Natural-product classes and biosynthesis — NPClassifier vs ClassyFire, class counts in COCONUT/LOTUS/NPAtlas, per-class MS/MS difficulty, building-block mass ladders

TL;DR. NPClassifier (7 pathways / 70 superclasses / 672 classes, trained on 73,607 NPs) is a biosynthesis-flavoured ontology that beats the rule-based ClassyFire on 47 of 62 NP classes, and SIRIUS's CANOPUS predicts both ontologies from a spectrum. Natural-product databases are not balanced: terpenoids are ~39 % of LOTUS unique structures and ~46 % of plant-derived ones, while NPAtlas (microbial) is 34 % polyketides / 27 % peptides for bacteria. No published benchmark gives a clean per-NPClassifier-class accuracy for structure elucidation; the closest evidence (CASMI 2022, CANOPUS hold-outs, flavonoid-glycoside isomer work, lipidomics guidelines) says positional/skeletal isomers inside a class — not the class itself — are where MS/MS ranking fails, and terpenoids (median 8 stereocentres, 98 % chiral) are the worst-case family for a 2D-InChIKey14 metric that collapses stereo but not skeleton isomers.

1. NPClassifier vs ClassyFire vs CANOPUS

1.1 NPClassifier (Kim et al. 2021, J Nat Prod 84:2795, DOI 10.1021/acs.jnatprod.1c00399; PMC8631337)

An independent 2025 re-implementation on 77,705 NPClassifier-labelled molecules (Comput Struct Biotechnol J, PMC12800374) reports test macro-F1 of 0.967 (pathway), 0.916 (superclass), 0.803 (class) for GNNs versus 0.965 / 0.870 / 0.688 for an MLP on counted Morgan FPs — i.e. the class level is where structure-based classification itself starts to fail, before any spectrum is involved.

1.2 ClassyFire (Djoumbou Feunang et al. 2016, J Cheminform 8:61, DOI 10.1186/s13321-016-0174-y; PMC5096306)

• ChemOnt: 4,825 categories (4,146 organic, 678 inorganic, + root), up to 11 levels, average depth 5. Kingdom 2 → Superclass 31 → Class 764 → Subclass 1,729 → levels 5–11: 2,296 more.
• Fully rule-based: >9,000 SMARTS patterns, ~4 % Markush, logical expressions, ~200 regex on IUPAC names. Deterministic: same structure → same answer, no probabilities.
• Validation: 99.97 % agreement on an 800-compound / 21,102-assignment test set; 77 M compounds classified.
• Admitted gaps: "cannot comprehensively cover biosynthetic origins"; alkaloids and peptidomimetics are defined as unions of sub-categories; complex fused ring systems incompletely covered.
• Practical consequence: a diterpenoid is "Lipids and lipid-like molecules → Prenol lipids → Diterpenoids" in ClassyFire, but "Terpenoids → Diterpenoids → Labdane diterpenoids" in NPClassifier (see the COCONUT row for CNP0241103 below). ClassyFire's "Lipids" superclass therefore swallows most terpenoids, steroids and fatty acyls — a trap when reading per-"lipid" statistics.

1.3 CANOPUS (Dührkop et al. 2021, Nat Biotechnol 39:462, DOI 10.1038/s41587-020-0740-8) — what it predicts from a spectrum

