Enveda's public chemistry reconstructed: platform, isolation, analog patents, what 'hypothesised NPs' and 'synthetics plausibly in nature' mean, and a confidence-labelled model of the hidden 400

TL;DR. Enveda's public record (14 MS/cheminformatics papers and preprints, 24 patent-family publications, blog posts and press releases opened here) describes a company that LC-MS/MS-profiles plant extracts on Bruker timsTOF Pro / Pro 2 instruments, isolates actives with a 2D-LC mass-targeted fraction collector, and then runs classical medicinal chemistry on natural-product scaffolds (withaferin A, isoandrographolide, nootkatone), adding halogens, heteroaryls, fluoromethyl and amine groups. "Hypothesised natural products" in the hidden test are, on this evidence, structures that Enveda's own de novo / fingerprint tools (MS2Mol, CSI:FingerID, NMR networking) proposed for extract features and that were then confirmed or synthesised, while "synthetics plausibly in nature" are most likely such scaffold analogs and metabolite-mimetics. The consequence for the challenge: roughly half of the hidden 400 should be reachable in COCONUT, the med-chem-decorated slice is unreachable in COCONUT by database policy (fluorine and sulfonamide filters) and only partly reachable in PubChem (patent-derived structures), and the dominant miss should be "right scaffold, wrong substituent", which argues for shift-tolerant analog retrieval plus single-step enumeration.

1. Enveda's publications and preprints on the platform

Opened via Europe PMC, OpenAlex (166 works with an "Enveda" affiliation string; 49 when restricted to spectra/metabolomics/natural-product terms), arXiv export API and the papers themselves. Only MS/cheminformatics items are listed; the knowledge-graph and clinical papers are omitted.

Not found anywhere: a "diffusion models for MS" paper, a "spectral library of plants" paper, or any paper describing the isolation platform's throughput. The ENV-956 code in the brief does not appear on Enveda's pipeline page (programs listed: ENV-294, ENV-308, ENV-6946, plus "17 development candidates / 6 IND-enabling assets"); treat ENV-956 as unverified.

2. Patents (Google Patents, assignee "Enveda", 24 results across 3 pages, read 2026-10-07)

Negative result: no Enveda patent claims a "natural product analog library", an "in silico metabolome", biotransformation-rule enumeration, or a generative "hypothesised natural product" method. The analog chemistry that is public is conventional, scaffold-by-scaffold med-chem on three terpenoid natural products.

3. What the site, blog, press and job boards say about scale and instruments

Reading the three library numbers together (inference, confidence medium): ~90–200 k is the set of characterised compounds in Enveda's extract-derived screening library; 1.5 M+ is the set of features with predicted structures that PRISM/MS2Mol make "searchable". No source gives a pure-compound (isolated, weighed) library size, and no job ad mentions isolation chemists — isolation appears to be in-house hardware (the fractionation patent) rather than a large wet-chemistry team, with med-chem scaled in Hyderabad. No evidence for timsTOF HT or fleX; "timsTOF" in the test schema is consistent with Pro/Pro 2.

4. Decoding the four hidden-test categories

The competition wording (BRIEF.md; Kaggle page itself is a JS shell): "natural products, hypothesised NPs, analogs, synthetics plausibly in nature; none previously published". Business Wire (via Yahoo Finance, 2026-09-15) and the Kaggle mirror say the molecules are "real or potential natural molecules and their analogs", generated "on Enveda's isolation and profiling platforms", "about 400 molecules … about 2,500 spectra", monoisotopic mass "157 to 1,159 Da, median about 348 Da", 1–16 spectra per molecule (median 3), all Bruker timsTOF.

Key implication of "none previously published": it is a statement about spectra and, for most categories, structures. Only the "natural products" slice can have a prior structure record (COCONUT/LOTUS/PubChem) without a prior spectrum.

