Skip to content

Medicinal Mushroom Compound × Chokepoint Mapping — Computational Analysis (comp-014)

Status

Phase 1 (Scope) — Committed 2026-05-06. Phases 2–6 are queued, not executed. This page documents the scope and the rationale for the multi-phase plan; it is NOT a results page. Per Open Enzyme convention, this page will be revised in place when each subsequent phase lands, with a status promotion (scoped → in progress → complete) tracked in the YAML frontmatter and the computational-experiments.md tracking index.

Cross-experiment update from comp-018 (2026-05-08): The upstream complement modulator sweep — broader than comp-014 (all compound classes, not just fungal) — surfaced Ganoderma triterpene anticomplement evidence at the C3 convertase node: ganoderic acid Sz IC50 44.6 µM, ergosterol IC50 52 µM, ergosterol peroxide IC50 126.8 µM (Seo HW et al., Arch Pharm Res 2009;32(11):1573-9, PMID 20091270, doi:10.1007/s12272-009-2109-x). SAR insight: Δ7(8), Δ9(11)-lanostadiene type triterpene + 3-keto required for activity. Mechanistic closure of the comp-014 β-glucan structure-dependence finding: the same Ganoderma fruiting body contains lectin-pathway-activating β-glucans AND classical-pathway-INHIBITING triterpenes/sterols. Whether a reishi extract is anti- or pro-inflammatory at the complement level depends on which fractions dominate. Implication: triterpene-enriched ethyl-acetate-soluble fractions (not whole extract) are the gout-favorable preparation. comp-018 also re-confirms the C5aR1 antagonist class is empty (third independent scan — comp-014 + §1.21 + comp-018) but documents that the upstream-CP0 modality at the C3 convertase node is dietary-tier viable via rosmarinic acid (rosemary, lemon balm, spearmint) — the chokepoint-hacker move comp-014 Phase 5 was scoped to find. See upstream-complement-modulator-sweep-computational.md.

§1 Question

Across all known characterized fungal natural products — aggregated from sources that span the global research corpus, not just Western pharma — which compounds map onto Open Enzyme's existing chokepoints, and which fungal species are the highest-leverage producers?

The framing question per chokepoint, per Brian (2026-05-06): "What open biological problem might fungal chemistry answer at this chokepoint that we don't yet have a fermentable lever for?" — not "is fungus useful here?" The inversion is the same load-bearing one applied in modality-chokepoint-matrix.md — guard against path-dependent narrowing.

§2 Why this experiment exists

Three converging reasons:

  1. Path-dependent narrowing risk if we default to ChEMBL-only. comp-013 already surfaced the named gap: 5 of 9 candidate TCM compounds had no ChEMBL bioactivity records of any kind, forcing reliance on animal-model in vivo evidence to assign verdicts. The gap is downstream of which compounds got into Western-pharma curation pipelines, not whether the compounds have evidence. Fungal compounds will hit the same coverage gap or worse — most medicinal-mushroom clinical evidence sits in Chinese (CNKI, Wanfang) and Japanese (J-STAGE) corpora that ChEMBL has never seen. Per Open Enzyme/CLAUDE.md §Global-multilingual research, treating these as "language barrier" is the wrong frame: in 2026 the AI substrate reads them natively.

  2. Fungi are biochemically distinctive in chemistries that adjacent kingdoms don't produce in volume. Ergothioneine biosynthesis is essentially fungal-and-mycobacterial; epipolythiodioxopiperazine (ETP) compounds with internal disulfide bridges are fungal-specific; lanostane triterpenoid diversity (Ganoderma genus alone has 400+ characterized) dwarfs plant triterpenoid diversity at the species level. A breadth-first aggregation that ignores this is biased; one that includes it might surface chokepoint coverage that no single existing OE wiki page would predict.

  3. The "redox/disulfide" angle in Brian's framing points at a chokepoint Open Enzyme has not formally mapped on the patient side. The DAF SCR1-4 disulfide-count incident (2026-05-06, CLAUDE.md §Pre-commit verification gate) shows disulfide chemistry is already load-bearing for engineering. The patient-side mechanism — TXNIP/thioredoxin axis as an NLRP3 priming layer; PDI as ER folding capacity for disulfide-rich therapeutic proteins; ergothioneine as a uniquely-fungal thiol antioxidant with cardiovascular evidence — has not been worked up. comp-014 may produce the evidence to formalize it as a named chokepoint.