• Predicts 2,497 ClassyFire classes from MS/MS. Architecture (Fig. 7 caption): kernel SVMs (CSI:FingerID) predict the molecular fingerprint from the spectrum; a DNN then predicts classes from fingerprint + formula, with Platt calibration; the DNN is trained on fingerprints computed from the structure database, "without the need for spectral training data".
• Reported: "average accuracy of 99.7 % in cross-validation" (dominated by true negatives across 2,497 mostly-absent classes — see Gotchas). Extended Data Fig. 1: MCC sunburst for the 782 of 2,497 classes with ≥50 positive examples. A search-result snippet of the paper reports 370 classes with MCC ≥0.8 and 943 of 1,270 classes with MCC ≥0.5 (examples: phosphocholines 0.954, flavonoid O-glycosides 0.918, pregnane steroids 0.867) — not verified verbatim by me; confidence medium.
• Hold-out experiment (Extended Data Fig. 2, verbatim): "We remove all flavonoid glycosides (the subclass) from the MS/MS training data … CANOPUS exhibits only a small drop (68 % to 97 %) in correct classifications. In contrast, direct prediction … misses almost all of them (8 %) afterwards." Cause: isoflavonoid glycosides in training led the direct model to use "presence of a sugar residue to infer that a MS/MS spectrum is not a glycoside". This is the only quantified "novel-class" evaluation I found; note it is class-level (glycoside yes/no), not pathway-level.
• NPClassifier from spectra: SIRIUS docs (v6 "Methods background") state CANOPUS now predicts both ontologies — ClassyFire ("more than 2500 compound classes") and NPClassifier ("more general" but "may align better with biosynthetic pathways"). No accuracy figures for the NPClassifier head are published in the docs or in a paper I could open; the docs also warn there is "no universally accepted threshold" for the posterior probabilities and that for compounds far from training data "CANOPUS may return lower probabilities for all compound classes". Confidence that the NPClassifier head reaches the ClassyFire head's quality: low/unknown.

1.4 CASMI 2022 (Fiehn lab, UC Davis) — community numbers for class vs structure

Results page lists 21 submissions. "Priority" set (compounds not in public libraries), best per category:

"Bonus" set: classes 75–77 % (Dührkop 123/165, Wang 149/194), structures ≤29 %. The gap "formula 94 % → class 69 % → structure 26 %" is the cleanest published quantification of how much is lost at each rung of the ladder on a novel-compound set. No per-class breakdown is given on the results page (the methods doc is a Google-Drive link I did not open).

2. Class distribution per database (NPClassifier pathway)

All counts below are my own computation on the raw downloads (files, dates and scripts noted), because none of the three papers prints a pathway table. Multi-pathway labels are exploded (a hybrid counts once in each pathway), so columns can sum to >100 %.

2.1 COCONUT (CSV-lite release 2026-10, downloaded 2026-10-07; 695,133 structures in the Sept-2024 paper release)

738,794 rows in the 2026-10 CSV-lite (stereo-variants are separate rows; 489,862 unique base CNP identifiers). COCONUT stores exactly one NPClassifier pathway per entry (pathway-count distribution: 1 → 686,525; 0 → 52,269; never 2), unlike LOTUS/NPAtlas — hybrids are collapsed to one label. Stereocentres computed with RDKit 2026.03.6 (FindMolChiralCenters, includeUnassigned; 174 parse failures).

Overall: MW median 433 Da; heavy atoms median 31 (mean 36.2); 83.6 % of rows with ≥1 stereocentre (median 3); MW bins: ≤200 4 %, 200–300 13 %, 300–400 24 %, 400–500 22 %, 500–600 11 %, 600–800 11 %, 800–1,000 9 %, 1,000–1,500 4.5 %, >1,500 0.8 %. Top superclasses: Terpenoids → Diterpenoids 45,548, Triterpenoids 45,442, Sesquiterpenoids 42,836, Steroids 39,445, Monoterpenoids 13,260; Alkaloids → Tryptophan 29,816, Lysine 18,592, Tyrosine 15,474, Pseudoalkaloids 15,354; Shikimates → Flavonoids 38,636, Coumarins 25,992, Lignans 8,496, Isoflavonoids 7,985; Fatty acids → Glycerolipids 48,475 (of which Triacylglycerols 44,564 = the single largest NPClassifier class in COCONUT), Glycerophospholipids 9,536; Polyketides → Macrolides 10,182, Polycyclic aromatic PKs 7,750. ClassyFire cross-tab: "Lipids and lipid-like molecules" (235,316 rows) = 147,366 NPClassifier-terpenoids + 62,956 fatty acids + 10,920 alkaloids + 7,462 polyketides — the ClassyFire "lipid" superclass is 63 % terpenoid.