5. What enveda-180 and enveda-np-examples are

• enveda-180 = Enamine HLL-200 (Hit Locator Library, 200,000 compounds) run pooled on a timsTOF Pro 2, PASEF, nominal 20/40/60 eV, both polarities; 184,330 compounds with curated spectra, 1,158,293 spectra, 14 surviving adducts of 17 targeted (README read here: "drug-like compounds from the Enamine REAL space", 8 positive + 9 negative adducts; Zenodo detail per the q2 sibling file, which I did not re-open). Why Enveda bought it: their own blog benchmarks the in-house library against Enamine HLL-460, so HLL-type sets are their reference "synthetic" chemistry. Confidence high. The 0.03 % COCONUT overlap is partly definitional — 23 % of enveda-180 carries fluorine and COCONUT deletes fluorinated and sulfonamide entries at import (sibling q3 worker, COCONUT 2.0 paper).
• enveda-np-examples (1,184 timsTOF spectra, 250–251 structures, 99.6 % in COCONUT by InChIKey14 per one write-up, and all with spectra in other public libraries per another) is therefore almost certainly a purchased natural-product standard set run on the same timsTOF method, not isolated material. The best public twin is TIMS-Bench's "Plant natural products dataset" (MassIVE MSV000101619, 2026), released by the same group (Misra, Domingo-Fernández, Dorrestein) as ground truth for feature-finding benchmarks. Confidence medium (MassIVE page unreachable; count of compounds in MSV000101619 not confirmed).
• Neither slice is drawn from the hidden-test population: enveda-180 is the wrong chemistry, np-examples is the right chemistry but the wrong novelty regime (every structure is public).

6. A confidence-labelled model of the hidden 400

Assumptions stated so they can be attacked. Shares are priors informed by the category wording, the median mass of 348 Da (too light for a glycoside-dominated set; too heavy for a fragment-like synthetic set), the 1,159 Da ceiling (saponins/glycosylated terpenoids or peptides exist), and Enveda's disclosed scaffold preferences (withanolides, labdanes, sesquiterpenes, flavonoids from Acanthaceae/Solanaceae/Rutaceae-type material).

This reproduces, from the chemistry side, the back-solved public number (recall@25 ≈ 0.55 at MRR 0.417) and the sibling estimate (45–55 % in COCONUT). Two refinements it adds: (i) the slice that is structurally in PubChem but not in COCONUT is small but non-zero and consists of patent-type analogs — a PubChem search restricted to candidates with NP-likeness > 0 and a Murcko scaffold shared with a COCONUT entry would capture it without the dilution of the full 90 M; (ii) the "nowhere public" slice is dominated by one-step congeners and decorated scaffolds, which is exactly what scaffold-anchored enumeration reaches.

Instrument fleet implication: the timsTOF Pro / Pro 2 are 4500 V heated-ESI Q-TOFs with PASEF; the hidden spectra are described as merged multi-CE rows (lists in collision_energy_ev), consistent with enveda-180's 20/40/60 eV scheme — so enveda-180 is the right instrument exemplar even though it is the wrong chemistry.

Forced analogy: music theory — "variations on a theme"

Statement. A scaffold is the theme; each analog is a variation. A transposition preserves interval structure and changes every absolute pitch: in MS/MS terms, substituent changes that keep the fragmentation pathway (bond cleavages, charge site) but shift fragment masses by a constant Δ on one side of the cleavage. Timbre is the intensity envelope across fragments — which pathways dominate at a given CE. The question is whether a model trained on the parent recognises a transposed variation, and whether Enveda's library design makes "scaffold-correct, substituent-wrong" the dominant failure.

