Natural-product chemical space — COCONUT / LOTUS / NPAtlas sizes, overlaps, and what they can and cannot cover for CASMI 2026

TL;DR

COCONUT's October-2026 release holds 738,794 records but only 479,690 unique InChIKey14s — and InChIKey14 is the only identity the CASMI 2026 metric scores, so anyone quoting "695k candidates" is overcounting the real pool by 1.54×. I downloaded and recomputed everything below rather than trusting press numbers: NPAtlas adds only 138 new InChIKey14s on top of COCONUT (99.59% already contained), the competition's enveda-180 library overlaps COCONUT by 51 of 182,941 molecules (0.028%), and — a finding I could not find stated anywhere — enveda-180 is Enveda's own MS/MS + ion-mobility library built from the Enamine REAL make-on-demand space, per Enveda's public enveda/enveda-180 repository. Candidate coverage, not spectral modelling, is the binding constraint: a typical natural product has 27 COCONUT InChIKey14 competitors inside ±5 ppm (median; p90 = 193), and sesquiterpene/diterpene formulas routinely carry 600–1,300.

1. Database sizes, dates, formats, licences

All counts in the "recomputed" column were derived by me on 2026-10-07 from the actual download files, not from the papers.

Union, in the metric's own currency. COCONUT(IK14) ∪ NPAtlas(IK14) = 479,828 — NPAtlas contributes 138 genuinely new skeletons. COCONUT ∪ all 275,810 training structures = 729,389 IK14. (A public write-up quotes "about 712k structures: COCONUT plus every structure in the training set" — same construction, slightly older COCONUT snapshot.)

COCONUT's internal provenance (top contributors, recomputed)

54.5% of COCONUT records cite exactly one collection; the long tail reaches 51 collections for a single molecule. ANPDB and NANPDB are byte-identical in counts (5,897 / 5,306 / 0 sole-sourced) — the same African dataset entered twice under two names.

2. How much of COCONUT is "real"

• Organism-linked: 273,549 / 738,794 = 37.0%. Only 32.7% carry a DOI; 7.0% carry a CAS number. So roughly two-thirds of COCONUT has no stated producing organism, which is the operative definition of "a natural product" in this competition's framing.
• Annotation level (1★ worst … 5★ best, plus an undocumented level 0): 0 → 76,679; 1 → 239,701; 2 → 148,594; 3 → 163,128; 4 → 57,742; 5 → 52,950. Only 7.2% are 5★; 42.8% are level ≤1. 74,164 records (10.0%) are simultaneously level 0 and nameless.
• COCONUT's inclusion filter is chemical, not evidentiary. The paper states that after the ChEMBL Checker step, "synthetic compounds such as fluorine-containing compounds, sulfonamides … or drug derivatives, and entries containing multiple organic structures (multi-component entries) are removed." The filter leaks both ways: 1,315 fluorine-containing records survive (0.18%), and genuine natural organofluorines and natural sulfonamides are excluded by construction.
• NP Classifier pathways (records): Terpenoids 209,865 · Alkaloids 173,145 · Shikimates/phenylpropanoids 116,239 · Fatty acids 83,412 · Polyketides 61,388 · Amino acids & peptides 33,829 · Carbohydrates 8,647 · unclassified 52,269.
• Enumerated-looking contamination exists but is small. The single largest formula bucket is C6H5N3O with 1,761 IK14s at 10 heavy atoms each, all stereo-free, mostly nameless, and reading as exhaustively generated bicyclic heteroaromatics. By contrast the other mega-buckets are real: C20H30O3 (1,283), C15H20O4 (1,186), C15H24 (1,001) — terpenoid isomer space genuinely is that dense.

