QSPR MODELING OF THE SOLUBILITY OF DRUG-LIKE COMPOUNDS IN SUPERCRITICAL CARBON DIOXIDE USING INTELLIGENT COMPUTER TECHNIQUES.

Abstract

The solubility enhancement process of drugs in supercritical carbon dioxide is considered as one of the most powerful methods. However, the major obstructs of this technique can be resumed in the limit of using tedious and expensive experiments. The objective of the current study is to propose a new predictive model to correlate the solubility of drug-like compounds in supercritical carbon dioxide. A QSPR Model was first applied to a total number of 133 pharmaceutical compounds that have been collected, converted to the SMILES notation, and then their molecular descriptors were calculated using the software Padel. The obtained 1544 molecular descriptors have been reduced to 11 significant descriptors: AATS3v, MATS2e, GATS4c, GATS3v, GATS4e, GATS3s, nBondsM, AVP-0, SHBd, MLogP, and MLFER_S. Adding to these two independent intensive state properties (temperature and pressures), those 13 features are used as input variables forming the final dataset of 3590 experimental data points. In the form of a comparative study of four machine learning techniques: two ANNs, one DFO-SVR, and one ANFIS model their predictive/correlative capacity and feasibility have been examined. The models developed are all statistically valid with a global AARD between 0.73% and 2.46%, where the DFO-SVR model shows the best performance results and is the best-fit candidate for predicting future data and can be used to predict the solubility of pharmaceutical compounds, particularly for those that have not been tested as well as new compounds.