Two COCONUT-specific warnings fall out of the table: (i) the alkaloid pathway has mean NP-likeness 0.07 (vs 2.51 for terpenoids), i.e. COCONUT's 173 k "alkaloids" are heavily diluted with synthetic-looking N-heterocycles from aggregated sources, which is why alkaloids are 23 % of COCONUT but only 11–13 % of LOTUS/NPAtlas; (ii) the fatty-acid pathway's MW median of 807 Da is an artefact of 44.6 k enumerated triacylglycerols.

Paper facts (Chandrasekhar et al. 2025, NAR 53:D634, DOI 10.1093/nar/gkae1063): 695,133 unique structures from 63 sources; 82,220 without stereocentres, 539,350 with preserved stereo, 73,563 with undefined stereo → 88.2 % carry ≥1 stereocentre; stereoisomers are separate entries; NP-likeness is multimodal (peaks at −2…0, 0…1, ~2), which the authors read as "the need for a comprehensive curation process … to remove non-natural products".

2.2 LOTUS (frozen dataset 2026-04-13, Zenodo record 19360665; 709,740 structure–organism pairs; 227,298 unique InChIKeys; 151,885 unique InChIKey14)

Overall: 85.9 % of unique LOTUS structures have ≥1 stereocentre; median 5, mean 6.5; median exact mass 405 Da. Multi-pathway labels: ~6 % of unique structures.

Pathway share by producing kingdom (unique structure×kingdom, exploded, %):

Paper facts (Rutz et al. 2022, eLife 11:e70780): >750,000 pairs, >290,000 unique 2D structures at publication; plants 28,439 organisms / 95,191 unique structures (545 chemical classes), fungi 4,003 / 22,594 (417 classes), animals 2,716 / 15,242, bacteria 1,555 / 15,895 (385 classes); authors "favored [NPClassifier] over ClassyFire simply because it was tailored to describe NPs".

2.3 NPAtlas (JSON download dated 2024-09-13; 36,454 compounds: 22,995 fungal, 13,459 bacterial)

10.8 % of NPAtlas entries carry two or more pathways (Peptides$Polyketides 1,997; Alkaloids$Peptides 891; Polyketides$Terpenoids 628). By origin: bacteria = 34.5 % polyketides, 26.9 % peptides, 5.5 % terpenoids; fungi = 28.3 % polyketides, 26.9 % terpenoids, 8.3 % peptides — matching the paper's qualitative Fig. 3 statement (van Santen et al. 2022, NAR 50:D1317, PMC8728154: "bacterial classes are dominated by compounds of peptidic and hybrid PKS-NRPS origin, whereas most fungal classes include large numbers of polyketide and terpenoid natural products"). Top bacterial superclasses: Oligopeptides 4,356; Macrolides 1,748; Polycyclic aromatic polyketides 1,018. Top fungal: Sesquiterpenoids 2,354; Meroterpenoids 1,922; Peptide alkaloids 1,386; Cyclic polyketides 1,331.

2.4 The 77,705-molecule NPClassifier benchmark set (PMC12800374, Table 2) for comparison

Fatty acids 29.5 %, Terpenoids 29.1 %, Shikimates/Phenylpropanoids 12.1 %, Alkaloids 10.9 %, Polyketides 10.9 %, Peptides 5.3 %, Carbohydrates 1.1 % — note how fatty acids dominate this curated training-style set but are only 4–5 % of LOTUS/NPAtlas; a model trained on NPClassifier-like data will be over-exposed to lipids relative to the NP databases.

3. Which classes are intrinsically hard for MS/MS — evidence table

Search of CASMI 2022, CANOPUS, COSMIC, MassSpecGym, MIST, DiffMS, MSNovelist and two 2026 benchmark papers: none publishes a per-NPClassifier-class or per-ClassyFire-superclass structure-retrieval accuracy table. MassSpecGym (arXiv 2410.23326; 231 k spectra / 29 k molecules; retrieval Hit@1 14.6 % for MIST, 2.5 % fingerprint-FFN, 0.37 % random with ≤256 candidates) gives none; Che et al. 2026 (arXiv 2601.00941) break down only by adduct; De Waele et al. 2026 (arXiv 2602.16507) only by loss function. What exists:

Reading across: the class is rarely the problem; the within-class isomer family is. The 2D-InChIKey14 metric used by CASMI 2026 forgives stereo (removing most of the terpenoid stereo burden) but not positional/skeletal isomers, so flavonoid-glycoside regio-isomers, terpene skeletons and acyl-position isomers all count as misses.

