Machine Learning-Based Retention Time Prediction Tool for Routine LC-MS Data Analysis

The continuous evolution of machine learning (ML) methodologies and the aggregation of data have culminated in the development of highly accurate qualitative structure – property relationship (QSPR) models. For instance, liquid chromatography–mass spectrometry (LC-MS) data analysis widely used during chemical synthesis can be considerably improved by predicting retention time (RT). With hundreds of thousands of syntheses per year at Enamine Ltd., a large amount of high quality data has been produced from this analysis. In this paper we report on the development of an RT prediction model based on the GATv2Conv+DL graph neural network (NN) using internal data and evaluation of the selected NN architecture with METLIN SMRT dataset. The final model shows a mean absolute error (MAE) value of 2.83 s for the 120-s LC-MS method. It was integrated into the existing in-house LC-MS viewing tool. We also performed a thorough analysis of RT prediction errors and suggested the range between RT – 11.34 s and RT + 10.68 s, as this interval contained over 95% of the data.