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

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Graph Neural Networks (GNNs) excel at predicting small molecule retention times (RT).
• Limited training data for specific chromatographic systems hinders GNN performance due to costly RT measurement experiments.
• Transfer learning offers a solution by leveraging abundant data from related tasks.

Purpose of the Study:

• To develop an improved transfer learning method for enhanced RT prediction in chromatography.
• To utilize a pretrained GNN with a small target dataset for accurate molecular RT prediction.

Main Methods:

• Employed a graph isomorphism network as the GNN architecture.
• Pretrained the GNN on the METLIN-SMRT dataset.
• Fine-tuned the GNN on the target dataset using a limited-memory BFGS optimizer with learning rate decay.

Main Results:

• The proposed transfer learning method demonstrated superior predictive performance compared to existing methods.
• Effectiveness was particularly pronounced when using small training datasets.
• The approach showed strong generalization across various chromatographic systems.

Conclusions:

• The improved transfer learning method effectively predicts molecular RT with limited data.
• This approach addresses the challenge of data scarcity in chromatographic system analysis.
• Future work could involve integrating multiple small datasets for further performance enhancement.