4. Building blocks and mass signatures (monoisotopic; C 12.000000, H 1.007825, N 14.003074, O 15.994915, S 31.972071)

Near-isobaric traps that any ladder-finder must handle: C2H4 (28.0313) vs CO (27.9949) vs N2 (28.0061) — 36 mDa apart, resolved at timsTOF ~0.03 Da/40 k resolution but not at unit mass; C5H8 ≡ 5·CH2 − H2 exactly (so a +68.06 ladder is also a lipid homologue series). Leu/Ile, Gln/Lys (128.0586 vs 128.0950, Δ36 mDa) and hydroxyproline/Leu-type overlaps persist.

5. Descriptor distributions per pathway (from §2 computations + literature)

• Stereocentres: LOTUS unique structures — terpenoids 98.4 % chiral (median 8), carbohydrates 99.5 %, peptides 97.4 % (median 7), vs alkaloids 72 %, shikimates 74 %, polyketides 75 % (median 2–3). Overall 85.9 % (LOTUS), 88.2 % (COCONUT, from 82,220/695,133 achiral). NPAtlas: terpenoids 96.8 %, peptides 98.5 %, shikimates only 59 %.
• Mass: median neutral exact mass ~405 Da (LOTUS) / MW ~436 Da polyketides, ~716 Da peptides (NPAtlas). Peptides and PKS–NRPS hybrids are the only pathways with q90 >1,300 Da; terpenoid q90 ≈ 830 Da (saponins). Capecchi & Reymond 2021 (J Cheminform 13:82, PMC8524952) on COCONUT: MW range 81 (1,2-dihydropyridine) to 2,901 Da (lacticin 481); "plant NPs dominated the intermediate molecular weight range (200 < MW < 800), while fungal NPs were most abundant in the low molecular weight range (MW ≤ 200) and bacterial NPs in the high MW range (MW ≥ 800)"; high-Fsp3 region = terpenes, steroids, glycosides; low Fsp3 = polyphenols.
• Heavy atoms / Fsp3 per pathway — see the COCONUT table in §2.1 (computed from the CSV-lite heavy_atom_count and fractioncsp3 columns).

Forced analogy: music theory — "instrument families"

Statement. An NPClassifier pathway is an orchestral family (strings, woodwind, brass, percussion). Each family's sound is defined by its overtone series — the physical repeat unit of the vibrating body. The biosynthetic building block (C5H8, C2H2O, CH2, amino-acid residue, hexose) is the overtone series: the spectrum of a molecule should show a "harmonic" ladder at integer multiples of its unit, and that ladder (plus the mass defect, i.e. the "formant") should let you name the family from the sound alone, before you name the instrument (class) or the note (structure).

Testable idea 1 — explicit ladder features. Compute, per spectrum, the count and summed intensity of peak pairs (and neutral-loss-to-precursor differences) separated by 68.0626, 42.0106, 14.0157, 162.0528/146.0579/132.0423, and by the 20 residue masses, within ±5 mDa (timsTOF tolerance). Use these ~30 numbers as (a) a pathway prior to re-weight the COCONUT candidate pool, and (b) an auxiliary head when training the spectrum→fingerprint model. Test: does adding the ladder features raise MRR@25 on a held-out slice of enveda-np-examples (99.6 % in COCONUT) stratified by pathway? Expected: large gains for peptides/depsipeptides and glycosides, small for alkaloids, near-zero for terpenoids.

Testable idea 2 — family-aware re-ranking. Run CANOPUS (SIRIUS 6) NPClassifier pathway probabilities on the test spectra, then multiply candidate scores by P(pathway | spectrum) × P(candidate's NPClassifier pathway). Validate on the CASMI-2022-like split: pathway accuracy from spectrum must exceed the LOTUS prior (39 % terpenoids) to add information.

