Transfer learning in DeepLC improves LC retention time prediction across substantially different modifications and...

Area of Science:

• Proteomics
• Analytical Chemistry

Background:

• Peptide retention time prediction is valuable but limited by experimental variations.
• Inaccurate predictions hinder peptide identification, validation, and DIA spectral library generation.
• Current mitigation strategies like calibration or bespoke models offer only partial success.

Purpose of the Study:

• To demonstrate transfer learning as a robust solution for accurate peptide retention time prediction.
• To overcome limitations imposed by variations in experimental parameters in liquid chromatography (LC).
• To improve the adaptability and performance of prediction models across different peptide modifications and LC conditions.

Main Methods:

• Leveraging pre-trained model parameters for transfer learning.
• Applying transfer learning to peptide retention time prediction models.
• Evaluating model performance on peptide modifications and LC conditions distinct from training data.

Main Results:

• Transfer learning successfully overcomes limitations caused by experimental parameter variations.
• Highly performant models were achieved even for peptide modifications and LC conditions different from the original training.
• The adaptability of transfer learning proved effective across a wide range of proteomics workflows.

Conclusions:

• Transfer learning offers a highly robust solution for accurate peptide retention time prediction.
• This approach significantly enhances the reliability of predictions for various proteomics applications.
• Transfer learning broadens the applicability of retention time prediction models in diverse LC-MS workflows.