Deep vs. Shallow Learning-based Filters of MSMS Spectra in Support of Protein Search Engines

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Mass spectrometry-based proteomics generates large MSMS spectra datasets.
• Current protein search engines face challenges with processing time and identifying non-identifiable spectra.
• Efficient filtering of MSMS spectra is crucial for accelerating proteomic data analysis.

Purpose of the Study:

• To evaluate the efficacy of machine learning algorithms in filtering non-identifiable MSMS spectra.
• To compare the performance of deep learning models against various shallow learning algorithms.
• To assess the impact of simple feature sets on MSMS spectra filtering.

Main Methods:

• A deep learning algorithm was compared against nine shallow learning algorithms.
• Experiments were conducted using ten diverse datasets from different sources and species.
• Performance was evaluated based on the exclusion of non-identifiable spectra and the retention of identifiable ones.

Main Results:

• Deep learning models demonstrated significant power in filtering MSMS spectra.
• The deep learning model excluded approximately 50% of non-identifiable spectra with only a 9% loss of identifiable spectra.
• Shallow learning algorithms (Random Forest, SVM, Neural Networks) effectively filtered spectra, removing 70% of non-identifiable ones at the cost of 25% identifiable spectra loss.

Conclusions:

• Machine learning, particularly deep learning, offers an efficient solution for filtering non-identifiable MSMS spectra.
• Shallow learning algorithms provide valuable alternatives, especially for highly abundant proteins.
• The choice of algorithm may depend on protein concentration and expression levels in specific proteomic studies.