Deep MS/MS-Aided Structural-Similarity Scoring for Unknown Metabolite Identification

Abstract

Tandem mass spectrometry (MS/MS) is the workhorse for structural annotation of metabolites, because it can provide abundance of structural information. Currently, metabolite identification mainly relies on querying experimental spectra against public or in-house spectral databases. The identification is severely limited by the available spectra in the databases. Although, the metabolome consists of a huge number of different functional metabolites, the whole metabolome derives from a limited number of initial metabolites via bioreactions. In each bioreaction, the reactant and the product often change some substructures but are still structurally related. These structurally related metabolites often have related MS/MS spectra, which provide the possibility to identify unknown metabolites through known ones. However, it is challenging to explore the internal relationship between MS/MS spectra and structural similarity. In this study, we present the deep-learning-based approach for MS/MS-aided structural-similarity scoring (DeepMASS), which can score the structural similarity of unknown metabolite against the known one with MS/MS spectra and deep neural networks. We evaluated DeepMASS with leave-one-out cross-validation on MS/MS spectra of 662 compounds in KEGG and an external test on the biomarkers from male infertility study measured on Shimadzu LC-ESI-IT-TOF and Bruker Compact LC-ESI-QTOF. Results show that the identification of unknown compound is valid if its structure-related metabolite is available in the database. It provides an effective approach to extend the identification range of metabolites for existing MS/MS databases.

Deep metabolome: Applications of deep learning in metabolomics

“…Small molecule structure identification remains one of the biggest challenges in metabolomics (particularly for MS-based methods). Typically, retention time, accurate mass and mass spectra acquired from various analytical platforms are searched against reference databases [57] , [58] , [59] such as HMDB [60] , METLIN [61] and MassBank [62] to name a few. Similarities between unknow and reference compounds’ data are typically estimated based on correlation [63] , weighted cosine similarity [64] and Euclidean distance [65] which are used to rank the matching candidate hits [66] .…”

“…al. [59] , addresses the limitation of availability of spectra in the reference databases by increasing the chance to identify unknown compounds by augmenting the search results based on structural similarity to related known metabolites. The developed method leverages structural similarity between biochemical reactant and product pairs’ substructures and their resultant mass spectra [70] .…”

Network-based strategies in metabolomics data analysis and interpretation: from molecular networking to biological interpretation

“…Moreover, the recent success of molecular networking strategies has fired an exciting search for new methods of establishing structural relationships between mass spectra features. The recent development of MS2LDA [72] is a great example of how alternative metrics can be complementary to the already well-established spectra similarity, and the popularization of machine learning algorithms is providing some promising results in this area [123,124].…”

Spec2Vec: Improved mass spectral similarity scoring through learning of structural relationships

“…This is because it allows spectra to be queried against structural databases that are typically orders of magnitude larger than spectral ones. Despite the fact that database searching is still the gold standard for metabolite annotation 11 , methods such as MAGMa 12 , SIRIUS 13 , CSI:FingerID 5 , IOKR 14 , DeepMASS 15 , and MetDNA 16 have been effective at widening the search space and clearly demonstrate that useful structural information can be learnt from MS2 spectra. Building upon the insight that structural information can readily be learnt from spectra, we present a novel spectral similarity score based upon learnt embeddings of spectra.…”