3. Overlaps, recomputed in InChIKey14 space

| Set A | |A| (IK14) | in COCONUT | % | |---|---|---|---| | NPAtlas v2024_09 | 33,497 | 33,359 | 99.59% | | enveda-np-examples (host-measured NPs, timsTOF) | 250 | 249 | 99.6% | | masaryk | 416 | 353 | 84.9% | | mona | 11,681 | 9,450 | 80.9% | | msdial | 9,127 | 7,326 | 80.3% | | riken | 15,892 | 12,552 | 79.0% | | drug_plus | 2,539 | 1,949 | 76.8% | | massbank | 9,180 | 5,795 | 63.1% | | gnps | 45,750 | 16,912 | 37.0% | | spectraverse | 9,631 | 2,092 | 21.7% | | pluskal_ms2 | 46,821 | 8,128 | 17.4% | | enveda-180 | 182,941 | 51 | 0.028% | | All of train.parquet | 275,810 | 26,111 | 9.47% |

The 9.47% figure is the one to internalise: the training set and the candidate database are almost disjoint, because enveda-180 (66% of all training structures) is synthetic.

4. enveda-180: identified, not guessed

Confidence: high. Enveda publishes the generating code at github.com/enveda/enveda-180 (created 2026-05-19, MIT, 8 stars), titled "Enveda-180: High throughput generation of a large open multimodal small molecule LC-MS/MS and ion mobility library". The README states the project "generates a large open multimodal MS/MS + ion mobility spectral library for drug-like small molecules from the Enamine REAL space", runs over 14 "enamine" batches of Bruker .d raw files, and its environment variables are literally ENAMINE_PROCESSING_DATA_PATH / ENAMINE_PREP_DATA_PATH. Citation given as Krettler, Kind, DeBloois, Killian, Prince, Taylor, Gadiya, Simpson, Domingo-Fernández, Allen, Colluru, Healey (2026), ChemRxiv; the DOI printed in the README (10.26434/chemrxiv.15004319) did not resolve for me (404 at doi.org, absent from Crossref) — treat the preprint pointer as unverified. Data DOI badge: 10.5281/zenodo.20436851 (Zenodo blocked this host, so unverified); a third-party library index lists Enveda-180's data licence as CC-BY. Not MCE, not MolPort — Enamine REAL.

Targeted adducts in the pipeline: positive [M+H]+ [M+Na]+ [M+NH4]+ [M+K]+ [2M+Na]+ [2M+H]+ [M-H2O+H]+ [M-H4O2+H]+; negative [M-H]- [M+Cl]- [M+FA-H]- [M+Br]- [M+Hac-H]- [2M-H]- [2M+FA-H]- [2M+Hac-H]- [M-H2O-H]-; cation [M]+.

My own read of train.parquet corroborates it compositionally: enveda-180 = 1,153,785 spectra / 182,941 IK14, 100% timsTOF, 77.0% positive / 23.0% negative. Its molecules are 99.96% nitrogen-containing, 23.0% fluorine-containing, 28.3% sulfur-containing, with a startlingly tight mass distribution (MW median 330.1, IQR 304–348). That is the fingerprint of amide/sulfonamide combinatorial make-on-demand chemistry. It also explains the 0.028% overlap mechanically: COCONUT deliberately deletes fluorinated compounds and sulfonamides, so a large slice of enveda-180 could not appear in COCONUT even if nature made it.

5. Chemical space: NP vs synthetic, from the actual files

COCONUT descriptors computed by the COCONUT pipeline itself (all 738,794 records): fraction sp3 median 0.58 (mean 0.572, Q1 0.36, Q3 0.80); AlogP median 3.28; rings median 4; aromatic rings median 1; TPSA median 85.7; rotatable bonds median 5; HBD/HBA median 2/7; QED median 0.41; NP-likeness (Ertl) median 1.25 (Q1 0.14, Q3 2.25); 13.0% contain a sugar; 15.8% flagged glycoside. The headline NP-vs-synthetic contrast in the literature — natural products carry fewer aromatic rings and more stereocentres while logP/TPSA are not dramatically different (Ertl and co-workers; NP-likeness as a Bayesian NP-vs-synthetic separator) — is reproduced here mainly in the halogen and nitrogen channels, which are far sharper discriminators in this particular pair of datasets than MW or RDBE.

I did not obtain verified Enamine REAL / ChEMBL / ZINC descriptor tables from the primary literature: the PDF reprints I fetched (Ertl JCIM 2008; Ertl CS3-2018 slides) came back as unparsed binary, and the session's web-search budget was exhausted before I could substitute another source. enveda-180 is used above as an empirical proxy for Enamine REAL, which is defensible given its stated provenance but is not the same thing as REAL's published statistics. Confidence on the Enamine/ChEMBL column: medium-low.

