LipidIN
Summary¶
LipidIN is a lipid annotation framework built around a 168.5-million-fragment, five-level hierarchical spectral library covering 121 lipid subclasses (including C=C double-bond position, resolved via Paternò-Büchi reaction and electron-activated dissociation). It pairs an ultra-fast expeditious querying (EQ) module with a lipid categories intelligence (LCI) module that exploits retention-time regularities to remove false positives and predict unannotated lipids, and adds a "reverse lipidomics" neural network — the Wide-spectrum Modeling Yield network (WMYn) — that regenerates high-accuracy, platform-transferable fingerprint spectra to boost recall [1].
Architecture¶
- Five-level hierarchical library: (1) precursor ion/head group, (2) fatty-acyl/alkyl side chain, (3) head-group neutral loss, (4) side-chain neutral loss, (5) regenerated fragment fingerprints — assembled from >100 published datasets plus Paternò-Büchi and electron-activated-dissociation experiments for C=C-location resolution.
- Expeditious querying (EQ) module: a non-informative-prior greedy/hash-table search with linear time complexity O(mΔ), versus the O(nmΔ) of flash entropy search. Reported throughput exceeds 100 billion queries/second, and one billion MS2 comparisons complete in 0.23 ms against a one-million-spectrum MassBank library (~60,000× faster than flash entropy on the same hardware).
- Lipid categories intelligence (LCI) module: statistically formalizes three intraclass relative-retention-time regularities — equivalent carbon number (ECN), intra-subclass unsaturation parallelism (IUP), and equivalent separated carbon number (ESCN) — derived from >100 published datasets (median ~90% conformity), and uses confidence-interval fitting plus best-first search to reject false positives and predict candidates lacking MS/MS fragments.
- Wide-spectrum Modeling Yield network (WMYn): a three-stage "reverse lipidomics" model that (1) maps sparse MS2 m/z-intensity matrices into a shared 512-dimension space per lipid, (2) applies a transformer-style multi-head self-attention encoder for inter-feature learning, and (3) down/upsamples to regenerate the 5th-level fingerprint library, enabling cross-instrument transfer. SiLU activation outperformed ReLU in small-sample training.
Performance¶
- Speed and coverage: EQ+LCI achieved >90% recall@Top-20 on a benchmark dataset (vs. ~70% for EQ alone against the same MS-DIAL library), outperforming MS-DIAL, LipidSearch, Spectral Entropy, and LipidMatch; LipidIN annotated 8,923 lipids across four datasets (RBL-2H3 cells, a mixture dataset, human sera, zebrafish tissue) at an average 5.69% estimated FDR.
- False-positive removal: the LCI retention-time model outperformed Lipid Data Analyzer (LDA) across almost all lipid subclasses.
- Reverse lipidomics validation: fingerprints regenerated by WMYn from 333 clinical sera samples matched three independent injections of 15 lipid reference standards with mean spectral (entropy) similarity of 0.9826, and retention-time deviation of ~0.03 min. WMYn generalized across instruments, matching reference standards from an Agilent LC-qTOF system at ~0.9 similarity despite being trained on Orbitrap Exploris 240 data.
- External benchmark: on a published aging-associated mouse lipidome atlas, LipidIN recovered 93.64% of the reported 2,704 lipids using its 5-level library, versus lower recall using the MS-DIAL public database alone. On NIST SRM 1950, LipidIN produced the highest Top@1 annotation count (434) among LipidIN, MS-DIAL, Entropy Search, Lipid Hunter, and Lipid Annotator.
Clinical Applications¶
- Breast cancer biomarker panel: across two independent cohorts (1,393 and 333 sera samples), LipidIN annotated 4,747 lipids across 53 subclasses. A 10-lipid panel selected via random forest and modeled with LightGBM classified breast cancer vs. healthy controls at 96.93% accuracy in the discovery cohort and 79.61% in the validation cohort. WGCNA linked hexosylceramide and ceramide levels to comorbid diabetes mellitus, and specific ceramide/lysophosphatidylcholine species to tumor grade and size.
- Lung metastasis discrimination: in a third cohort (105 sera samples spanning breast nodules, breast cancer without metastasis, breast cancer with lung metastasis, and female lung cancer), simple fold-change comparisons found no discriminating lipids, but double-bond-positional resolution (via Paternò-Büchi reaction) revealed group-specific fatty-acyl double-bond positions, e.g., elevated PC 18:1(Δ15) in lung-metastatic breast cancer.
Availability¶
Source code: github.com/LinShuhaiLAB/LipidIN; hierarchical fragment library and code archive on Zenodo; reproducible capsule on CodeOcean. Lipidomics data are deposited at Metabolomics Workbench (MTBLS10170) and the National Genomics Data Center (PRJCA028507).