Graphormer-IR: Graph Transformers Predict Experimental IR Spectra Using Highly Specialized Attention.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Infrared (IR) spectroscopy is crucial for chemical and forensic analysis.
• Developing accurate and rapid *in silico* spectral prediction methods is essential for real-time experimental feedback.
• Current computational models face challenges in speed and accuracy for complex spectral features.

Purpose of the Study:

• To develop a highly accurate and efficient computational model for predicting IR spectra.
• To utilize Graphormer, a graph neural network (GNN) transformer, for IR spectral prediction solely from Simplified Molecular Input Line Entry System (SMILES) strings.
• To investigate the impact of enhanced node embeddings on spectral prediction accuracy.

Main Methods:

• Employed Graphormer, a GNN transformer architecture, for IR spectral prediction.
• Utilized a dataset of 53,528 high-quality IR spectra across five experimental media.
• Incorporated novel architectural features including a global node for phase encoding, learned node feature embeddings, and a 1D smoothing convolutional neural network (CNN).

Main Results:

• Graphormer-IR achieved a mean test spectral information similarity (SISμ) of 0.8449 ± 0.0012, outperforming the state-of-the-art Chemprop-IR.
• Augmenting node embeddings with additional descriptors improved the SISμ score to 0.8523 ± 0.0006.
• The model effectively captured long-range interactions, anharmonic peak positions, and stretching frequencies of uncommon functional groups.

Conclusions:

• Graphormer-IR demonstrates superior performance in predicting IR spectra compared to existing methods.
• The model's architecture excels at capturing complex molecular interactions and spectral details.
• This approach offers significant potential for accelerating computational chemistry and experimental spectroscopy.