Retention Time Prediction with Message-Passing Neural Networks

Abstract

Retention time prediction, facilitated by advances in machine learning, has become a useful tool in untargeted LC-MS applications. State-of-the-art approaches include graph neural networks and 1D-convolutional neural networks that are trained on the METLIN small molecule retention time dataset (SMRT). These approaches demonstrate accurate predictions comparable with the experimental error for the training set. The weak point of retention time prediction approaches is the transfer of predictions to various systems. The accuracy of this step depends both on the method of mapping and on the accuracy of the general model trained on SMRT. Therefore, improvements to both parts of prediction workflows may lead to improved compound annotations. Here, we evaluate capabilities of message-passing neural networks (MPNN) that have demonstrated outstanding performance on many chemical tasks to accurately predict retention times. The model was initially trained on SMRT, providing mean and median absolute cross-validation errors of 32 and 16 s, respectively. The pretrained MPNN was further fine-tuned on five publicly available small reversed-phase retention sets in a transfer learning mode and demonstrated up to 30% improvement of prediction accuracy for these sets compared with the state-of-the-art methods. We demonstrated that filtering isomeric candidates by predicted retention with the thresholds obtained from ROC curves eliminates up to 50% of false identities.

Retention Time Prediction through Learning from a Small Training Data Set with a Pretrained Graph Neural Network

“…The Projection method was similar to those presented in Domingo-Almenara et al 4 's study. The combination of FE and Adam was similar to those presented in Yang et al, 17 Yang et al, 18 and Osipenko et al 19 's studies.…”

“…It is the most widely used optimizer for training neural networks on relatively large data sets. Using the Adam optimizer, training was performed based on the configuration described in the “Pretraining” subsection except that the mini-batch size was reduced to 32, which is similar to the configurations used in previous studies. − For holdout validation, the target training data set was split in an 8:1 ratio.…”

“…A GNN can directly process the graph representation. , Previous studies have attempted to build GNNs as RT prediction models to achieve an improved predictive performance. Various GNN architectures have been evaluated for their suitability for RT prediction, including graph convolutional network (GCN) and its variants, , message passing neural network (MPNN), , and many others …”

Generic and accurate prediction of retention times in liquid chromatography by post-projection calibration

“…To the best of our knowledge, the largest experimental RT dataset at present is METLIN (http://metlin.scripps.edu) containing 80,038 molecules, which was determined on the Zorbax Extend-C18 reverse-phase (RP) column by using 0.1% formic acid in water and acetonitrile as mobile phases. This dataset has been reported multiple times for establishing structure-based models, with a median prediction error of 1.6% −6.2% [19][20][21][22] . However, these models usually worked for speci c LC setups, mainly due to the variance in RT between different CMs.…”

“…These predictions are extremely accurate over alternative methods, with a median prediction error of less than 3.7%. For the metabolomics community, such approaches had been used to share the RTs observed in publicly available databases and the RTs predicted by QSRR models with other CMs/laboratories 9,[19][20][21][22]31 .…”