provide at least a rough estimate for a wide variety of molecules.

• Fragment based prediction (group contribution; ClogP, etc.)

It has been shown that the log P of a compound can be determined by the sum of its non-overlapping molecular fragments (defined as one or more atoms covalently bound to each other within the molecule). Fragmentary log P values have been determined in a statistical method analogous to the atomic methods (least squares fitting to a training set). In addition, Hammett type corrections are included to account of electronic and steric effects. This method in general gives better results than atomic based methods, but cannot be used to predict partition coefficients for molecules containing unusual functional groups for which the method has not yet been parameterized (most likely because of the lack of experimental data for molecules containing such functional groups).

• Data mining prediction

A typical data mining based prediction uses support vector machines, decision trees, or neural networks. This method is usually very successful for calculating log P values when used with compounds that have similar chemical structures and known log P values.

• Molecule mining prediction

Molecule mining approaches apply a similarity matrix based prediction or an automatic fragmentation scheme into molecular substructures. Furthermore there exist also approaches using maximum common subgraph searches or molecule kernels.

• Estimation of log D (at a given pH) from log P and pK_a:
  • exact expressions:
  • approximations for when the compound is largely ionized:
  • approximation when the compound is largely un-ionized:

• Prediction of pK_a

  For prediction of pK_a, which in turn can be used to estimate log D, Hammett type equations have frequently been applied. See for a recent review of newer methods.