Predictive Free Wilson analysis

A large combined data set of 48 propafenones was then analyzed by both Free-Wilson analysis and a combined Hansch/Free-Wilson approach using an artificial neural network (ANN). With this approach it was possible, in contrast to conventional MLR analysis, to correctly predict the MDR-reversing activity of 34 compounds of the data set after the ANN was trained by only 14 compounds. Best results were obtained using those descriptors showing the highest statistical significance in MLR analysis [150]. 

This represents 3X3X3X6X6X3 = 2916 possible compounds. A good Free-Wilson analysis was obtained from only 42 of the possible analogs the preparation of these 42 compounds enables one to predict with a fair assurance the approximate antimalarial activity to be expected for almost 2900 unprepared analogs. 

Although the predictive power of a model is considered to be a criterion for the relevance of QSAR models, the main purpose of Hansch analysis and related approaches such as Free-Wilson analysis concerns not prediction, but a better understanding of the chemical problem. 

Only a small number of new analogs can be predicted from a Free Wilson analysis (14 other compounds in the case of eq. 70). 

The problems of this data set are easily understood if a Free-Wilson analysis is applied. " The training set compounds ( 1-21) can be described by a simple one-parameter regression equation (equation 1 the term 4,5-C=C- indicates the presence or absence of a cycloaliphatic 4,5-double bond in ring A of the steroids). The internal predictivity of this model (Q = 0.726 spress = 0.630) and the test set predictivity (n = 10 = 0.477 spress = 0.733) are even slightly... 

Hernandez-Gallegos, Z., Lehmann, P. A. (1990) A Free-Wilson/Fujita-Ban analysis and prediction of the analgesic potency of some 3-hydroxy- and 3-methoxy-M-alkylmorphinan-6-one opioids. J Med Chem 33(10), 2813-2817. 

Sciabola, S., Stanton, R. V., Wittkopp, S., et al. (2008) Predicting kinase selectivity profiles using Free-Wilson QSAR analysis. J Chem Inform Model 48(9), 1851-1867. 

The basic principles on which the Hansch multiple parameter method for structure-activity correlation depends are described. These are compared with the basic features of the Free-Wilson method for assigning additivity constants to structural features of related compounds. An example is given for which the two methods of analysis have led to similar structure-activity relationships. Factors which determine the particular method to use in a new situation are discussed. The Free-Wilson method is presented in considerable operational detail with special emphasis on the detection and avoidance of situations which lead to singularity problems in solution of the matrix. Favorable analyses, which result in additivity constants that can be correlated with known physical constants, may lead to predictions for new compounds not covered in the original matrix. 

The main shortcomings of the Free-Wilson approaches are that 1) structural variation is necessary in at least two different sites 2) a relatively large number of variables is necessary to describe a relatively small number of compounds 3) the models can be used to predict a maximum number of compounds equal to - n, where n is the number of compounds effectively used in the model 4) predictions of substituents not included in the analysis are usually not reliable. 

One of the disadvantages of the Free-Wilson method is that— unlike regression equations based on physicochemical parameters— it cannot be used to make predictions for substituents not included in the original analysis. The technique may break down when there are linear dependencies between the structural descriptors, for example, when two substituents at two positions always occur together, or where interactions between substituents occur. Advantages of the technique include its ability to handle data sets with a small number of substituents at a large number of positions, a situation not well handled by other analytical methods, and its ability to describe quite unusual substituents since it does not require substituent constant data. A number of variations and improvements have... 

---