On the feasibility of deep learning applications using raw mass spectrometry data

Area of Science:

• Proteomics
• Bioinformatics
• Machine Learning

Background:

• SWATH-MS is a powerful proteomic technique but has complex data analysis requirements.
• Current methods may discard valuable biological information by focusing only on a subset of peptides.
• There is a need for streamlined analysis of mass spectrometry data.

Purpose of the Study:

• To demonstrate deep learning's capability to extract features directly from raw mass spectrometry (MS) data.
• To develop a method that eliminates the need for extensive data processing and expert curation.
• To assess the potential of deep learning for distinguishing tumor from normal prostate biopsies using MS data.

Main Methods:

• Utilized transfer learning with pre-trained deep learning models for natural image classification.
• Applied models to raw MS images to generate feature vectors.
• Trained a classifier using these feature vectors to differentiate between tumor and normal prostate tissue samples.

Main Results:

• Achieved a classification performance of 0.876 AUC in distinguishing prostate tumor from normal biopsies.
• Demonstrated that deep learning can effectively learn from raw MS data without extensive preprocessing.
• Investigated the impact of image preprocessing and the inclusion of secondary MS2 spectra on classification performance.

Conclusions:

• Deep learning, combined with transfer learning, offers a powerful approach for analyzing raw mass spectrometry data.
• This methodology can potentially simplify proteomic data analysis and reduce the need for expert intervention.
• The findings suggest a promising future for deep learning in real-time diagnostic applications using MS data.