Which variations keep timbre (prediction, from fragmentation chemistry):

Evidence the analogy pays in practice. One public CASMI-26 write-up (mirrored locally; the GitHub page was not opened) reports its largest single jump — MRR 0.145 → 0.335 on the public board — from "Reference library matching & analog propagation": adduct-shifted modified cosine against training spectra with N_ANALOG = 200 timsTOF representatives, i.e. shift-tolerant matching of hidden spectra to variations already in train. The same write-up measured the best-analog Tanimoto of correctly-vs-incorrectly ranked molecules at 0.753 vs 0.689 and concluded that the hidden test's difficulty "comes from natural-product chemistry with sparse analogs, not from acquisition conditions". That is the analogy stated as a measurement: when a close variation exists in the library, the model hears the theme.

Testable idea for MRR@25 (concrete). Build a transposition table and a timbre test:

1. From train, extract all pairs of structures whose formulas differ by exactly one of {CH2, O, C2H2O, C6H10O5, C5H8O4, C6H10O4, HF−H, Cl−H, Br−H} and whose Morgan-r2 Tanimoto ≥ 0.6 (same scaffold). Expect thousands of pairs from GNPS/MassBank/RIKEN plant chemistry and tens of thousands inside enveda-180's Enamine R-group siblings.
2. For each pair, compute modified cosine with and without the Δ shift at matched CE. The gap (shifted − plain) per Δ is the empirical "timbre retention" curve; the hypothesis predicts retention ≥ 0.6 for +CH2/+O/+hexose and < 0.3 for aryl/heteroaryl/NMe2 variations.
3. Deploy the result as a ranking feature: for a hidden spectrum, retrieve library spectra at precursor ± Δ for the high-retention Δs only, score by shifted cosine, and propagate the matched structure's scaffold as a prior over COCONUT candidates (and as the seed for one-step enumeration, deduplicated by RDKit-2026.03.3 tautomer-canonical InChIKey14).
4. Validation metric: on enveda-np-examples with the structure deleted from the pool, measure the fraction of misses whose top-1 shares the Murcko scaffold with the truth. If that fraction exceeds ~50 %, "scaffold-correct, substituent-wrong" is the dominant failure mode and the expansion is worth the dilution; if < 25 %, the misses are scaffold errors and the effort belongs in the fingerprint head instead. Expected size of the win if the premise holds: +0.02–0.05 MRR@25 (10–25 molecules of 400 moved into the top-25), consistent with the sibling's tailoring-delta estimate.

Where the analogy holds, partially holds, breaks.

• Holds: biosynthetic tailoring (methyl, hydroxyl, acetyl, sugar) are true transpositions — the fragment ions of the core are reused, and the shift-tolerant matching gain on the public board is direct evidence.
• Partially holds: stereochemistry is "the same piece by a different performer" and the metric ignores it, so stereo-variants cost nothing — a gift the analogy predicts.
• Breaks on regio-isomers: a theme with six hydroxyls has six O-methyl "transpositions" that are isobaric and near-isospectral. Music has one transposition per key; chemistry has many per Δ and the judge (InChIKey14) rewards only one. This is the degeneracy that MRR@25 punishes, and it is why ungated enumeration costs score.
• Breaks on med-chem analogs: heteroaryl, NMe2, CH2F and nitrile variations change the protonation site and therefore the pathway — new timbre, not a transposition. For this slice the analogy says "do not expect the parent's spectrum to find the child"; a forward model (GrAFF-MS-class, ICEBERG) scoring enumerated children is the only route, and COCONUT cannot hold the answer at all if it carries fluorine.
• Breaks on the prior: Enveda chose a programme the library does not own ("none previously published"); a concert of unknown pieces is still recognisable by theme only if the themes are in the repertoire — the 35–45 % "nowhere" slice tests exactly that.

