HDSE-MS: Tandem Mass Spectrum Prediction for Small Molecules via Hierarchical Distance Structural Encoding

Area of Science:

• Computational Chemistry
• Analytical Chemistry
• Bioinformatics

Background:

• Tandem mass spectrometry (MS/MS) is crucial for small molecule identification in metabolomics.
• Limited experimental spectral libraries hinder accurate MS-based identification.
• Existing spectrum prediction methods struggle with generalizability due to representation and enumeration limitations.

Purpose of the Study:

• To develop an advanced MS/MS spectrum prediction model for enhanced small molecule identification.
• To improve the molecular representation capabilities of spectrum prediction methods.
• To address the generalizability issues in current predictive models.

Main Methods:

• Proposed HDSE-MS, integrating a message passing neural network (MPNN) with a Transformer architecture.
• Employed hierarchical distance structural encoding (HDSE) using graph coarsening for multilevel cluster structures.
• Encoded hierarchical distances as structural biases within the Transformer to model substructures and long-range dependencies.

Main Results:

• HDSE-MS demonstrated superior performance on NIST23, MoNA, and MassBank datasets with high spectral entropy similarities (0.759, 0.567, 0.483).
• Achieved a Rank of 220.8 and Top-1 accuracy of 0.098 on the CASMI2022 test set.
• Exhibited strong predictive accuracy, robust generalization, and scalability in experiments.

Conclusions:

• HDSE-MS significantly advances MS/MS spectrum prediction for small molecule identification.
• The model's hierarchical structural encoding enhances molecular representation and predictive power.
• HDSE-MS offers a scalable and generalizable solution for metabolomics research.