Mol2Raman: a graph neural network model for predicting Raman spectra from SMILES representations.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy is crucial for molecular analysis but computationally expensive to predict.
• Accurate prediction of Raman spectra is hindered by complex structure-spectrum relationships.

Purpose of the Study:

• Introduce Mol2Raman, a deep learning framework for direct Raman spectra prediction from SMILES strings.
• Enable accurate prediction of peak positions and intensities for diverse molecular structures.

Main Methods:

• Utilize Graph Isomorphism Networks with edge features (GINE) to encode molecular topology.
• Train the model on a dataset of over 31,000 DFT-calculated Raman spectra.
• Compare Mol2Raman against fingerprint-based and Chemprop models.

Main Results:

• Mol2Raman achieves high fidelity in predicting Raman spectra, outperforming existing methods.
• The model accurately predicts spectral features for structurally dissimilar molecules and enantiomers.
• Demonstrate fast inference times (22 ms/molecule) suitable for high-throughput screening.

Conclusions:

• Mol2Raman provides a scalable, accurate, and interpretable platform for Raman spectral prediction.
• The open-access web application facilitates real-time predictions without specialized hardware.
• This framework advances molecular design, materials discovery, and spectroscopic diagnostics.