Gotchas / what could fool us

1. "1.5 M+ natural product compounds" is not a pure-compound library. Three incompatible numbers (90 k, ~200 k, 1.5 M) appear in Enveda's own text within 15 months; the largest is almost certainly features-with-predicted-structures. Do not infer that 1.5 M isolated NPs exist and might be in the test.
2. COCONUT's import filter makes fluorinated/sulfonamide analogs unreachable by policy, not by chance. Enveda's analog patents put F at C-4, on side-chain CH2F, and as aryl-F; a pool without a small F-bearing side set cannot score those molecules at all.
3. Patent-exemplified analogs may be in PubChem. If Enveda's test analogs are the patent compounds, a scaffold-restricted PubChem lookup could recover them; if they are unpublished SAR compounds, nothing will. Neither is verifiable; budget for both.
4. MS2Mol's EnvedaDark (226 NPs "absent from major databases") is the host's own template for a benchmark. Expect the hidden set to be built with the same mindset — a deliberately out-of-database tail — and expect the host's own de novo model to score ~21 % close-match, not exact matches, on it.
5. enveda-np-examples is a replicate test, not a novelty test. All its structures have public spectra; leave-one-spectrum-out gives 0.93, structure-removed gives ~0.05 (sibling q2). It is the right instrument/chemistry proxy only when the structure is deleted from the pool.
6. Halogen isotope features learned on enveda-180 (36 % halogenated in the decoy) will fire on the analog slice but not on the NP slice. They are not dead weight here (unlike the sibling's assumption) — but they will be miscalibrated: Cl/Br in Enamine amides vs I/F on terpenoid scaffolds.
7. Dimer adducts (28.5 % of enveda-180) do not occur in the decoy's 7 adducts; a model that learned [2M+Na]+ behaviour wastes capacity.
8. Multi-CE merged rows. Hidden rows apparently merge several CEs (2,500 acquisitions → ~1,500 rows); shift-tolerant cosine against single-CE library rows must be done per CE or against merged 20/40/60 eV triplets built from enveda-180.
9. The PRISM "paper" does not exist. All PRISM numbers come from one blog post and press releases; there is no architecture, tokeniser or evaluation detail to reproduce, and the "600 M internal spectra" are unavailable.
10. The Enveda-180 preprint DOI (chemrxiv.15004319) does not resolve and MS2Mol's PDF is behind a 403; thresholds behind "close match"/"meaningful similarity" remain unverified.
11. ENV-294's scaffold is inferred, not stated. The chain is: ENV-294 = RAC2 modulator (press release) → Rac2 patent inventor Jadhav → same inventor on the isoandrographolide family → Enveda's Acanthaceae/andrographolide paper. Plausible, unproven; a test-set analog could equally be withanolide- or nootkatone-derived.
12. Mass range 157–1,159 Da with median 348 Da says the set is not glycoside-dominated; models that over-weight hexose/rutinoside losses (a SHAP finding in one write-up) will overfit the np-examples proxy.
13. TIMS-Bench's licence is CC BY-NC-ND; its ground-truth data may not be redistributable inside a winning open-source method.

Implications for the challenge

1. Treat the hidden set as four sub-populations and score each on validation. Build a held-out set that mirrors §6: (a) 250 np-examples with structures deleted from the pool; (b) ~300 public-library plant NPs with one-step congeners present in the pool but the molecule absent; (c) ~200 enveda-180 molecules whose nearest train neighbour has Tanimoto ≥ 0.7 (the "analog present" regime); (d) ~100 F/heteroaryl-bearing NP-like molecules from train (the "unreachable in COCONUT" regime). Report MRR@25 per block.
2. Implement shift-tolerant analog propagation first (adduct- and Δ-shifted modified cosine against all timsTOF train spectra, then scaffold-prior over COCONUT); it produced the largest public gain reported (0.145 → 0.335 in one write-up) and is cheap.
3. Run the timbre-retention measurement (analogy §) once on train pairs and publish the per-Δ retention table in research/q3_natural_product_chemistry/; gate single-step enumeration on it.
4. Add a small non-COCONUT side pool: train structures with NP_Score > 0 plus enumerated F/Cl/Br/heteroaryl children of the top-k COCONUT scaffolds retrieved per query; score them with a forward model, not with fingerprint cosine alone.
5. Scaffold-restricted PubChem lookup instead of full PubChem: candidates in the ppm window that share a Murcko scaffold with any COCONUT entry or with the top analog hit. This targets patent-type analogs without the 90 M-pool dilution that hurt public entrants.
6. Use enveda-180 for instrument priors and R-group-sibling training pairs, not for chemistry: it is the HLL-200 Enamine set on the test instrument, and its siblings are the ideal "variation" training data for step 3.
7. Expect bimodal scores by class: flavonoid/alkaloid NPs (small formula buckets, catalogued) near MRR 0.6–0.8; terpenoid/withanolide analogs (large isomeric buckets, decorated, uncatalogued) near 0–0.1. Optimise the ranker on the second group.
8. Watch for TIMS-Bench's MSV000101619 as an external timsTOF plant-NP ground-truth set; if its compound list is public, compute its overlap with np-examples and with the hidden set's mass histogram.

