[Application of Kohonen Self-Organizing Feature Maps in QSAR of human ADMET and kinase data sets]. / Kohonen-féle önszervezodo tulajdonságtérkép használata humán ADMET és kináz adatok QSAR predikciójában.

QSAR predictions have been proven very useful in a large number of studies for drug design, such as kinase inhibitor design as targets for cancer therapy, however the overall predictability often remains unsatisfactory. To improve predictability of ADMET features and kinase inhibitory data, we present a new method using Kohonen's Self-Organizing Feature Map (SOFM) to cluster molecules based on explanatory variables (X) and separate dissimilar ones. We calculated SOFM clusters for a large number of molecules with human ADMET and kinase inhibitory data, and we showed that chemically similar molecules were in the same SOFM cluster, and within such clusters the QSAR models had significantly better predictability. We used also target variables (Y, e.g. ADMET) jointly with X variables to create a novel type of clustering. With our method, cells of loosely coupled XY data could be identified and separated into different model building sets.

Comparison of predictive ability of water solubility QSPR models generated by MLR, PLS and ANN methods.

ADME/Tox computational screening is one of the most hot topics of modern drug research. About one half of the potential drug candidates fail because of poor ADME/Tox properties. Since the experimental determination of water solubility is time-consuming also, reliable computational predictions are needed for the pre-selection of acceptable "drug-like" compounds from diverse combinatorial libraries. Recently many successful attempts were made for predicting water solubility of compounds. A comprehensive review of previously developed water solubility calculation methods is presented here, followed by the description of the solubility prediction method designed and used in our laboratory. We have selected carefully 1381 compounds from scientific publications in a unified database and used this dataset in the calculations. The externally validated models were based on calculated descriptors only. The aim of model optimization was to improve repeated evaluations statistics of the predictions and effective descriptor scoring functions were used to facilitate quick generation of multiple linear regression analysis (MLR), partial least squares method (PLS) and artificial neural network (ANN) models with optimal predicting ability. Standard error of prediction of the best model generated with ANN (with 39-7-1 network structure) was 0.72 in logS units while the cross validated squared correlation coefficient (Q(2)) was better than 0.85. These values give a good chance for successful pre-selection of screening compounds from virtual libraries, based on the predicted water solubility.

Lipophilicity of aminopyridazinone regioisomers.

Ten pairs of pyridazinone regioisomers were prepared, and their lipophilicity was described by the logarithm of the octanol/water partition coefficient (log P) determined experimentally and calculated with prediction methods. The 4- and 5-(substituted amino)-3(2H)-pyridazinone regioisomers were synthesized by nucleophilic substitution of one of the chloro atoms of 4,5-dichloro-2-methyl-3(2H)-pyridazinone or its 6-nitro derivative. Structures of new compounds were proven by spectroscopic methods. The experimental log P values were obtained by a shake flask method in octanol and a Sörensen buffer (pH 7.4) solvent system. A consequent difference was found in the lipophilicity of regioisomers. For each isomer pair, the log P value of the 4-isomer was significantly (average by 0.75 log unit) higher than that of the 5-isomer. Some quantum chemical calculations as well as X-ray analysis of two pairs of regioisomers were also carried out to gain insight into the structural differences of regioisomers. The log P values were calculated by the fragmental approach KOWWIN and a QSPR analysis (3DNET). The a priori KOWWIN gave poor agreement, but with the programs KOWWIN with EVA (experimental value adjusted) and 3DNET, the results were generally in agreement with experiment.

Validation subset selections for extrapolation oriented QSPAR models.

One of the most important features of QSPAR models is their predictive ability. The predictive ability of QSPAR models should be checked by external validation. In this work we examined three different types of external validation set selection methods for their usefulness in in-silico screening. The usefulness of the selection methods was studied in such a way that: 1) We generated thousands of QSPR models and stored them in 'model banks'. 2) We selected a final top model from the model banks based on three different validation set selection methods. 3) We predicted large data sets, which we called 'chemical universe sets', and calculated the corresponding SEPs. The models were generated from small fractions of the available water solubility data during a GA Variable Subset Selection procedure. The external validation sets were constructed by random selections, uniformly distributed selections or by perimeter-oriented selections. We found that the best performing models on the perimeter-oriented external validation sets usually gave the best validation results when the remaining part of the available data was overwhelmingly large, i.e., when the model had to make a lot of extrapolations. We also compared the top final models obtained from external validation set selection methods in three independent and different sizes of 'chemical universe sets'.