Retention time prediction of emerging contaminants via transfer learning with graph neural networks

This study tackles the bottleneck challenge of retention time (RT) prediction for environmental organic contaminants in liquid chromatography-mass spectrometry (LC-MS) non-targeted analysis by proposing a graph neural network (GNN) transfer learning approach based on the METLIN-SMRT dataset. To overcome the limitations of laborious experimental determination and data scarcity, we systematically evaluated 5 GNN models, 3 pre-training optimizers, 3 training strategies, and 2 optimizers for transfer learning through a phased optimization workflow. Results show that GNNs effectively capture structure-RT relationships by encoding molecular graph topologies. By transferring knowledge from the source domain (METLIN-SMRT, containing natural products, metabolites and drug-like molecules) to a target domain of 1051 environmental pollutants, the optimal model - GIN3 with a finetune strategy and L-BFGS optimizer- achieved an R² of 0.894, significantly surpassing the best traditional machine learning method (R²=0.816). Gradient-based attribution analysis revealed how transfer learning shifts attention to key structural motifs, offering interpretability into model decision. Applicability domain assessment further confirmed the reliability of predictions. This work demonstrates three advantages of graph-based transfer learning: superior generalization (train-test decline of 8 % vs. 11-19 % for conventional methods), representation autonomy, and statistical robustness (SD = 0.020 vs. 0.039). By leveraging cross-domain chromatographic knowledge, the model overcomes dataset size constraints and enables accurate RT prediction with minimal target data, offering an efficient solution for rapid contaminant screening-especially in emergencies involving unknown pollutants or scarce standards-thus reducing experimental reliance and advancing intelligent chromatographic analysis.