§3 Scope (Phase 1)

§3.1 Phase 5 anchor species (18 — sanity-check, NOT the breadth gate)

Reframed 2026-05-06 per Brian's correction: breadth means species-agnostic pull from LOTUS/NPAtlas/KNApSAcK/NPASS/TCMSP filtered by toxicity-filter.json, not a curated medicinal-fungus list. The 18 species below serve two roles: (1) Phase 2 sanity-check — every one of them MUST appear in the Phase 2 output, else the pipeline has a bug; (2) Phase 5 deep-dive seed — they are the well-studied anchor when Phase 4 chokepoint intersection produces hits, since they have established CNKI/J-STAGE clinical-evidence corpora to ingest first.

Selected for substantive characterized chemistry + human-use precedent + regulatory tractability. Exclusions: psychoactive (Schedule I — Psilocybe), toxic (Amanita, Claviceps), culinary-only (truffles). Full rationale per species in experiments/comp-014-medicinal-mushroom-compound-mapping/inputs/candidate-species.json.

Common name Scientific Distinctive chemistry Redox-chokepoint relevance
reishi / lingzhi Ganoderma lucidum 400+ ganoderic acid triterpenoids, β-glucans —
cordyceps (cultivated) Cordyceps militaris cordycepin (3'-deoxyadenosine — purine analog, abundant in militaris) —
caterpillar fungus (wild) Ophiocordyceps sinensis adenosine + cordyceps acid (mannitol); chemistry diverges from C. militaris —
lion's mane Hericium erinaceus erinacines, hericenones (NGF inducers) —
turkey tail / yun zhi Trametes versicolor PSK / PSP (FDA-approved adjuvant in Japan) —
chaga Inonotus obliquus betulinic acid, lanostanes, melanin —
maitake Grifola frondosa D-fraction β-glucan —
shiitake Lentinula edodes lentinan, eritadenine (cholesterol-lowering) —
oyster mushroom Pleurotus ostreatus lovastatin (originally discovered here), ergothioneine 2.4 mg/g DW [Phase 7-1c correction: P. ostreatus is NOT the highest fungal EGT producer — that's P. citrinopileatus at 7.0 mg/g DW; see Phase 7-1c strain scan output and medicinal-mushroom-complement-track.md candidate species table] PRIMARY
golden oyster Pleurotus citrinopileatus ergothioneine — highest known fungal producer (7.0 mg/g DW), apex strain P. citrinopileatus 303 → 641.76 mg/L submerged with H₂O₂+vit C oxidative stimulus per Li 2025; β-glucans [added 2026-05-06 as Phase 7-1c propagation back-fix] PRIMARY
almond mushroom Agaricus blazei β-glucans (agaritine = toxicity flag) —
antrodia Antrodia camphorata antrocins, antroquinonol (Phase II oncology) —
sang hwang Phellinus linteus hispolon —
fu ling / poria Wolfiporia cocos pachymic acid, lanostane triterpenes —
zhu ling Polyporus umbellatus polyporusterones (TCM diuretic) —
tremella Tremella fuciformis acidic heteropolysaccharides —
wood ear Auricularia auricula-judae acidic polysaccharides (anticoagulant) —
koji (chassis) Aspergillus oryzae kojic acid, ergothioneine producer, secreted enzymes secondary
A. terreus Aspergillus terreus lovastatin (industrial origin), terrein —

Phase 2 expansion criteria documented in the JSON — KNApSAcK / TCMSP / MIBiG may surface less-studied species worth adding.

§3.2 Chokepoint targets

18 entries, drawn from canonical wiki sources and joined by UniProt accession. 17 canonical, 1 proposed (Phase 5 confirm/reject).

Canonical chokepoints carried over from existing wiki:

  • Urate axis: uricase substrate (luminal sink), URAT1, GLUT9, ABCG2, OAT1/OAT3, Xanthine Oxidase
  • NLRP3 axis: NLRP3, ASC, Caspase-1, IL-1β, TNFα
  • Complement (CP0): DAF/CD55 (engineering chassis target, not patient target), C5aR1
  • Vessel-wall: Lp-PLA2
  • Transcriptional / regulatory: HDAC6 (per comp-007), PPARγ (per abcg2-modulators.md), Nrf2/KEAP1

Proposed addition (Phase 5 decision): redox/disulfide-modulator chokepoint — candidate human targets PDI (P4HB), PDIA3, Thioredoxin (TXN), TXNIP (Q9H3M7), Glutaredoxin (GLRX). The TXNIP entry is bidirectional with the NLRP3 chokepoint: TXNIP is a direct NLRP3 activator under oxidative stress, so fungal compounds suppressing TXNIP would suppress NLRP3 priming.

