QSRR for small pharmaceutical compounds in RPLC: A Machine learning approach

Reversed-Phase Liquid Chromatography(RPLC) is a common liquid chromatographic mode used for the control of pharmaceutical compounds during their drug life cycle. RPLC methods are used to characterize drug substances by separating e.g.the principal active ingredient from its impurities or matrix excipients. Nevertheless, determining the optimal chromatographic conditions that enable this separation is time-consuming and requires a lot of lab work. Quantitative Structure retention Relationship models(QSRR) are helpful for doing this job with minimal time and cost since they allow for predicting retention times of known samples without performing experiments.
In the current work, we have developed QSRR models and compared the strength of various machine learning algorithms to predict the retention times of small pharmaceutical compounds.The dataset comprised small pharmaceutical compounds covering a wide range in terms of physicochemical properties related to the selected retention mechanism. The regression models were based on a combination of linear and non-linear algorithms such as Multiple Linear regression(MLR), Support Vector Regression(SVR), Random Forest(RF), and Gradient Boosted Regression(GBR). Models were built for five pH conditions, i.e., at pH 2, 3.5 and 5 at a gradient time of 20 minutes (0-95 \% methanol) using a C18 T3 column. In the end, the model predictions were combined using stacking, and the performances of each model were compared. Here, QSRR models of the retention prediction have been built using structured derived physicochemical molecular descriptors, under consideration of the OECD principles in regulation for their acceptability and validation check during model construction and assessment. The KNN-based application domain filter was established to assess the reliability of the prediction for further compound prioritization. The best model was selected based on comparative values of $R^{2}$ and RMSE values on 10-fold cross-validation. Then the model was assessed on a holdout test data. Out of all models, the stacked model at pH5 outperformed with $R^{2}=0.92$ and RMSE= 2.02 minutes on 10-fold cross-validation and $R^{2} = 0.94$ and RMSE = 1.93 minutes on the test set.\\
This study can be insightful for analytical chemists working with RPLC platforms to begin with the computational prediction modeling like QSRR to improve predictive confidence of studies that will be based on separation of small molecules.We suggest that the current strategy can be successfully applied as a complementary tool to the existing ones for retention prediction of compounds even when a small dataset is available.