Prediction of Retention Time by Combining Multiple Data Sets with Chromatographic Parameter Vectorization and...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Retention time (RT) is crucial for mass spectrometry-based compound identification.
• RT prediction is challenging due to sensitivity to experimental conditions and data sparsity.
• Existing machine learning models often lack generalizability across different chromatographic systems.

Purpose of the Study:

• To develop a robust machine learning method for accurate retention time prediction.
• To overcome data sparsity and improve model generalizability in chromatography.
• To enable efficient transfer learning for retention time prediction across diverse experimental setups.

Main Methods:

• Proposed a Multi-Dataset Learning with Transfer Learning (MDL-TL) approach.
• Vectorized chromatographic parameters (CPs) using word2vec and autoencoders.
• Integrated CPs into compound representation for joint multi-dataset training and fine-tuning.

Main Results:

• MDL-TL significantly outperformed five deep learning and four machine learning methods.
• Achieved superior performance in mean absolute error, median absolute error, mean relative error, and R² across 28 datasets (14 RP-LC, 14 HILIC).
• Demonstrated effective transferability to new chromatographic systems through fine-tuning.

Conclusions:

• MDL-TL offers a promising solution for accurate and generalizable retention time prediction.
• The method effectively leverages multi-dataset learning and transfer learning principles.
• MDL-TL enhances the reliability of compound identification in mass spectrometry-based analyses.