PREDICTION OF ORAL DRUG BIOAVAILABILITY BASED ON CHEMICAL STRUCTURE

Abstract

Prediction of bioavailability is a crucial aspect of drug development and formulation design. Aim. To investigate the possibility of predicting the bioavailability of oral drug compounds based on molecular descriptors by means of feedforward neural network. Materials and methods. Software ChemOffice 2020 was used for calculation molecular descriptors from the drug structures. Software Matlab R2022b was used for building the multivariate regression between molecular descriptors and bioavailability as well as feedforward neural network. For each of the 145 drug molecules 16 molecular descriptors were calculated. Of the 145 compounds, 10 were randomly selected for use as a test subset, 5 compounds – as a validation subset. The remaining 130 compounds were used to train feedforward neural network. Results and discussion. It is established that there are 7 the most informativeness molecular descriptors for oral drug bioavailability prediction: molecular weight, hydrogen bond acceptors, hydrogen bond donors, partition coefficient, Balaban index, molecular topological index, Wiener index (multiple coefficient of determination is equal to 0,3701). Optimal number of hidden neurons for effective realization feedforward neural network was found by experimental way and its equal to 16. Trained feedforward neural network with 16 hidden neurons was used for predicting bioavailability values for test and validation subsets. Conclusions. Feedforward neural network is the effective tool for prediction of oral drug bioavailability based on chemical structure. This is evidenced by high values of determination coefficient between predicted bioavailability values and experimental bioavailability values for test and validation subsets (0,7084 and 0,8432, correspondingly). Obtained results can be useful at the stage of experiment planning or drug design.