Descriptors of lipophilicity are all distribution coefficients between a buffer (at different pH values) and a lipid-like phase, such as octanol, reversed phase (RP) highperformance liquid chromatography (HPLC) column stationary phase, lipid membranes, and so forth, as stated in Section 26.3. The most prominent member of this distribution coefficient family is the coefficient between octanol and water. For the neutral form of the desired molecule, this coefficient is called the partition coefficient (Kow). More commonly, its log P value is used. Also available are logarithms for the distribution between octanol and buffer at various pH values, log DpH (distribution coefficient), or log K, which is the logarithm of the capacity factor from RP-HPLC measurements. Some recently developed methods are distribution coefficients that have stationary phases on RP-HPLC columns built up by covalently bound single lipid molecules - log IAM (immobilized artificial membranes) [74] - or those using liposomes [75], or complete lipid bilayers immobilized on silica beads - log MA (membrane affinity) [76]. Only for log P and log D do commercially available calculation methods exist.

In general, one can divide the prediction methods for log P (log D) into three groups:

1 Atom contribution methods (A log P) [77]. With this approach, methods are gathered that fragment a molecule into its heavy atoms. Once the contribution of an atom fragment (defined before analysis) to the molecule's log P value is evaluated by a correlation analysis with measured log P values, the log P of a new structure is calculated by summing up the contribution of all atoms. In some cases, the contributions are somehow weighted when an atom is an exposed peripheral one compared with more shielded interior atoms.

2 Molecular property methods (B log P) [77]. These methods use structural features, such as number of nitrogen and oxygen atoms or ring structures available, or semiempirical molecular descriptors derived from quantum chemical calculations or solvatochromic parameters, for example hydrogen-bonding acceptor/ donor ability, and correlate these to known log P values of experimental datasets of varying size.

3 Fragmental contribution methods (C log P) [77]. These methods are the most established in the literature and are based on the contribution of molecular frag ments (substituents). Additional constants or correlation factors may be employed depending on the interconnectivity of the fragments.

All of these methods have their own advantages and disadvantages, e.g. the fragmental approaches give higher correlation coefficients than the atomic contribution methods but sometimes in the fragmental approaches a parameterization is not possible because a fragment is missing from the training set. A benchmark can be found in the literature [77].

It is our belief that a method using computer-automated structure evaluation [78], which is now commercially available [79], is a special case. This method encodes the structural fragments into binary bitstrings, known as fingerprints. Two ways of fingerprinting have been described:

1 a structural key approach with predefined functional groups, and

2 an approach of hashing of unique structural paths.

The latter has the advantage that the fingerprints generated are characterized by the nature of the chemical structures in the experimental dataset. Various conditions can be selected to determine the uniqueness of a fragment. The fragment length can be predefined or can be kept flexible in a length region, for example between 1 and 5 atom units, and the program itself finds the most relevant fragments by a multilinear regression analysis with measured properties. Alternatively, neuronal networks can be trained with these fingerprints as input patterns. These methods are very useful if a large dataset of measured properties, such as log P, log K, log D, and log MA values, is available. The advantage is that for distribution coefficients such as log K and log MA, where neither commercial programs nor public databases are available, a calculation prior to synthesis is possible even for very large libraries.

Was this article helpful?

0 0

Post a comment