| Resumo: | One of the most important factors, affecting the pharmacokinetic profile of a drug is binding to plasma protein. As such, this study aimed at the development of a quantitative structure–activity relationship model, to predict the fraction unbound in plasma (fub) for four species, using artificial neural network ensemble (ANNE). To this end a database of 363 drugs was used, and molecular descriptors were determined. The dataset was divided in two groups, a train and an external validation, to avoid overfitting. The ANNE optimization reduced the descriptors required to determine the fub to 37, and 150 ANN were randomly selected, trained and the optimal configuration was collected. The different ANNE were built by averaging the output values of the selected ANN and the best ANNE was selected. The model created was able to predict, with a small amount of error, the fub values (root mean square error of 0.16798 and 0.193705 for train and test dataset respectively), however, it tends to underestimate this value (mean error of -0.00291 and -0.015780 for train and test dataset respectively). The ANNE interpretation showed that the main characteristics of that affect fub were the molecule charge, size, structure and lipophilic and hydrophilic affinity.
|
|---|