6. How much of the hidden test could even be in COCONUT?

The hidden set is ~400 molecules: confirmed NPs, hypothesised NPs, NP analogues, and synthetics "plausibly in nature", explicitly "none previously published" in public spectral libraries. "Not previously published as a spectrum" ≠ "not previously published as a structure" — that distinction is the whole estimate.

Assumptions, stated so they can be attacked:

1. enveda-np-examples (250 molecules, host-measured on the same timsTOF) is the host's own released sample of the test chemistry. 99.6% of it is in COCONUT. That is an upper bound, because it is a demo set and was probably drawn from well-characterised material.
2. Pye et al.'s retrospective (52,395 microbial + marine NPs, 1941–2015, from AntiMarin plus manual curation of 29,062 articles) found median maximum Tanimoto to prior compounds plateauing at ~0.65 since the mid-1990s, with 76.8% of the corpus inside fewer than 6,500 Tversky clusters and >23% singletons. Their reply to Skinnider & Magarvey adds the sharpest single number: in 2015, 1,576 compounds from 484 papers, and 49% of them had another compound in the same paper at Tanimoto > 0.9. New NP reports are overwhelmingly congeneric series.
3. Skinnider & Magarvey's counter-letter (in-house DB of 32,380 NPs) showed those same trends appear after randomly permuting discovery year and even when NPs are swapped for ZINC compounds, and that relative to that null, post-1990 NP discovery is more diverse than expected (KS p = 6.2 × 10⁻¹⁴). Contested: the raw novelty-decline trend is a database-growth artefact; the field's novelty rate is not actually collapsing.
4. COCONUT at 479,690 IK14 already contains an order of magnitude more structures than Pye's 52,395, because it ingests plant/TCM/food space.

Estimate. Partitioning the hidden 400 by the host's own four categories and applying the above:

So roughly 45–55% of the hidden set is plausibly present in COCONUT at InChIKey14, with a wide band (35–70%). This is strikingly consistent with the independently back-solved figure in the mission brief — recall@25 ≈ 0.55 at an MRR of 0.417 — and it says the ceiling is candidate presence, not ranking skill. Adding the 275,810 training structures lifts the pool to 729,389 IK14 but adds mostly Enamine-like chemistry that cannot help.

Candidate density (the other half of the ceiling). Using COCONUT IK14 exact masses: for a random COCONUT molecule the number of competing IK14s is median 27 at ±5 ppm (Q1 8, Q3 83, p90 193, p99 689, max 1,761), 22 at ±1 mDa, 35 at ±10 ppm, 52 at ±20 ppm. For the 250 NP test proxies specifically: median 37 at ±5 ppm (Q3 90, p90 144, max 385). Tightening from ±20 ppm to ±1 mDa cuts the median pool by 58% — which is exactly why the public write-ups report that a ~1 mDa / 10 ppm window is critical and that PubChem-scale pools hurt. If the formula is solved exactly, the pool is the formula bucket: median 21 IK14s, Q3 72, p90 186, p99 689 (molecule-weighted).

Forced analogy: music theory — "the repertoire"

The analogy, stated precisely. A candidate database is a sheet-music library. The hidden test is a concert of 400 pieces, many never engraved. A transposition — the same piece shifted into another key — is the musical operation that preserves interval structure (the scaffold) while moving every absolute pitch (the substituent masses). In chemistry the corresponding move is a single biosynthetic tailoring step: O-methylation (+CH₂), hydroxylation (+O), acetylation (+C₂H₂O), prenylation (+C₅H₈), glycosylation (+C₆H₁₀O₅ / +C₅H₈O₄), desaturation (−H₂). The question: if the library lacks the exact piece, can it supply a transposition that the ranker can still score?

