Infrared spectra prediction using attention-based graph neural networks

Abstract

Infrared (IR) spectroscopy is an analytical technique that is used in broad disciplines of fundamental and applied research areas. Along those lines, it is of significant interest to develop an efficient computational method for the prediction of IR spectra based on chemical structures. In this work, we investigated the performance of attention-based graph neural networks. Our study showed that AttentiveFP model, which incorporates the message passing and graph attention mechanisms, exhibits the highest performance among the tested graph neural network models and a benchmark descriptor-based model. The implication is that the capability of AttentiveFP model to learn interactions between neighboring atoms and distant atoms allows for improving the performance of IR spectra prediction model. The mean Spearman correlation coefficient between the IR spectra predicted by AttentiveFP model and actual spectra was 0.911 and the correlation coefficient values were above 0.8 in 88% of the test cases. Our findings demonstrate the utility of the graph attention mechanism for the development of graph neural network-based machine learning models for IR spectra prediction with improved performances.