pDeep3: Toward More Accurate Spectrum Prediction with Fast Few-Shot Learning

Area of Science:

• Computational chemistry
• Proteomics
• Mass spectrometry

Background:

• Deep learning significantly improved spectrum prediction accuracy.
• Existing methods face limitations due to variations in fragmentation types and instrument settings.

Purpose of the Study:

• To address limitations in deep learning spectrum prediction by employing few-shot learning.
• To improve prediction accuracy and efficiency across different mass spectrometry instruments and settings.

Main Methods:

• Implemented a few-shot learning approach for online data fitting.
• Evaluated the method on ten datasets using various instruments (Velos, QE, Lumos, Sciex) with different collision energies.
• Tested performance on collision-induced dissociation (CID) and higher-energy collision dissociation (HCD) spectra.

Main Results:

• Few-shot learning achieved higher prediction accuracy with negligible computational resources.
• On an untrained Sciex-6600 instrument, prediction accuracy increased from 69.7% to 86.4% in approximately 10 seconds.
• For CID spectra, accuracy improved from 48.0% to 83.9% when using a model trained on HCD spectra.
• The method demonstrated robustness to data quality and efficiency in closing accuracy gaps.

Conclusions:

• Few-shot learning is an effective strategy to enhance deep learning-based spectrum prediction.
• The approach significantly improves accuracy and generalizability across diverse mass spectrometry data.
• This method offers a computationally efficient solution for improving spectrum prediction accuracy.