Derivation, tested on real data. Two measurements:

1. Is NP space actually transposition-dense? For every COCONUT InChIKey14, I asked whether another COCONUT formula exists at exactly one tailoring delta away. Results (fraction of all 479,690 IK14s with such a relative): +CH₂ 84.1%, −H₂ 80.2%, +O 79.8%, +C₂H₂O 78.0%, +CH₂O 77.7%, +C₅H₈ 68.9%, +C₃H₂O₃ (malonyl) 58.0%, +C₅H₈O₄ 46.6%, +C₆H₁₀O₅ 33.7%, +C₆H₈O₆ 25.5%, +SO₃ 4.1%. Any of the eleven: 95.7%. Natural-product space is, in this sense, almost entirely composed of transpositions of other entries.
2. Does it work for the test chemistry? For the 250 host-measured NP molecules, 247/250 (98.8%) have at least one tailoring-delta relative present in COCONUT (+O 237, −O 234, +CH₂ 233, +C₂H₂O 229, +H₂ 229, +C₆H₁₀O₅ 200, −C₆H₁₀O₅ 93).

The concrete, testable idea. For each query, after the ±5 ppm COCONUT lookup, also retrieve COCONUT molecules at mass − Δ for Δ ∈ {CH₂, O, CH₂O, C₂H₂O, C₅H₈, C₆H₁₀O₅, C₅H₈O₄, C₃H₂O₃, H₂} (and mass + Δ for losses), then enumerate the corresponding single-site variants of each retrieved neighbour with RDKit (O-methylate/acetylate/glycosylate each free OH; dehydrogenate each saturated ring bond) and append them to the pool, de-duplicated by InChIKey14 and filtered back to the ppm window.

How many hits could it recover? Honest arithmetic. If ~50% of the hidden set is already in COCONUT, the 200 missing molecules are dominated by hypothesised NPs and analogues, which are by construction one or two tailoring steps from something known. If the derivative-relative rate for missing molecules matches the 98.8% measured on present ones, the enumeration reaches most of them — but reaching is not ranking. A molecule with 6 free hydroxyls yields 6 O-methyl regio-isomers; across ~9 deltas and a median of ~27 in-window parents, the pool per query inflates from ~27–37 to roughly 500–3,000 structures, an order-of-magnitude dilution. Only one of those is right, and all its regio-isomers have identical formulas and very similar fragment sets. Expected recovery: +10 to +25 molecules of 400 reach the top-25, i.e. ΔMRR@25 ≈ +0.01 to +0.04, and only if (a) the expansion is applied selectively (when the in-window COCONUT pool is small or the fingerprint model's best in-pool similarity is low) and (b) the ranker can use neutral-loss evidence to break regio-isomer ties. Applied indiscriminately, it will cost score, for the same reason PubChem-scale pools cost score.

Where the analogy holds, partially holds, and breaks.

• Holds: the repertoire really is mostly transpositions (95.7% / 98.8% measured), and the "engraved library" metaphor correctly predicts that a 2D-connectivity metric is key-blind in one respect and key-sensitive in another.
• Partially holds: the metric ignores stereochemistry — musically, it scores the piece regardless of which instrument plays it. That is a real gift: 134,616 of COCONUT's IK14s have ≥2 stereo variants, and all of them collapse to one answer.
• Breaks, hard, on regiochemistry. Transposition in music is a single operation per key; O-methylation is one operation per hydroxyl, and the resulting pieces are not merely similar — they are isobaric and near-isomeric in MS/MS. Music has no equivalent of "six different transpositions that all sound identical to the judge". That degeneracy is precisely what MRR@25 punishes.
• Breaks again on the prior. A concert programme is curated by taste; the hidden 400 was curated to be absent from public spectral libraries. The analogy would need a conductor deliberately choosing pieces the library does not own — which is exactly what Enveda did.

