Purification of Pharmaceuticals via Retention Time Prediction: Leveraging Graph Isomorphism Networks, Limited Data, and Transfer Learning.
Journal:
Journal of separation science
Published Date:
Jun 1, 2025
Abstract
The design-make-test cycle for drug discovery is highly dependent on the purification of synthesized compounds. Prior to evaluation of suitability, ultrahigh-performance liquid chromatography is used for an initial standard analysis, where retention times of analytes are measured with a shorter standard gradient method and used to select the appropriate gradients for a final purification method. To circumvent this preliminary screening experiment for small molecule libraries, retention time prediction had been achieved previously by the use of commercial modeling methods. However, these retention time prediction models can have limited applicability when built from smaller datasets and are less effective when constructed from disparate data collected under differing chromatography conditions. Having thousands of measured retention times from high-throughput physiochemical screening, we sought to leverage these data for the construction of predictive models for a standard preliminary method enabling high-throughput purification of macrocyclic peptide libraries. Utilizing 4549 analytes and their retention times from high-throughput physiochemical screening, a structure-to-retention-time model was built using a graph isomorphism network, a form of artificial neural network architecture. Once fitted to high-throughput screening data, the model was re-trained with standard gradient method data, a technique known as transfer learning. Through transfer learning, a training set of 80 analytes yielded a neural network model that, when evaluated against a test set of 24 analytes, displays high performance metrics with a coefficient of determination (R) of 0.82 and mean average error of 0.088 min, or 1.26% of the gradient time. Comparatively, the best commercial quantitative structure-retention relationship model poorly performed, with an R of 0.11 and mean average error of 0.202 min. This model has been deployed internally as a Dash app to help democratize the use of the developed models and is being used for selecting purification methods based on analyte structure.