Open questions

• Does MSV000101619 (TIMS-Bench plant NP dataset) contain the same 250 compounds as enveda-np-examples, and what vendor set are they (MassIVE page unreachable this session)?
• What are the Tanimoto thresholds behind MS2Mol's "close match" and "meaningful similarity" (PDF 403)?
• Are the hidden-test "analogs" patent-exemplified compounds (then partially PubChem-reachable) or unpublished SAR compounds?
• Which scaffold is ENV-294 — labdane lactone (inferred), withanolide, or other?
• Does Enveda's TIMS-Bench / dereplication work define "hypothesised NP" as an MS2Mol-proposed structure, an NMR-networking proposal, or a synthesised confirmation?
• What is the hidden set's class mix by NPClassifier — computable only after the leaderboard reveals structures, but the mass histogram (median 348 Da) could be compared now against COCONUT class-wise mass distributions.

Sources

Opened and read on 2026-10-07 (web-search budget was exhausted after one query; everything below came from direct fetches or local files):

• Enveda PRISM blog: https://enveda.com/prism-a-foundation-model-for-lifes-chemistry/
• Enveda platform page: https://enveda.com/platform/ ; pipeline: https://enveda.com/pipeline/ ; careers: https://enveda.com/careers/ ; news index: https://enveda.com/news/ ; sitemaps: https://enveda.com/sitemap.xml , post-sitemap.xml, news-sitemap.xml
• Enveda blog posts: "The chemical code opportunity" (Colluru, 2023-09-28) https://enveda.com/the-chemical-code-opportunity/ ; "Breaking barriers: simultaneous target and drug discovery at scale" (Gordon, 2025-01-30) https://enveda.com/breaking-barriers-simultaneous-target-and-drug-discovery-at-scale/ ; "Hacking the unknown" (2025-01-30) https://enveda.com/hacking-the-unknown/
• Enveda press releases: MS2Mol launch 2023-06-21 https://enveda.com/news/enveda-biosciences-unveils-new-ai-algorithm-to-unlock-natures-chemical-code/ ; ENV-294 IND 2024-11-13 https://enveda.com/news/enveda-announces-first-candidate-discovered-from-nature-using-ai-advances-to-clinical-trials-2/ ; ENV-6946 IND 2025-12-23 https://enveda.com/news/enveda-advances-third-asset-to-the-clinic-with-us-fda-ind-clearance-and-phase-1-initiation-of-env-6946/ ; ENV-294 late-breaker 2026-09-29 https://enveda.com/news/enveda-announces-late-breaker-oral-presentation-of-new-mechanistic-and-biomarker-data-from-the-phase-1b-trial-of-env-294-a-potential-first-in-class-oral-therapy-for-moderate-to-severe-atopic-dermatit/ ; Series E 2026-09-23 https://enveda.com/news/enveda-raises-311-million/
• Business Wire CASMI announcement via Yahoo Finance: https://finance.yahoo.com/healthcare/articles/enveda-launching-next-alphafold-contest-120700676.html
• TechCrunch 2024-06-14: https://techcrunch.com/2024/06/14/enveda-raises-55m-to-combine-ancient-remedies-with-ai-for-drug-discovery/ ; BioSpace 2023-02-14: https://www.biospace.com/enveda-s-matured-platform-drives-toward-three-inds-in-2023 ; STAT 2026-08-18 (ENV-308, lac-phe; paywalled beyond lede): https://www.statnews.com/2026/08/18/enveda-pill-mimics-exercise-phase-one-trial-leptin-weight-loss-maintenance/