Gotchas / what could fool us

1. "695,133" and "479,690" are not the same number and the metric only knows the second. COCONUT 2.0 counts stereoisomers as distinct molecules ("Since different stereoisomers of a molecule can have distinct biological activities, they are treated as different molecules"), whereas COCONUT 1.0 deduplicated on stereo-free InChIKeys (411,621 in 2020). A pool size printed as 739k is 1.54× the number of answers it can actually produce. Deduplicate on IK14 before measuring recall, or every local recall number is inflated.
2. COCONUT's chemical filter is a hidden exclusion rule. Fluorinated compounds, sulfonamides, "drug derivatives" and multi-component entries are deleted at import. Any hidden-test molecule in the "synthetic plausibly in nature" bucket that carries F or a sulfonamide is structurally unreachable from COCONUT — not rare, not unlucky, impossible. Keep a small F/sulfonamide-bearing side pool from the training structures.
3. The 0.03% enveda-180 ↔ COCONUT overlap is partly definitional. 23.0% of enveda-180 contains fluorine and 28.3% sulfur; COCONUT excludes exactly that chemistry. Reading 0.03% as "Enamine space is chemically alien to nature" overstates the case; some of it is a database policy.
4. The whole training set is only 9.47% COCONUT-covered. Validating a retrieval pipeline on random held-out training structures measures performance on Enamine chemistry. One public write-up reports that random held-out training structures "predicted the opposite ordering of three rankers and were wrong on the board", while the 250 enveda-np-examples tracked the leaderboard. Validate on those 250, with the structure removed from the pool.
5. Terpenoid formula buckets are a wall. C20H30O3 1,283 IK14s, C15H20O4 1,186, C15H22O3 1,168, C15H24 1,001. If the hidden molecule is a sesquiterpene or diterpene, perfect formula determination still leaves ~1,000 candidates and MRR@25 depends entirely on fragment-level discrimination among near-identical skeletons. Expect the score to be bimodal by compound class.
6. Lipids are worse and are a different failure. C54H104O6 (731), C53H102O6 (727), C55H106O6 (715) … triacylglycerol formulas carry 650–730 IK14s each, and MS/MS barely distinguishes acyl positional isomers.
7. C6H5N3O with 1,761 entries at 10 heavy atoms is almost certainly enumerated, not isolated. Stereo-free, largely nameless. If a hidden molecule happens to sit near m/z 135.04, the pool is poisoned. Consider a filter on "formula bucket size ≫ expected for heavy-atom count".
8. Annotation level is not evidence quality in the way it looks. Level 0 exists but is undocumented in the paper (which describes 1–5). 42.8% of records are level ≤1, and 63% have no organism at all. Using annotation_level ≥ 3 as a prior is defensible; using it as a hard filter discards 52.2% of the pool, including genuine but poorly-indexed NPs.
9. NPAtlas is CC BY-NC. It buys you 138 new IK14s for a non-commercial licence obligation on a competition whose winning methods are to be open-sourced. Probably not worth the paperwork.
10. COCONUT provenance is circular. The paper itself flags "circular imports between databases. Many of these aggregate data from other collections and are often, in turn, aggregated". Super Natural II + Supernatural3 + Wikidata NPs + ZINC NP are not four independent votes on whether a structure is natural. The duplicated ANPDB/NANPDB pair proves the point. Do not weight candidates by "number of collections".
11. LOTUS's public snapshot is from 2021-12-20. Anything the hidden set draws from 2022–2026 literature is in Wikidata but may not be in any frozen file you ship. COCONUT's Oct-2026 release is the freshest shippable aggregate.
12. The novelty-rate literature is genuinely contested (Pye et al. vs Skinnider & Magarvey vs the Pye reply). Do not build a coverage estimate on "NP novelty is declining" as if it were settled; the measured, uncontested number is the 49% intra-paper Tanimoto > 0.9 congener rate.
13. Zenodo and ChemRxiv were unreachable from this host (403 / Cloudflare), and the ChemRxiv DOI printed in Enveda's own README does not resolve. Several secondary figures above (SuperNatural 3.0 = 449,058; NPASS = 96,481; Enveda-180 Zenodo DOI and CC-BY data licence) rest on single search snippets or third-party listings, not on documents I opened. Marked accordingly.

