ChemEmbed: A deep learning framework for metabolite identification using enhanced MS/MS data and multidimensional molecular embeddings

Abstract

Machine learning tools have become essential for annotating the vast number of unidentified MS/MS spectra in metabolomics, addressing the limitations of current reference spectral libraries. However, these tools often struggle with the high dimensionality and sparsity of MS/MS spectra and metabolite structures. ChemEmbed introduces a novel approach by combining multidimensional and continuous vector representations of chemical structures with enhanced MS/MS spectra. This enhancement is achieved by merging spectra from multiple collision energies and incorporating calculated neutral losses from 38,472 distinct compounds, providing richer input for a convolutional neural network (CNN). ChemEmbed achieves top-ranked candidate annotations in over 42% of cases and identifies the correct compound within the top five in more than 76% of cases in a test dataset. Against external benchmarks such as CASMI 2016 and 2022, ChemEmbed outperforms SIRIUS, the current state-of-the-art in computational metabolomics. In a validation experiment with the Annotated Recurrent Unidentified Spectra (ARUS) dataset-- including over 25,000 spectra from human plasma and 68,000 from urine-- ChemEmbed successfully identified 24 previously unannotated compounds. By aligning with the advanced capabilities of modern mass spectrometry instrumentation, ChemEmbed balances accuracy, computational efficiency, and scalability, making it a powerful solution for high-throughput metabolomics applications.