Phase 5 verdict tiers (three-way decision, not binary — refined in Phase 1 peer review): - ADMIT: ≥20 fungal compounds across ≥3 mechanistic classes → formal chokepoint addition to modality-chokepoint-matrix.md as a new column. - PRELIMINARY: 10–19 compounds across ≥2 mechanistic classes → provisional column in the matrix with explicit "needs Phase 5b expansion" tag. - REJECT: <10 compounds OR dominated by one mechanistic class → fold surfaced compounds into the existing NLRP3 sub-chokepoint map as an additional priming axis (TXNIP/redox), without elevating to a standalone chokepoint.

One canonical-wiki gap surfaced in Phase 1: OAT4 (SLC22A11, UniProt Q9NS40) is a known apical urate exchanger and benzbromarone target but is not currently in any wiki page. Added to chokepoint-targets.json for comp-014 scope and flagged for closure in a separate sweep — abcg2-modulators.md or gout-pathophysiology.md should incorporate OAT4 as part of the canonical urate-transporter inventory.

§3.3 Data sources

12 compound and bioactivity databases: - LOTUS, NPAtlas, KNApSAcK, NPASS, TCMSP, TCMID, TCM Database@Taiwan, HIT, BATMAN-TCM, COCONUT (compound coverage) - ChEMBL, PubChem BioAssay, SwissTargetPrediction, STITCH (bioactivity / target mapping) - MIBiG, antiSMASH-DB (genomic-encoded compound space — for hypothesis generation about uncharacterized chemistry)

6 multilingual literature corpora: PubMed, CNKI, Wanfang, J-STAGE, CiNii, KISS, RISS. Phase 5 ingestion follows Open Enzyme/CLAUDE.md §Translation protocol — two-model cross-check with inline annotations on disagreement.

Per-language model pairing for Phase 5 (specified now to avoid Phase-5-time drift): - Chinese-language sources (CNKI, Wanfang): Model A = Claude (Anthropic) or Gemini (Google); Model B = DeepSeek or Qwen (Chinese-vendor — native-language depth catches idiomatic and classical-TCM-terminology nuance Western-trained models miss). - Japanese-language sources (J-STAGE, CiNii): Model A = Claude; Model B = Gemini OR another competent Japanese-fluent model. No Japanese-vendor model is currently a clear best-in-class option — two competent Western-trained models is acceptable per CLAUDE.md. - Korean-language sources (KISS, RISS): Model A = Claude; Model B = GPT-5 OR Gemini. No Korean-vendor model is currently in the standard frontier-model set — two competent Western-trained models is acceptable per CLAUDE.md. If a Korean-fluent vendor model becomes available before Phase 5 lands, prefer it for Model B.

Translation cost is small (~$0.05 per paper at 2026 API pricing) and applies only to the Phase 5 deep-dive subset (top 3-5 species), not the full breadth pass.

Full source list with URLs, access methods, expected yield, and license terms in experiments/comp-014-medicinal-mushroom-compound-mapping/inputs/data-sources.json.

§4 Method (planned for Phases 2-6)

§4.1 Phase 2 — Breadth aggregation (species-agnostic)

Pull every fungal compound record from LOTUS, NPAtlas, KNApSAcK, NPASS, TCMSP, COCONUT — Kingdom = Fungi filter at the database level, NO pre-curated species list. Apply toxicity-filter.json: include any species with FDA GRAS / EFSA QPS / pharmacopoeia / clinical-trial precedent OR any species in compound DBs without coexisting mycotoxin flag (default-include with safety_review_pending for grey-zone species); hard-exclude WHO Fungal Priority Pathogens 2022, documented mycotoxin producers, DEA Schedule I/II producers. Dedup by InChIKey across all sources. Expected yield: 500-2000 unique species after filtering (vs. 18 in the curated anchor list). Output: a unified table of fungal compounds × producer species × source provenance × safety status.

§4.2 Phase 3 — Target mapping

For each compound, query empirical bioactivity in this priority order: 1. ChEMBL via mcp__plugin_chembl_ChEMBL__get_bioactivity (highest curation, Western-pharma-skewed) 2. HIT (curated TCM ingredient → target) 3. PubChem BioAssay (broader corpus, less curated) 4. SwissTargetPrediction (similarity-based prediction — used only when 1-3 return nothing; flagged as predicted, not empirical)

Output: compound × target table with empirical/predicted flag and source provenance per row.

§4.3 Phase 4 — Chokepoint intersection

Join compound × target × chokepoint-targets.json. For each compound, score: - Empirical chokepoint hits: count of validated bioactivity entries against UniProt accessions in chokepoint-targets.json - Predicted chokepoint hits: count of SwissTargetPrediction hits against the same - Site fit: does the compound's known bioavailability profile (where available from LOTUS / NPASS) match the chokepoint's anatomical site? - Mechanism diversity: does the compound hit ≥2 mechanistically distinct chokepoints (rare and high-leverage)?