Does the analogy hold?

• Holds for peptides/NRPS (residue ladder = harmonic series; b/y ions are literally partials), polyketide macrolides (42.01 Da and sugar losses), glycosides (162/146/132 Da — the sugar is a detachable "mute" whose loss is heard first), and fatty acyls (14.016 Da homologue spacing). CANOPUS's 2,497-class prediction at "99.7 % accuracy" and its 97 % (→68 % when the subclass is unseen) glycoside recall are the quantitative evidence that family can be heard; CASMI 2022's class 69 % vs structure 26 % says the family is far more audible than the note.
• Partially holds for shikimates/flavonoids: the overtone is a fixed C6–C3 aglycone "fundamental" (271/287/303 m/z) rather than a repeating unit — more like a fixed-pitch percussion instrument than a string.
• Breaks for terpenoids: the unit C5H8 is incorporated once as a skeleton and then rearranged (cyclisations, Wagner–Meerwein shifts), so there is no ladder to hear; the spectrum is dominated by generic H2O/CH3/C4H8 losses and C8–C11 hydrocarbon ions shared by hundreds of skeletons. Terpenoids are the largest family in the NP orchestra (39–46 % of plant NPs) and the most "mute" — the analogy predicts, correctly, that this is where candidate ranking will collapse. It also breaks for alkaloids, where the "family" is defined by a nitrogen source rather than a repeat unit.
• Mixed ensembles: 10.8 % of NPAtlas and ~6 % of LOTUS structures carry two pathways (NRPS–PKS hybrids median mass ~796 Da, meroterpenoids, alkaloid-terpenoids median 11 stereocentres). Their spectra superimpose two ladders; a single-family classifier is forced into a wrong binary choice, exactly as a timbre classifier fails on a chord played by two sections. NPClassifier's DAG (multi-label) was designed for this; a softmax pathway head is not.

Verdict: the analogy pays off as a feature-engineering prescription (ladders as explicit timbre descriptors, multi-label not softmax) and as a difficulty forecast (terpenoids are the silent family); it mis-predicts nothing I could test, but it is weakest exactly where the challenge is hardest.

Gotchas / what could fool us

1. "Lipids" means different things. ClassyFire's superclass "Lipids and lipid-like molecules" contains terpenoids, steroids, prenol lipids and fatty acyls; NPClassifier splits them into Terpenoids vs Fatty acids. Any per-"lipid" accuracy from a ClassyFire-based benchmark is mostly a terpenoid number.
2. CANOPUS's 99.7 % is micro-accuracy over 2,497 mostly-negative labels. The honest numbers are MCC per class on the 782 classes with ≥50 positives; the paper's own hold-out shows a 97 → 68 % drop when a subclass is unseen, and SIRIUS docs warn probabilities deflate for out-of-distribution compounds — i.e. exactly the hidden-test situation ("none previously published").
3. No NPClassifier-head accuracy is published for CANOPUS. Treat pathway-from-spectrum as a prior of unknown calibration until measured on enveda-np-examples.
4. Database priors are kingdom priors. Terpenoids are 46.5 % of plant structures but 6.2 % of bacterial ones; peptides 1.5 % vs 27.5 %. A COCONUT-wide prior will systematically mis-rank microbial-origin test compounds (Enveda's own libraries are plant-heavy; the hidden set is "natural products, hypothesised NPs, analogs" — origin mix unknown).
5. 15.4 % of NPAtlas has no NPClassifier pathway at all, and 3.3 % of LOTUS; "(none)" is itself a class (often atypical or very small molecules). Do not silently drop them when building class-stratified validation.
6. Hybrid labels are not rare (10.8 % NPAtlas): a single-label pathway model will be wrong ~1 in 10 on microbial NPs by construction.
7. Multi-pathway labels use different separators across files (LOTUS " $ ", NPAtlas JSON lists, COCONUT CSV its own) — a naive value_counts() reports "Alkaloids $ Terpenoids" as a separate pathway and under-counts both parents.
8. Stereo is forgiven, isomers are not. InChIKey14 matching ignores the median-8 stereocentres of terpenoids, but a wrong skeleton, wrong sugar position or wrong acyl position is a miss. The isomer burden per formula is therefore highest for terpenoid formulas (C15H24O2, C20H30O3, C30H48O3 …) — count COCONUT candidates per formula to see this directly.
9. COSMIC's audit: 48/315 "novel" annotations were proteinogenic peptides; peptide-rich test subsets inflate library-like performance without being hard.
10. In-source water loss ([M+H−H2O]+ formula accuracy 0.16 vs 0.99 for [M+H]+ in Che et al. 2026) hits exactly the polyol-rich families (terpenoids, steroids, glycosides); adduct mis-assignment masquerades as class difficulty.
11. Class-imbalance in training sets ≠ databases: the 77,705-molecule NPClassifier-style set is 29.5 % fatty acids; LOTUS is 5 %. A fingerprint model's "easy" lipids may be over-represented in train and under-represented in the hidden test.
12. CASMI 2022 class scoring dropped compounds ClassyFire could not key, so its 69 % class accuracy is on an easier-than-full subset.
13. The flavonoid-glycoside isomer ViT reports accuracy on its spectra (one platform); 84.7 % is not transferable to timsTOF merged-CE spectra without re-testing.

