Predicting Intestinal Permeation of Drugs Through Neural Network Analysis Based on Five Molecular Descriptors

Abstract

The oral route is generally preferred for drug administration because of its ease and good patient compliance. In the search for new drugs for oral administration a major problem encountered is obtaining drug structures which, as well as being potent in viiro, possess favourable pharmacokinetic profiles which enable them to pass easily through the relevant body membranes, especially the gastrointestinal epithelia to effect their action. However, many of the compounds derived from combinatorial synthesis and high throughput screening have inappropriate properties for oral absorption, such as low solubility and low permeability, so that rate of success in drug development for delivery through this route is low. Mastlo of MacroModel (of Schriidinger Inc.) was used to build the molecules of 74 drugs. No partial charges were given to any atom a priori. This was to enable the true charges resulting from their mutual interactions with one another, and the solvent atoms to be introduced later, to manifest duringlafler modeling. At the end of thisstage, therefore, only one structure existed for each drug. Thereafter, the structures were all initially minimized with a conjugate gradient method and then with one of ho hewton Matrix methods. These generated much more stable forms of the structures built by minimizing their internal energies. The minimizations were all done at body temperature (310 K) and in aqueous medium with a dielectric constant of 78. Exhaustive conformer searches were done employing two main algorithms, the Monte-Carlo (MC) search, the MClStochastic Dynamics and mixed-mode MCISD. The latter two were only employed for those molecules unable to converge with the MC search, needing more extensive treatment. The resulting low-energy conformers were hen sorted with Schrbdinger's Xcluster. Depending on which was more appropriate, from the resulting graphs obtained, torsional angles or root mean square vii (rms) displacements of atoms were used as the basis for screening the conformers of low energy, for each drug, into naturally occurring classes from which the lowest energy conformer in each class was selected as representative of the class of very similar conformers. This was necessary due to the hundreds of low-energy conformers generated for most drugs. Clustering analysis reduced them to less than 20 (less than 1.2 for most) representative low energy conformers. These formed the matter of further investigations. Using MOPAC2002, the dipole moment (DP), polarizability (Pol) and -sum charges of each conformer (the representative) were determined. A Java program extracted all the charges of each atom of each conformer, the confomer DP. energy and Pol from the varied output files into a Microsoft Excel file. Here, a Visual Basic program was used to synchronize these dispersed values into Osum charges, N-sum charges, H (attached to 0 & N) -sum charges, DP and Pol for each drug. The Boltzmann distribution coefficient of each conformer was generated and utilized in averaging their values to give a representative value for each drug with respect to each descriptor. This skewed the average (as obtains in reality) towards vdues of lower energy conformers. A neural network model (NNM) was generated with Qlyuda Forecaster XL, permitting it to auto-optimize the resulting NNM. A multiple linear regression analytic (MLRA) model was generated with Excel. It was done ;IS a basis of comparison (based on R~ and error coefficients) with the non-linear NNM An analysis of variance (ANOVA) study on the significance of the MLRA model silowed it to be significant (PC 0.05). The NNM was found to be better than the multiple linear regression analytic models. This was borne out by the R2 values - 0 7185 and 0.4192 - and predictive Root Mean Square Error values (RMSE) - 0.4481 and 0 6387 - for the NNM and MLRA model respectively. The NNM results were viii also found to be superior to previous studies with Polar Surface Area (PSA) as the sole molecular descriptor and PSA with molecular weight as dual descriptors – RMSE 0.622 and 0.606 respectively. From the NNM, the relative importance of each of the descriptors in determining the permeability was H: 48.369%, 0: 23.683%, Dipole: 13.741%, N: 7.268% and Polarizability: 6.939%; the MLRA enabled a general trend lo be deducible after statistical/mathematical transformations (positive gradients: 0 and N, negative gradients: H, Dipole and Polarizability). Polynomial trend lines of the untransformed data (mvestigated up to the sixth order) supported the MLRA result.