Prediction of HPLC conditions using QSPR techniques: an effective tool to improve combinatorial library design

The purification and characterization of compounds resulting from parallel synthesis or combinatorial chemistry has not yet been optimized to operate as a completely automated high-throughput process. Liquid chromatography/mass spectroscopy (LC/MS) is most commonly employed to carry out the characterization and identification of combinatorial compounds. This desired level of automation can only be accomplished if the separation conditions for every compound in the combinatorial array are known prior to the analysis. This study presents a quantitative structure retention relationship (QSRR) approach to predict the retention time of structurally diverse solutes under 75 different LC/MS conditions. Sixty-two compounds were analyzed using 15 commonly used HPLC columns under 5 different gradient conditions. The solute retention time was used as the dependent variable, and more than 1000 molecular descriptors were calculated for this compound set to generate QSRR models. After the elimination of highly correlated variables and those with zero variance, two different genetic algorithms were applied to identify the most significant descriptors. Following the variable selection, the identified descriptors were used to create QSRR models for each separation condition. The calculated stepwise multiple linear regression models have been proven to be statistically significant and highly predictive, with an average coefficient of determination (R2) of 0.86, an average cross-validated r2 of 0.62, r2 = 0.76, and an average F value of 27.29. The QSRR models can be used to design "analysis-friendly" library purification plates, in which compounds are arranged on the basis of their predicted separation condition and can also be used during the library design phase to flag compounds not amenable to the separation methods in use.