Implications for the challenge

1. Build the validation split stratified by NPClassifier pathway and by hybrid/non-hybrid, with the LOTUS plant-vs-microbe mix as two separate scenarios; report MRR@25 per pathway, not just overall. Expect terpenoids to be the floor.
2. Add explicit mass-ladder features (68.0626, 42.0106, 14.0157, 162.0528, 146.0579, 132.0423, 176.0321, residue masses; ±5 mDa) to the spectrum encoder and as a cheap rule-based pathway prior; measure the per-pathway delta.
3. Run SIRIUS 6 CANOPUS (NPClassifier + ClassyFire heads) on the visible and hidden-style spectra and use pathway/superclass posteriors to re-weight COCONUT candidates — but calibrate first on enveda-np-examples, because the NPClassifier head has no published accuracy.
4. For every candidate formula, pre-compute the number of COCONUT candidates and their pathway composition; when the pool is terpenoid-dominated (>60 %) with >100 isomers, invest the 25 slots in skeleton diversity (cluster by Murcko framework / NPClassifier class and fill round-robin) rather than fingerprint-similarity order.
5. Treat the glycoside flag as a first-class prediction: a predicted sugar loss (162/146/132/176) should hard-filter candidates on np_classifier_is_glycoside (COCONUT column) before ranking; CANOPUS shows glycoside-ness is learnable even for unseen subclasses (68 % vs 8 %).
6. For positional flavonoid-glycoside and acyl-position isomers, do not spend model capacity trying to separate them from a single spectrum: order them by database prior (3-O before 7-O, etc., by COCONUT frequency) and accept the MRR hit; this is where the 25-slot budget pays.
7. Use NPClassifier multi-label (DAG) outputs everywhere; never force a softmax over 7 pathways — ~10 % of microbial NPs are hybrids.
8. Normalise COCONUT, LOTUS and NPAtlas pathway strings to one separator before building any prior (see Gotcha 7); the per-pathway JSONs beside this file are already exploded.
9. Expect in-source dehydrated precursors in the terpenoid/steroid/glycoside subsets; test adduct handling on [M+H−H2O]+ rows of train before trusting pathway predictions on them.

