Modeling of lipophilicity by means of correlation weighting of local graph invariants

Results of using descriptors calculated with so-called correlation weights (CWs) of local graph invariants for modeling of lipophilicity are reported. Presence of different atoms (i.e. H, C, N, O, F, Cl and Br) in the molecular structure and the presence of different values of vertex degrees in labeled hydrogen-filled molecular graphs have been examined as local invariants. By means of Monte Carlo optimization procedure such values of the CWs which produce as large values as possible of correlation coefficient between lipophilicity and values of the descriptors on training set have been found. The model of lipophilicity obtained by this approach for structures of training set is a reasonable model for the prediction of the parameter values for compounds from test set.