Abstract
In the domain of metabolomics, the accurate identification of compounds is paramount. However, this process is hindered by the vast number of metabolites, which poses a significant challenge. In this study, a novel approach to compound identification is proposed, namely a molecular-fingerprint prediction method based on the graph attention network (GAT) model. The method involves the processing of fragmentation-tree data derived from tandem mass spectrometry (MS/MS) data computation and the subsequent processing of fragmentation-tree graph data with a technique inspired by natural language processing. The model is then trained using a 3-layer GAT model and a 2-layer linear layer. The results demonstrate the method's efficacy in molecular-fingerprint prediction, with the prediction of molecular fingerprints from MS/MS spectra exhibiting a high degree of accuracy. Firstly, this model achieves excellent performance in receiver operating characteristic (ROC) and precision-recall curves. The factors that have the most influence on the resultant performance are identified as edge features using different training parameters. Then, better performance is achieved for accuracy and F (1) score in comparison with MetFID. Secondly, the model performance was validated by querying the molecular libraries through methods commonly used in related studies. In the results based on precursor mass querying, the proposed model achieves comparable performance with CFM-ID; in the results based on molecular formula querying, the model achieves better performance than MetFID. This study demonstrates the potential of the GAT model for compound identification tasks and provides directions for further research.