Open questions

• What is the measured accuracy of CANOPUS's NPClassifier pathway/superclass head on timsTOF merged-CE spectra of novel NPs? (Nobody has published it; a 2-hour experiment on enveda-np-examples would settle it.)
• Has anyone published per-NPClassifier-class retrieval accuracy for CSI:FingerID / MIST / DiffMS on MassSpecGym? (Not found; the 2026 benchmark papers break down only by adduct/loss.)
• COCONUT 2.0 supplementary material may contain a pathway table I could not open (reCAPTCHA on PMC); my counts are from the 2026-10 CSV, which is 1 year newer than the paper release — the two may differ by tens of thousands of entries.
• Terpenoid-specific MS/MS benchmarks (e.g. sesquiterpene lactone or diterpenoid isomer panels on Q-TOF) exist in the phytochemistry literature; the WebSearch budget ran out before I could locate one with numbers.
• Does the hidden test's "hypothesised NPs / analogs / synthetics plausibly in nature" shift the pathway prior toward low-stereo, low-MW chemistry (alkaloid/shikimate-like) rather than terpenoid-like? The enveda-180 synthetic library (0.03 % COCONUT overlap) suggests yes for the synthetic slice.

Sources (opened)

• Kim HW et al. NPClassifier. J Nat Prod 2021;84:2795. DOI 10.1021/acs.jnatprod.1c00399 — full text via PMC8631337.
• Djoumbou Feunang Y et al. ClassyFire. J Cheminform 2016;8:61. DOI 10.1186/s13321-016-0174-y — PMC5096306.
• Dührkop K et al. CANOPUS. Nat Biotechnol 2021;39:462. DOI 10.1038/s41587-020-0740-8 — main text + Fig. 5–7 and Extended Data Fig. 1–2 captions (nature.com); bioRxiv 10.1101/2020.04.17.046672 was rate-limited (429) and not read.
• SIRIUS 6 documentation, "Methods background" (CANOPUS section): https://v6.docs.sirius-ms.io/methods-background/
• Chandrasekhar V et al. COCONUT 2.0. Nucleic Acids Res 2025;53:D634. DOI 10.1093/nar/gkae1063 — academic.oup.com page + Europe PMC full text (PMC11701633).
• COCONUT bulk download page https://coconut.naturalproducts.net/download → coconut_csv_lite-10-2026.zip (199 MB, dated 2026-10-01).
• Rutz A et al. LOTUS. eLife 2022;11:e70780. DOI 10.7554/eLife.70780 — PMC9135406; frozen dataset Zenodo record 19360665 (260413_frozen_metadata.csv.gz, 2026-04-13).
• van Santen JA et al. NPAtlas 2.0. Nucleic Acids Res 2022;50:D1317. DOI 10.1093/nar/gkab941 — PMC8728154; data: https://www.npatlas.org/static/downloads/NPAtlas_download.json (2024-09-13).
• Fiehn lab, CASMI 2022 results: https://fiehnlab.ucdavis.edu/casmi/casmi-2022-results
• Bushuiev R et al. MassSpecGym. arXiv 2410.23326 (HTML).
• Che X, Du X, Xu D. Comparative analysis of formula and structure prediction from tandem mass spectra. arXiv 2601.00941 (2026-01-02).
• De Waele G et al. Small molecule retrieval from tandem MS: what are we optimizing for? arXiv 2602.16507 (2026-02-18).
• Hoffmann MA et al. COSMIC. Nat Biotechnol 2022;40:411. DOI 10.1038/s41587-021-01045-9 — PMC8926923 (Europe PMC XML).
• Köfeler HC et al. Quality control requirements for the correct annotation of lipidomics data. Nat Commun 2021;12:4771. DOI 10.1038/s41467-021-24984-y — PMC8346590.
• Capecchi A, Reymond J-L. Classifying natural products from plants, fungi or bacteria using COCONUT and machine learning. J Cheminform 2021;13:82. DOI 10.1186/s13321-021-00559-3 — PMC8524952.
• Leveraging molecular graphs for natural product classification. Comput Struct Biotechnol J 2025 — PMC12800374 (Europe PMC XML; Tables 2, 4, 6).
• Determination of flavonoid glycoside isomers using vision transformer and tandem MS. PMC11644359 (2024).
• Atomic masses: IUPAC/NIST monoisotopic values as listed in §4; masses computed by script (scratchpad), not copied.
• Not opened / unverified: Jena CANOPUS blog (503), bioRxiv CANOPUS (429), Phytochemistry Reviews 2025 review (Springer cookie wall returned abstract only), arXiv 2110.15047 terpene PDF (binary garbled), ResearchGate figure pages (403).