Deep Learning for Bidirectional Translation between Molecular Structures and Vibrational Spectra.

Area of Science:

• Computational chemistry
• Spectroscopy
• Artificial intelligence

Background:

• Molecular vibrational spectra and Simplified Molecular Input Line Entry System (SMILES) are crucial for chemical identification.
• Establishing a direct, bidirectional link between these two representations remains a challenge.
• Existing methods often lack accuracy or efficiency in spectral interpretation.

Purpose of the Study:

• To develop deep learning models for translating between molecular spectra and SMILES representations.
• To improve the accuracy and efficiency of spectral interpretation using artificial intelligence.
• To enable the recognition of functional groups and differentiation of isomers and homologues from spectral data.

Main Methods:

• Development of two deep learning models: TranSpec and SpecGNN.
• Implementation of techniques including model fusion, transfer learning, and multisource learning.
• Augmentation of datasets and application of molecular mass filtering.
• Utilizing SpecGNN for spectral simulation and candidate reordering.

Main Results:

• Initial TranSpec accuracy reached 55-63% for calculated spectra, but dropped to 11% for experimental IR data.
• Improved methods boosted TranSpec accuracy to 53.6% for experimental IR data.
• SpecGNN demonstrated superior spectral accuracy and computational efficiency compared to traditional quantum chemistry methods.
• Successful recognition of functional groups and distinction between isomers/homologues was achieved.

Conclusions:

• TranSpec and SpecGNN offer an efficient and accurate AI-driven framework for molecular structure and spectra interpretation.
• These models advance applications in spectroscopy and cheminformatics.
• The developed models provide a powerful tool for chemical structure elucidation from spectral data.