Deep graph convolutional network for small-molecule retention time prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in analytical chemistry

Background:

• Retention time (RT) is vital for molecule identification in liquid chromatography-mass spectrometry (LCMS).
• Graph neural networks (GNNs) show promise for RT prediction, but their depth has not been optimized.
• Accurate RT prediction aids in filtering molecular candidates with similar spectra but different RTs.

Purpose of the Study:

• To investigate the impact of GNN depth on RT prediction accuracy.
• To develop a deep GNN model for enhanced RT prediction in LCMS.
• To evaluate the model's performance across diverse chromatographic conditions and its utility in molecular structure identification.

Main Methods:

• Utilized deep graph convolutional networks (GCNs) with residual connections and edge information.
• Investigated the effect of increasing GNN depth up to 16 layers.
• Fine-tuned the DeepGCN-RT model on seven diverse LCMS datasets.
• Assessed the model's impact on molecular structure identification accuracy.

Main Results:

• A 16-layer GNN with residual connections significantly improved RT prediction.
• The DeepGCN-RT model achieved a mean absolute percentage error (MAPE) of 3.3% and mean absolute error (MAE) of 26.55 s on the SMRT test set.
• Fine-tuning reduced mean MAE by 30% across seven datasets compared to previous methods.
• DeepGCN-RT improved top-1 accuracy by 11% in molecular structure identification by reducing candidate structures by 30%.

Conclusions:

• Deeper GNNs, particularly with residual connections and edge information, enhance RT prediction accuracy in LCMS.
• The developed DeepGCN-RT model represents a significant advancement over existing methods for RT prediction and molecule identification.
• DeepGCN-RT demonstrates broad applicability across various chromatographic conditions and effectively assists in complex molecular structure elucidation.