3D-MPEA: A Graph Attention Model-Guided Computational Approach for Annotating Unknown Metabolites in Interactomics via Mass Spectrometry-Focused Multilayer Molecular Networking

Abstract

The spectral matching strategy of MS2 fragment spectrograms serves as a ubiquitous method for compound characterization within the matrix. Nevertheless, challenges arise due to the deficiency of distinctions in spectra across instruments caused by coelution peak-derived fragments and incompleteness of the current spectral reference database, leading to dilemma of multidimensional omics annotation. The graph attention model embedded with long short-term memory was proposed as an optimized approach involving integrating similar MS2 spectra into molecular networks according to the isotopic ion peak cluster spacing features to collapse diverse ion species and expand the spectral reference library, which efficiently evaluated the substance capture capacity to 123.1% than classic substance perception tactics. The versatility and utility of the established annotation procedure were showcased in a study on the stimulation of pork mediated by 2,2-bis(4-hydroxyphenyl)propane and enabled the global metabolite annotation from knowns to unknowns at metabolite-lipid-protein level. On the spectra for which in silico extended spectral library search provided a group truth, 83.5−117.1% accuracy surpassed 1.2−14.3% precision after manual validation. β-Ala-His dipeptidase was first evidenced as the critical node related to the transformation of α-helical (36.57 to 35.74%) to random coil (41.53 to 42.36%) mediated by 2,2-bis(4-hydroxyphenyl)propane, ultimately triggering an augment of catalytic performance, inducing a series of oxidative stress, and further intervening in the availability of animal-derived substrates. The integration of ionic fragment feature networks and long short-term memory models allows the effective annotation of recurrent unknowns in organisms and the deciphering of unacquainted matter in multiomics.