Molspectra: a general framework for multi-spectra prediction from molecular structures.

Area of Science:

• Computational chemistry
• Machine learning in spectroscopy
• Molecular informatics

Background:

• Accurate molecular structure elucidation relies on integrating multiple spectroscopic techniques.
• A scarcity of high-quality, multi-modal spectral data hinders practical applications.
• MolSpectra addresses this bottleneck by predicting spectra from molecular representations.

Purpose of the Study:

• To develop a universal deep learning framework (MolSpectra) for predicting multi-modal molecular spectra.
• To enable synergistic analysis of spectral data by integrating molecular structure and experimental metadata.
• To provide a flexible and accurate tool for spectral prediction across various techniques.

Main Methods:

• Utilized a Message Passing Graph Neural Network (MPGNN) with Sequential Signal Mixing Aggregation (SSMA) and Hierarchical Distance Structural Encoding (HDSE).
• Integrated an enhanced Transformer architecture with a structurally biased attention mechanism.
• Employed flexible prediction heads for node-level (NMR) and graph-level (IR, UV-Vis, EI-MS) spectral predictions, supporting metadata input.

Main Results:

• Achieved high prediction accuracy across multiple spectroscopic modalities, outperforming baseline models.
• Demonstrated maximum cosine similarities of 0.916 for IR and 0.861 for UV-Vis predictions.
• Reached a maximum cosine similarity of 0.880 and Top-1 accuracy of 0.530 for EI-MS, with MAEs as low as 0.153 ppm for 1H NMR chemical shifts.

Conclusions:

• MolSpectra offers a robust and versatile deep learning solution for predicting diverse molecular spectra.
• The framework's ability to handle multi-modal data and metadata enhances spectral analysis accuracy.
• MolSpectra facilitates accelerated and more accurate molecular structure elucidation in cheminformatics and related fields.