Deep learning for retention time prediction in reversed-phase liquid chromatography

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• Retention time prediction in High-Performance Liquid Chromatography (HPLC) is crucial for identifying unknown molecules in untargeted profiling.
• Existing methods utilize diverse molecular properties and machine learning algorithms for retention time prediction.

Purpose of the Study:

• To develop and evaluate a deep learning model for accurate retention time prediction of small molecules.
• To leverage the large retention time dataset from the Metabolite and Chemical Entity Database (METLIN) for model training.

Main Methods:

• Explored various deep learning architectures, including models trained on molecular fingerprints and SMILES strings represented as one-hot matrices.
• Utilized a one-dimensional convolutional neural network (1D CNN) with SMILES as input.
• Transferred a pre-trained 1D CNN model to five additional datasets to assess generalization capabilities.

Main Results:

• The 1D CNN model using SMILES input achieved the best performance.
• The model reported a mean absolute error of 34.7 s and a median absolute error of 18.7 s, outperforming previous benchmarks on the METLIN dataset.
• The pre-trained model demonstrated good generalization ability across different datasets.

Conclusions:

• A 1D CNN model trained on SMILES strings provides a highly accurate method for retention time prediction in HPLC.
• The model's performance and generalization capabilities highlight its potential for improving molecule identification in complex mixtures.