Implications for the challenge

1. Build the candidate pool from the COCONUT October-2026 lite CSV, deduplicated to InChIKey14 → 479,690 entries. It is CC0, 190 MB zipped, and ships cleanly inside an offline Kaggle notebook. Do not ship NPAtlas (CC BY-NC, +138 IK14) or anything proprietary.
2. Report every recall/coverage metric after IK14 deduplication. Expect a 1.54× drop from the naive record count; a pipeline that "covers 695k structures" actually offers 479,690 distinct answers.
3. Add the 275,810 training structures to the pool (union = 729,389 IK14) but down-weight the Enamine-like 66%. A cheap, strong prior: penalise candidates with ≥1 fluorine or a sulfonamide, since the test chemistry is NP-like (enveda-np-examples: 0.4% F, 3.6% S vs enveda-180: 23.0% F, 28.3% S).
4. Use ±1 mDa (or ±5 ppm, whichever is tighter at that mass) — not ±20 ppm. Median in-window COCONUT candidates: 22 at ±1 mDa vs 52 at ±20 ppm. Halving the pool roughly doubles the chance of landing inside top-25 at fixed ranker quality.
5. Validate exclusively on the 250 enveda-np-examples structures with each structure removed from the candidate pool. 99.6% of them are in COCONUT and 100% are timsTOF, so they are the only split in the released data that matches the hidden distribution on both chemistry and instrument.
6. Expect a bimodal score by compound class and instrument it. Log, per query, the in-window pool size. Queries with a pool of ≤25 are nearly free MRR; queries in the 600–1,800 terpenoid/lipid buckets are where the model has to earn it. Optimise the ranker on the large-pool subset.
7. Implement tailoring-delta candidate expansion, but gated. Trigger it only when the in-window IK14 pool is below some threshold (say <15) or the best in-pool spectral/fingerprint score is low. Deltas ranked by measured yield: +CH₂ (84.1%), −H₂ (80.2%), +O (79.8%), +C₂H₂O (78.0%), +CH₂O (77.7%), +C₅H₈ (68.9%). Expected gain ΔMRR@25 ≈ +0.01…+0.04; expected loss if ungated, larger than that.
8. Exploit the stereo-blind metric explicitly. 134,616 COCONUT IK14s have ≥2 stereo variants (max 72). Collapse them to one candidate and spend no model capacity on stereochemistry.
9. Build a formula-bucket-size feature and a contamination filter. Flag formula buckets whose IK14 count is wildly out of line with their heavy-atom count (the C6H5N3O/1,761 pattern) and demote those candidates; they are enumerated structures, not isolated natural products.
10. Use annotation_level and organism-presence as soft ranking priors, never as filters. 37.0% organism-linked, 32.7% DOI-linked, 7.2% five-star. A molecule that is 5★ and organism-linked and sole-sourced from NPAtlas/LOTUS is a much better prior than one from an un-sourced bulk import — but filtering on it deletes half the pool.
11. Treat the coverage ceiling as the project's main lever. My estimate is 45–55% of the hidden 400 reachable in COCONUT at IK14; the brief's back-solved recall@25 ≈ 0.55 agrees. A perfect ranker over an unexpanded COCONUT pool tops out near MRR ≈ 0.5. Getting past ~0.47 means expanding the pool intelligently (tailoring deltas, biosynthetic generators) or de-novo generation — not a better fingerprint head.
12. Do not quote Pye et al. as evidence that NP novelty is drying up in any write-up; cite the measured congener statistic (49% of 2015's 1,576 compounds had a same-paper relative at Tc > 0.9) and note the Skinnider–Magarvey null-model objection.

Open questions

1. What is LOTUS's current size at InChIKey14, and what does it add to COCONUT? Zenodo was 403 and the Wikidata SPARQL COUNT timed out. A single successful Wikidata dump or Zenodo fetch settles it; my proxy (the Wikidata Natural Products collection, 146,067 IK14 inside COCONUT) is a lower bound.
2. Is the Enveda-180 ChemRxiv preprint real and what does it report? The DOI in Enveda's own README 404s. The preprint would give the exact compound count, CE scheme, CCS coverage and the Enamine sublibrary identity.
3. Did Enveda build the hidden test from Enamine-adjacent "NP-like" sublibraries or from isolated natural material? enveda-np-examples (0.4% F) says isolated material; the host's "synthetics plausibly in nature" category says otherwise for part of it.
4. What fraction of the hidden set would a biosynthetically generated pool (e.g. enumerated glycosides/methyl ethers of COCONUT, or BGC-predicted structures) reach that COCONUT alone does not? Measurable offline against enveda-np-examples with each structure deleted from COCONUT.
5. Published Enamine REAL / ChEMBL descriptor tables. I used enveda-180 as a proxy; a primary citation would firm up §5.
6. What is COCONUT's level-0 annotation score? Undocumented in the paper, 76,679 records.

