Deep learning embedder method and tool for mass spectra similarity search.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Spectral similarity calculation is crucial for protein identification and mass spectra clustering in large-scale proteomics datasets.
• Traditional methods for spectral similarity can be computationally intensive and may not fully capture complex spectral features.

Purpose of the Study:

• To enhance spectral similarity comparison in proteomics using deep learning.
• To develop and evaluate a novel deep learning model (DLEAMSE) for spectral similarity.
• To introduce a bioinformatics tool (mslookup) for efficient searching of identified mass spectra.

Main Methods:

• Developed and assessed a deep learning embedder model (DLEAMSE) trained on high-quality spectra from PRIDE Cluster.
• Compared DLEAMSE performance against traditional methods like normalized dot product (NDP) in terms of accuracy and computational time.
• Created the mslookup tool for searching public repositories and spectral libraries, and released a human database.

Main Results:

• DLEAMSE demonstrated comparable accuracy to NDP for spectral similarity calculations.
• DLEAMSE's GPU implementation showed faster preprocessing and similarity computation (Euclidean distance on 32-D vectors) compared to NDP.
• The deep learning model's embedding step is a one-time process, allowing for faster future comparisons and large-scale data analysis.

Conclusions:

• Deep learning, exemplified by DLEAMSE, offers an effective approach to improve spectral similarity comparison in proteomics.
• The mslookup tool provides a valuable resource for researchers to search and manage identified mass spectra.
• This work facilitates more efficient analysis of large-scale proteomics data and promotes data sharing.