MODELLING AND PREDICTION OF LIPOPHILICITY FOR NATURAL COMPOUNDS WITH STRONG BIOLOGICAL ACTIVITY

Abstract

The goal of this study was to develop high statistical significant models for lipophilicity estimation for a group of 60 compounds with increased toxicity, belonging to alkaloids and mycotoxins. The multiple linear regression modelling was made by means of genetic algorithms as a function of 972 molecular descriptors, computed by ChemOffice and Dragon Plus software and completed by internet available module for Log P computation, ALOGPS 2.1. The compounds classification has been realized using principal component analysis and hierarchical cluster analysis. Data evaluation has been realized by various correlation matrices and relevant graphs. The modelling was made on the basis of 26 compounds with known log P_exp values and the results were validated by means of additional models developed for a series of 20 compounds. The other 6 compounds, which were excluded from the modelling process, were used afterwards as test set for prediction and comparison. The most descriptive models were those retaining four descriptors, and in all models were selected at least one computationally expressed log P value (miLogP, KOWWIN and ALOGP most often). All the obtained results are highly suggestive and offer a very pertinent idea regarding the lipophilicity range of natural compounds of increased toxicity. The models were validated considering various statistical parameters and different correlation matrices defined by high statistical significance.