Output: ranked candidate compound × chokepoint pairs.

§4.4 Phase 5 — Multilingual literature ingestion (parallel, all chokepoint-hit species)

Run multilingual primary-literature ingestion for every species with a chokepoint hit from Phase 4, parallelized via subagents. NOT capped at top 3-5 — Brian's correction 2026-05-06: capping was the exact path-dependent narrowing CLAUDE.md flags ("Never gate exploration by cost or wall-time-to-build"). Translation cost via DeepSeek is ~$0.05/paper; even thousands of papers totals a couple hundred USD, not a real gate.

Two-model cross-check pairing per CLAUDE.md §Translation protocol: - Chinese sources (CNKI, Wanfang): Claude (Sonnet 4.6 or Opus 4.7) + DeepSeek V4-Pro via OpenRouter. DeepSeek is the cheap (~$0.14/M tokens vs. Claude $15-75/M) Chinese-vendor model with native Mandarin fluency — same multi-vendor heterogeneity-guard pattern the wiki sweep daemon already uses across Pass 1 / Pass 2 / Pass 3. - Japanese sources (J-STAGE, CiNii): Claude + Gemini 2.5 Pro (or DeepSeek as backup — DeepSeek's Japanese is reasonable despite Chinese-vendor). - Korean sources (KISS, RISS): Claude + Gemini OR GPT-5.

The actual constraint on this phase, honestly stated: CNKI / Wanfang / KISS need institutional subscriptions for full text; J-STAGE / CiNii / PubMed are public. Paywall access is its own operational task — work with open-access subset and flag the paywalled gap explicitly per ingestion run.

Phase 5 deliverable: redox/disulfide chokepoint admit/PRELIMINARY/reject decision (criteria in §3.2).

§4.5 Phase 6 — Per-compound triage

comp-013-style IC50 occupancy + composite scoring. Verdicts: GUT-LUMINAL VIABLE / SYSTEMIC VIABLE / MECHANISM UNCLEAR / NON-VIABLE. Subset gets shio-koji protease stability comp-NNN follow-ups (the same comp-001/005/006/012 pattern for any compound that becomes a candidate engineering payload — though most fungal-compound hits will be small molecules where the relevant downstream decision is supplement-stack inclusion rather than koji engineering).

§5 Why this matters (platform context)

For the Open Enzyme platform thesis broadly: - The platform claim is "we explore all angles, every avenue, fully open." The matrix in modality-chokepoint-matrix.md operationalizes the "all modalities × all targets" search space. comp-014 systematically scans the fungal natural products row of that matrix — the row most likely to surface chemistry the wiki has not enumerated, because fungal pharmacognosy has historically been explored in non-English-corpus traditions.

For comp-013 (TCM gout compound triage): - comp-014 is the breadth follow-up to comp-013's hand-curated 9-compound triage. comp-013 established the methodology; comp-014 scales the candidate-pool acquisition. - Several comp-013 verdicts are likely re-tested under comp-014 if the same compound-species (e.g., berberine in Coptis) appears as a fungal-co-occurring metabolite — though the primary candidate pool for comp-014 is fungal natural products, not herbal.

For the "honest platform gap" at CP0 (complement priming): - validation-experiments.md §1.21 (executed 2026-04-27) confirmed via ChEMBL/NPASS/LOTUS/Open Targets scan that NO validated natural-product C5aR1 antagonists exist. comp-014's broader corpus — particularly KNApSAcK and HIT — may surface candidates the §1.21 scan missed, especially if there's CNKI clinical evidence for a fungus-derived anti-complement effect that never made it into the Western-curated databases. Even one validated fungal C5aR1 antagonist would partially close the avacopan-dependence gap. Highest single-leverage potential outcome of the breadth pass.

For the proposed redox/disulfide chokepoint: - If Phase 5 confirms the warrant for adding this chokepoint to the formal inventory, modality-chokepoint-matrix.md gets a new column. Existing rows (small molecules, peptides, engineered organisms) light up with whatever fungal-compound coverage the breadth pass surfaced, plus implications for the engineered-koji and engineered-LBP rows (since fungi themselves are producers of ergothioneine and related thiols — koji is already a producer).

§6 Limitations (Phase 1)

  1. Phase 1 has no actual data — only an execution plan. All quantitative claims about coverage, hit count, mechanism diversity will be filled in Phase 2-4. Anyone reviewing this page should treat Phase 1 as a hypothesis to be tested, not a finding.

