Abstract
INTRODUCTION: Metabolite annotation is a critical and challenging step in mass spectrometry-based metabolomic profiling. In a typical untargeted MS/MS-based metabolomic study, experimental MS/MS spectra are matched against those in spectral libraries for metabolite annotation. Yet, existing spectral libraries comprise merely a marginal percentage of known compounds. OBJECTIVE: The objective is to develop a method that helps rank putative metabolite IDs for analytes whose reference MS/MS spectra are not present in spectral libraries. METHODS: We introduce MetFID, which uses an artificial neural network (ANN) trained for predicting molecular fingerprints based on experimental MS/MS data. To narrow the search space, MetFID retrieves candidates from metabolite databases using molecular formula or m/z value of the precursor ions of the analytes. The candidate whose fingerprint is most analogous to the predicted fingerprint is used for metabolite annotation. A comprehensive evaluation was performed by training MetFID using MS/MS spectra from the MoNA repository and NIST library and by testing with structure-disjoint MS/MS spectra from the NIST library, the CASMI 2016 dataset, and in-house MS/MS data from a cancer biomarker discovery study. RESULTS: We observed that training separate models for distinct ranges of collision energies enhanced model performance compared to a single model that covers a wide range of collision energies. Using MetaboQuest to retrieve candidates, MetFID prioritized the correct putative ID in the first place rank for about 50% of the testing cases. Through the independent testing dataset, we demonstrated that MetFID has the potential to improve the accuracy of ranking putative metabolite IDs by more than 5% compared to other tools such as ChemDistiller, CSI:FingerID, and MetFrag. CONCLUSION: MetFID offers a promising opportunity to enhance the accuracy of metabolite annotation by using ANN for molecular fingerprint prediction.