CSU-EP: Contrastive Learning for Unifying Experimental and Predicted EI-MS Spectra

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• Spectral library matching is standard for mass spectrometry (MS) compound identification but limited by experimental library coverage.
• The experiment-to-prediction gap hinders the use of in-silico libraries due to discrepancies between experimental and predicted spectra.

Purpose of the Study:

• To develop a framework, CSU-EP, that bridges the experiment-to-prediction gap for improved compound identification.
• To create a unified spectral embedding database (CSU-EP-DB) using predicted spectra.

Main Methods:

• A two-stage training strategy involving self-supervised pretraining on predicted electron ionization (EI) mass spectra and contrastive learning for spectral unification.
• Generation of 2.24 million spectra-predicted spectra into the CSU-EP-DB.

Main Results:

• CSU-EP achieved a Recall@1 of 47.10% on the NIST replib benchmark, outperforming existing methods.
• Demonstrated effectiveness in identifying compounds not present in experimental libraries.
• Incorporating a molecular weight filter in a plasma metabolomics application boosted Recall@5 to 92.86%.

Conclusions:

• CSU-EP establishes a new paradigm for reliable and scalable compound identification using in-silico spectral libraries.
• The framework effectively bridges the experiment-to-prediction gap, enhancing MS-based compound discovery.
• A user-friendly web server is available for broad accessibility.