A Spectrum-to-Structure Diffusion Model for De Novo Small Molecule Generation.

Area of Science:

• Computational chemistry
• Cheminformatics
• Mass spectrometry analysis

Background:

• Identifying small molecule structures from mass spectra is crucial for metabolomics, proteomics, and drug discovery.
• Interpreting complex mass spectra to reconstruct molecular structures is a significant computational challenge.

Purpose of the Study:

• To develop an advanced computational framework, DiffNovo, for enhanced small molecule identification from mass spectral data.
• To improve the accuracy and efficiency of reconstructing molecular structures using fragmentation patterns.

Main Methods:

• Proposed DiffNovo, a spectrum-to-structure diffusion framework integrating a BART-based SMILES encoder-decoder and a high-resolution spectrum encoder.
• Utilized a conditional diffusion model with a forward noise-adding process and a reverse denoising process guided by spectral fragmentation patterns.

Main Results:

• DiffNovo demonstrated superior performance on the NIST20 dataset compared to existing methods.
• Achieved higher accuracy in SMILES-level identification, fingerprint-based similarity, and physicochemical property alignment.

Conclusions:

• DiffNovo effectively reconstructs molecular structures from mass spectra, showing high fidelity to spectral data.
• The framework offers a significant advancement in computational small molecule identification for various scientific domains.