Sources

Opened and read directly:

• COCONUT 2.0, Nucleic Acids Research 53:D634–D643 (2025), DOI 10.1093/nar/gkae1063 — full text via Europe PMC PMC11701633 (counts, curation pipeline, fluorine/sulfonamide filter, annotation score, 63 collections).
• COCONUT download page and data files: https://coconut.naturalproducts.net/download ; https://coconut.s3.uni-jena.de/prod/downloads/2026-10/coconut_csv_lite-10-2026.zip and coconut_csv-10-2026.zip (downloaded 2026-10-07; all recomputed COCONUT statistics in this document come from these two files).
• Sorokina & Steinbeck, "Review on natural products databases: where to find data in 2020", J. Cheminform. 12:20, DOI 10.1186/s13321-020-00424-9 — full text via Europe PMC PMC7118820 (411,621 COCONUT 1.0 molecules on stereo-free InChIKeys; 123 resources, 92 open, 50 downloadable; 40/50 overlap ≥50% with another DB).
• Rutz et al., "The LOTUS initiative…", eLife 11:e70780 (2022), DOI 10.7554/eLife.70780 — via PMC PMC9135406 (750k+ pairs, 290k+ structures, 40k+ organisms, CC0, Zenodo 10.5281/zenodo.5794106, DNP v29.2 comparison, KNApSAcK/NAPRALERT figures).
• NPAtlas download page and NPAtlas_download.tsv (v2024_09): https://www.npatlas.org/download (licence CC BY-NC 4.0; file downloaded and recomputed 2026-10-07). Zenodo record page https://zenodo.org/record/3530792 (v2024_09, 2024-09-13, file list, licence).
• Pye, Bertin, Lokey, Gerwick & Linington, PNAS 114:5601–5606 (2017), DOI 10.1073/pnas.1614680114 — full text via PMC PMC5465889 (52,395 unique molecules 1941–2015, AntiMarin + 29,062 curated articles, ~1,600/yr, median max Tc plateau ~0.65, 6,414 clusters / 40,229 compounds / 12,166 singletons / 76.8% / >23%).
• Skinnider & Magarvey, PNAS 114:E6271–E6272 (2017), DOI 10.1073/pnas.1708560114 — PMC PMC5547653 (32,380-NP in-house DB; permutation and ZINC null models; KS p = 6.2 × 10⁻¹⁴ and 7.6 × 10⁻¹¹).
• Pye et al. reply, PNAS 114:E6273 (2017), DOI 10.1073/pnas.1711139114 — PMC PMC5547657 (2015: 1,576 compounds from 484 papers; 49% with a same-article relative at Tc > 0.9).
• Enveda's own repository: https://github.com/enveda/enveda-180 — README (Enamine REAL provenance, 14 enamine batches, Bruker .d, adduct lists, author list, MIT code licence) and GitHub API metadata (created 2026-05-19).
• Kaggle write-ups mirrored on GitHub (Kaggle pages are JS shells; these are the mirrors I actually read): corentinlaudicina/enveda-casmi-2026 README ("adds 480k natural products"; "enveda-180 is not natural-product-like (0.03% of its molecules are in COCONUT, vs 99.6% of enveda-np-examples)"; LB figures) and shiferaxa/casmi26 README ("about 712k structures: COCONUT plus every structure in the training set"; validation-split finding).
• train.parquet (2,539,608 rows) from the competition data already staged at C:\Temp\casmi26\data\ — all per-ingest_lib spectra/structure counts, instrument/polarity splits, overlap percentages and formula-derived descriptors in §3–§5 were computed by me from this file on 2026-10-07.

Consulted but not verifiable from this host (marked low confidence where used): Zenodo records 10.5281/zenodo.20436851 and 10.5281/zenodo.21346580 (HTTP 403); ChemRxiv 10.26434/chemrxiv.15004319 (404 at doi.org, absent from Crossref); SuperNatural 3.0 (449,058) and NPASS (96,481) counts, from search snippets only; Ertl JCIM 2008 NP-likeness and Ertl CS3-2018 slides (PDFs returned unparsed binary); C&EN 2026-09-17 article (HTTP 406).