Peak2Patch: High-Fidelity Functional Group Identification through Attention-Based Fusion of Infrared and Mass Spectra

Area of Science:

• Computational chemistry
• Spectroscopy
• Machine learning

Background:

• Molecular structure identification from spectroscopic data is crucial for chemistry and biology.
• Infrared (IR) Spectroscopy and Mass Spectrometry (MS) offer rich structural information but require expert interpretation.
• Current machine learning methods often analyze single spectra, missing complementary multimodal insights.

Purpose of the Study:

• To develop a novel multimodal approach for predicting molecular functional groups using IR and MS data.
• To compare the efficacy of different neural network architectures (transformers vs. CNNs) for spectral encoding.
• To evaluate various data fusion strategies (early, middle, late) for combining spectral information.

Main Methods:

• Implemented transformer neural networks for encoding both sparse MS and dense IR spectra.
• Investigated early, middle, and late fusion techniques to integrate multimodal spectral features.
• Introduced Peak2Patch, an attention-based fusion scheme utilizing cross-attention for feature interaction between modalities.

Main Results:

• Transformer networks outperformed convolutional neural networks in encoding spectral data.
• Deep feature fusion demonstrated significant potential for combining multimodal information.
• Peak2Patch achieved substantial improvements in functional group prediction accuracy on a large dataset.

Conclusions:

• Multimodal spectral data fusion, particularly through attention-based mechanisms like Peak2Patch, enhances molecular functional group prediction.
• Transformer networks are highly effective for processing diverse spectral data types.
• This approach offers a powerful, automated solution for molecular structure elucidation.