• Google Patents query (assignee Enveda, 24 results): https://patents.google.com/xhr/query?url=assignee%3DEnveda ; patents read: US20250189496A1, WO2026039292A1, WO2026024676A1, WO2025076137A1, WO2024239001A2, JOP20250284A1, WO2024243089A3 (all at https://patents.google.com/patent//en)
• Europe PMC affiliation query: https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=AFF:%22Enveda%22 ; OpenAlex: https://api.openalex.org/works?filter=raw_affiliation_strings.search:Enveda (166 works; 49 with spectra/metabolomics terms) and the MS2Mol record https://api.openalex.org/works/https://doi.org/10.26434/chemrxiv-2023-vsmpx
• MS2Mol ChemRxiv DOI 10.26434/chemrxiv-2023-vsmpx (PDF 403; abstract from OpenAlex); GrAFF-MS arXiv 2301.11419 (export API); PRISM: no arXiv/Europe PMC record (queries returned 0)
• Engler Hart et al., GigaScience 2025, 10.1093/gigascience/giaf033 (PMC11970369); Domingo-Fernández et al., J Cheminform 2023, 10.1186/s13321-023-00778-w (PMC10636812); Engler Hart et al., JASMS 2024, 10.1021/jasms.3c00353 (Europe PMC full text XML PMC10854760); Engler Hart et al., J Cheminform 2024, 10.1186/s13321-024-00899-w (PMC11363525 XML); Stienstra et al., Commun Chem 2025, 10.1038/s42004-025-01839-x (PMC12816106); Domingo-Fernández et al., Anal Chem 2026, 10.1021/acs.analchem.6c00361 (PMC13373924)
• TIMS-Bench bioRxiv API record: https://api.biorxiv.org/details/biorxiv/10.64898/2026.05.23.724673 (full text blocked by Cloudflare 1015/429); MassIVE MSV000101619 (DOI 10.25345/c56t0h959) listed by OpenAlex, page not reachable
• GitHub: https://api.github.com/orgs/enveda/repos ; https://raw.githubusercontent.com/enveda/enveda-180/main/README.md ; enveda/acanthaceae-chemotaxonomy README; enveda/guild README; enveda/modbioterp-enveda README (unrelated protein-ML workshop)
• Lever job board: https://api.lever.co/v0/postings/enveda?mode=json
• Kaggle mirror: https://raw.githubusercontent.com/leesirleesirlee-star/Enveda_CASMI-2026_Molecule-ID-From-Mass-Spectra/main/docs/competition_brief.md (Kaggle page itself is a JS shell)
• Local: BRIEF.md; sibling deliverables research/q2_data_forensics/enveda180_and_np_examples_anatomy.md, research/q3_natural_product_chemistry/np_chemical_space_coconut_lotus_npatlas_sizes_overlap.md, research/q5_methods_lineage/denovo_generation_ms2mol_diffms_madgen_spec2mol_msnovelist.md; scratchpad mirrors of public write-ups (gnshx roadmap/method, corentinlaudicina README) grepped for the analog-propagation numbers — the GitHub pages were not opened by me
• Not reachable this session (do not cite from here): ChemRxiv full text, Zenodo API (session limit), ACS J Nat Prod labdane paper (403), Endpoints (403), Reuters (blocked), C&EN 2026-09-17 article (URL unknown)