Deciphering the Biological Function of Tandem Mass Spectra Using a Dual-Modal Metabolite Language Model
Liquid chromatography-tandem mass spectrometry untargeted metabolomics captures thousands of metabolic peaks from biological samples. Conventional identification-dependent metabolomics workflows require confident structural annotation before identified features can enter downstream pathway or functional analysis. However, only a small fraction of detected metabolic features can be assigned confident structural annotations, leaving a large fraction difficult to incorporate into downstream biological interpretation, a form of "functional dark matter". Here, we introduce Spec2Function, a function-centric framework that infers biological functions from tandem mass spectrometry (MS2) spectra. Spec2Function is trained using characterized spectrum-text pairs but does not require prior structural identification of the query spectra at inference time. At its core is Spec2Biotext, a dual-encoder contrastive model that embeds MS2 spectra and curated multi-aspect natural language functional texts into a unified functional space. Combined with a retrieval-augmented generation framework grounded in curated databases and biomedical literature, Spec2Function retrieves evidence-grounded biological knowledge and produces coherent functional summaries for individual spectra or groups of spectra. Across metabolite- and pathway-level benchmarks, Spec2Function achieved higher similarity to reference functional descriptions than the evaluated comparators and expanded the fraction of the metabolome accessible to functional interpretation. Application to a human liver cirrhosis cohort recapitulated known metabolic alterations and identified additional literature-supported functional modules. Applying Spec2Function to a 30-organ macaque aging atlas revealed organ-specific metabolic functions and previously unreported functional modules that were supported by independent human and mouse datasets. Together, these results position Spec2Function as a function-centric framework for scalable biological interpretation of structurally unannotated MS2 spectra.
Authors
- Feifan Zhang (ORCID: https://orcid.org/0009-0001-7505-3591)
- Yifan Fu (ORCID: https://orcid.org/0000-0001-9142-7648)
- Xiaotao Shen (ORCID: https://orcid.org/0000-0002-9608-9964)
- Chuchu Wang (ORCID: https://orcid.org/0000-0003-2015-7331)
- Jiramet Kinchagawat (ORCID: https://orcid.org/0000-0002-8294-0711)
- Yuting Wang
- Wenya Wang
- Yijiang Liu
- Qingpeng Kong
Institutions
- Nanyang Technological University (SG)
- Chinese Academy of Sciences (CN)
- Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
Publication Details
- Journal
- bioRxiv (Cold Spring Harbor Laboratory)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.64898/2026.11.16.766993
- Primary Topic
- Metabolomics and Mass Spectrometry Studies
- Type
- preprint
Funders
- Ministry of Education - Singapore
- Nanyang Technological University
- Lee Kong Chian School of Medicine, Nanyang Technological University