CSU-MS²: A Contrastive Learning Framework for Cross-Modal Compound Identification from MS/MS Spectra to Molecular...

Area of Science:

• Computational chemistry
• Bioinformatics
• Analytical chemistry

Background:

• Tandem mass spectrometry (MS/MS) is crucial for identifying compounds in complex mixtures.
• Conventional spectral matching methods are limited by library coverage and algorithms.
• Accurate compound identification is essential for metabolomics and drug discovery.

Purpose of the Study:

• To develop a novel framework, CSU-MS2, for unifying MS/MS spectra and molecular structures.
• To improve the accuracy and efficiency of compound identification in complex mixtures.
• To provide a versatile tool for cross-modal retrieval of molecular candidates.

Main Methods:

• CSU-MS2 employs cross-modal contrastive learning to bridge MS/MS spectra and molecular structures.
• An External Space Attention Aggregation (ESA) module dynamically aligns spectral and structural features.
• The framework is pretrained on in-silico data and fine-tuned on experimental MS/MS data.

Main Results:

• CSU-MS2 achieved a Recall@1 of 75.45% against a library of over 1 million compounds, surpassing existing methods.
• Validation on diverse datasets demonstrated robust generalizability, with Recall@10 of 91.67% for blood metabolites.
• The framework significantly outperforms CFM-ID, SIRIUS, MetFrag, and CMSSP in compound identification.

Conclusions:

• CSU-MS2 offers a transformative solution for high-throughput compound identification.
• The framework's unified embedding space enables direct retrieval of molecular candidates.
• Publicly accessible code, models, and a comprehensive database (SSFDB) facilitate broad adoption in metabolomics and related fields.