DeepMASS v.2: An enhanced deep learning platform for large-scale discovery and structural annotation of unknown plant metabolites.

Abstract

Determining the structures of unknown metabolites remains a fundamental bottleneck in plant metabolomics, as the vast chemical diversity of plant secondary metabolites far exceeds the coverage of existing spectral libraries. Here, we present DeepMASS v.2, a substantially enhanced platform for annotating unknown metabolites from liquid chromatography-tandem mass spectrometry data, designed to address this challenge at scale. DeepMASS v.2 leverages a semantic spectral representation model trained on millions of spectra from GNPS, NIST, and in-house resources. By integrating Spec2Vec-based embeddings with HNSW (hierarchical navigable small world) graph retrieval and a unified chemical space defined by molecular fingerprints, DeepMASS v.2 identifies structurally related neighbors of unknown spectra and ranks candidate structures according to their proximity to the predicted structural neighborhoods within chemical space. Benchmarking against Critical Assessment of Small Molecule Identification datasets and a curated natural product collection demonstrated that DeepMASS v.2 outperforms state-of-the-art in silico annotation tools, including SIRIUS, CFM-ID, MetFrag, and MS-Finder. Importantly, DeepMASS v.2 maintains strong performance for metabolites absent from spectral libraries, highlighting its capacity to annotate genuinely unknown compounds. Application of DeepMASS v.2 to large-scale plant metabolomics datasets demonstrated its ability to expand accessible metabolome coverage. Implemented as an intuitive web platform, DeepMASS v.2 provides the community with a scalable, interpretable, and high-throughput solution for structural annotation, enabling more comprehensive characterization of plant chemical diversity and accelerating natural product discovery in molecular plant science. The DeepMASS v.2 web server is publicly available at http://deepmass.cn.