Applying machine learning on high-performance thin-layer chromatography using the complementary developing solvents concept

Predicting chromatographic results is a difficult task for many analysts, especially in Thin-Layer Chromatography (TLC) where reproducibility is always a critical point. The availability of suitable equipment and rigorous standardization of parameters has transformed TLC into High-Performance Thin-Layer Chromatography (HPTLC) and made reproducibility of results a reality. With recent non-targeted screening methods using the concept of complementary developing solvents, HPTLC has become a medium to high throughput technique that generates large sets of data, allowing the construction of predictive models. In this study, we evaluated to which extend HPTLC R_F are decoded from molecular chemical properties. Various regressors (support vector machine, random forest, linear regression) trained with 178 reference substances predicted the R_F values of 20 reference substances belonging to different chemical classes. We show that the performance of the model is bound to the similarity between the training and the test sets. The proposed methodology further encourages the use of computational methods for evaluation of HPTLC data. Thus, the nature of an unknown zone within the chromatogram could be matched with potential candidates based on predicted R_F.