Skip to content

MetDNA3

Summary

MetDNA3 annotates metabolites in untargeted LC-MS metabolomics by interlocking two network layers — a knowledge-driven metabolic reaction network (MRN) and a data-driven experimental feature network — so that annotation can propagate recursively outward from a small set of chemical-standard-confirmed "seed" metabolites. A graph neural network (GNN) trained on known reaction pairs expands the MRN's coverage and connectivity well beyond raw KEGG/MetaCyc/HMDB, and pre-mapping the two layers before propagation removes redundant computation, giving a >10-fold speedup and materially higher annotation coverage and accuracy than its predecessor, MetDNA2 [1].

Curated Metabolic Reaction Network (MRN)

Starting from KEGG, MetaCyc, and HMDB reaction pairs (RPs), a GNN was trained to predict reaction relationships between arbitrary metabolite pairs (AUC = 0.9947 on a held-out test set of 2,995 pairs), using structurally similar known RPs as training signal and a two-step pre-screen to limit false positives. Unknown/hypothetical metabolites generated by BioTransformer further extended coverage. The resulting curated MRN grew from the source databases to 765,755 metabolites and 2,437,884 candidate reaction pairs — with a higher global clustering coefficient and degree distribution than the raw knowledge databases (e.g., 5,892 metabolites at degree 10 in the curated MRN vs. 39 in the databases-only network) — while retaining a chemical-space distribution consistent with genuine metabolites (unlike a PubChem-derived comparison network).

Two-Layer Network Topology and Recursive Annotation

  1. Pre-mapping: experimental features are matched to the MRN via MS1 m/z (yielding an MS1-constrained MRN), reaction relationships are projected onto the resulting feature set to build a feature network, MS2 similarity filters prune spurious feature-feature edges, and the refined topology is mapped back to produce a data-constrained MRN. On a NIST human urine dataset this pipeline reduced the MRN from 765,755 to 2,993 metabolites (~0.4%) and 2,437,884 to 55,674 reaction pairs (~2.3%), discarding the vast majority of biologically irrelevant network mass for that sample.
  2. Recursive propagation: seed metabolites (features matched to chemical standards by MS1, retention time, and MS2) seed both layers. Each round retrieves reaction-paired neighbor metabolites (knowledge layer) and MS2-similarity-paired neighbor features (data layer); a candidate is only accepted as a new annotation if it has a pre-mapped cross-network link between the two layers. Newly annotated pairs reseed the next round, continuing until no further links are found.

Because candidates are checked against pre-computed cross-network links rather than recomputing m/z matches and MS2 similarities at every round, MetDNA3 searched far fewer neighbor metabolites and evaluated far fewer feature pairs than MetDNA2 for the same data, cutting mean per-dataset computation time from 1,082 to 77 minutes (>14-fold) across three biological sample types (12 datasets).

Performance

Benchmarked across BV2 cells, mouse brain, mouse liver, NIST human plasma, and NIST human urine (HILIC/RP columns × positive/negative ionization, Orbitrap and Astral instruments):

  • Coverage: 1,652 unique level-1 (standard-confirmed) metabolites annotated directly; network-based propagation extended this to 12,508 level-3 annotations (9,410 known, 3,098 unknown), including two endogenous metabolites absent from human metabolome databases. Astral-acquired data reached a 26.9% annotation rate versus 13.2% for Orbitrap.
  • Coverage/accuracy vs. MetDNA2 (20 Orbitrap datasets, 10-fold cross-validation holding out 70% of level-1 annotations as validation targets): mean annotation coverage rose from 39.9% (MetDNA2) to 68.1% (MetDNA3), and Top-3 correct rate rose from 65.6% to 84.4%. MetDNA3's Top-N correct rate reached 68.0% (Top 1), 84.4% (Top 3), and 91.0% (Top 10), with an estimated FDR of 15.6% for Top-3 annotations.
  • Mechanistic validation: predicted (GNN-derived) reaction pairs drove 68.0% of propagated annotations in a representative dataset, with structural/MS2 similarity of predicted pairs comparable to reported pairs, and 83.9% of propagation occurring within the same structural superclass — consistent with propagation tracking genuine chemical relationships rather than network artifacts.
  • Cross-tool corroboration: level-3 annotations lacking chemical standards were checked against four independent, non-network-based annotation tools (CFM-ID, MetFrag, MS-FINDER, SIRIUS) for consistency, since standards-based validation is not possible for most propagated annotations.

Availability

Two-layer networking algorithm (R): github.com/ZhuMetLab/MrnAnnoAlgo3. GNN reaction-relationship predictor (Python): github.com/ZhuMetLab/ReactionPredictor. Both are CC BY-NC-ND 4.0. A free-registration web server providing the full pipeline is available at metdna.zhulab.cn.

See Also

  • MetGenX — a de novo structure-generation tool from the same group, complementing MetDNA3's network-propagation approach for metabolites not yet present in any structural database.

Citations

[1] Zhang, H., Zeng, X., Yin, Y., & Zhu, Z.-J. (2025). Knowledge and data-driven two-layer networking for accurate metabolite annotation in untargeted metabolomics. Nature Communications, 16, 8118.