  2. Database coverage may be uneven. KNApSAcK is decades-old and well-maintained but not API-first; some Asian-hosted databases (TCMID notably) have had maintenance gaps. Phase 2 will verify each source is actually accessible before fetching; sources that prove inaccessible will be documented as gaps in the Phase 2 provenance update, not silently dropped.

  3. ~~Translation cost is real.~~ REMOVED 2026-05-06 — this was the exact failure mode the umbrella CLAUDE.md Curiosity & First-Principles Framing rule warns against ("Never gate exploration by cost or wall-time-to-build. Both have collapsed in the AI era"). Translation via DeepSeek is ~$0.05/paper × thousands of papers = a couple hundred USD total. Not a gate. Phase 5 runs for every chokepoint-hit species in parallel via subagents, not the top-3-5 subset I originally imposed.

  4. SwissTargetPrediction outputs are hypotheses, not validated bioactivity. Phase 3's predicted-target rows must remain flagged as such all the way through to Phase 6 verdict assignment. Pretending a SwissTargetPrediction hit equals a ChEMBL IC50 would produce false confidence — same epistemic discipline as comp-013's animal-model-evidence rule (admissible but visibly distinguished from biochem IC50).

  5. The proposed redox/disulfide chokepoint is speculative until Phase 5. Naming it in the chokepoint-targets.json with _PROPOSED_ prefix is deliberate — it must not propagate into other wiki pages (modality-chokepoint-matrix, nlrp3-exploit-map, complement-c5a-gout) until Phase 5 confirms or rejects. This is the same discipline Brian's pre-commit grep-verify gate imposes on quantitative claims.

  6. Fungal taxonomy is messy. Ganoderma lucidum in Western literature often refers to a species complex (multiple genuine species sold under one name). Cordyceps sinensis was renamed Ophiocordyceps sinensis in 2007. Phase 2 will use NCBI Taxonomy IDs as the canonical join key; common-name and TCM-name fields are for human readers, not joins.

  7. Genome-coverage upstream provenance gap. antiSMASH-DB and MIBiG predicted-BGC entries are downstream of fungal genome assembly + annotation. For candidate species lacking an assembled genome in MycoCosm/JGI or NCBI RefSeq Fungi, predicted-BGC coverage is zero — not because the fungus encodes nothing relevant, but because no genome has been sequenced. Phase 2c will explicitly flag which candidate species lack assembled genomes; absence-of-evidence ≠ evidence-of-absence for those species.

  8. Paywalled comprehensive databases are excluded. Dictionary of Natural Products (DNP, Chapman & Hall), SciFinder (CAS), and Reaxys are the most comprehensive curated natural-product databases but are institutionally paywalled. Their exclusion from comp-014 is a coverage limitation, not a methodology defect — but it means a true "complete fungal compound space" remains out of reach until institutional access is acquired.

  9. Marine fungi are deliberately excluded from this scope. Marine Penicillium and Aspergillus species (and many less-studied marine basidiomycetes) produce ETP-class compounds with internal disulfide bridges (gliotoxin, sirodesmin, chetomin family) that would directly affect the redox/disulfide chokepoint thesis. The exclusion is for tractability — marine fungi sit in a different corpus (MarinLit) with different access conventions, and integrating them at Phase 2 would substantially expand scope. Effect on the redox chokepoint warrant: comp-014's evidence base for the chokepoint addition is terrestrial-fungi-only. A future comp-NNN extending to marine fungi could expand the warrant or convert a Phase-5-PRELIMINARY verdict to ADMIT.

  10. Cordyceps sinensis / militaris split as a join-key risk. The Phase 1 candidate file initially conflated these two genuinely distinct organisms (different taxonomy IDs, different chemistry — cordycepin abundant in militaris, sparse in sinensis). Updated 2026-05-06 to treat them as separate candidate rows. Phase 2 dedup on NCBI Taxonomy ID rather than common-name; any Phase 2 LOTUS / NPAtlas record listing only "Cordyceps sinensis" with the wild/cultivated origin ambiguous gets routed to a manual-resolution queue rather than auto-merged.

§7 Six Phase 2 follow-ups queued

Same pattern as engineered-lbp-chassis.md and sirna-urat1-modality.md — Phase 2 starts when Brian green-lights:

  1. Phase 2a — LOTUS + NPAtlas pull. Two open-license, well-API'd Western-hosted sources first. Establishes the deduplication infrastructure and the first compound × species table. Lowest-cost, fastest signal on whether the methodology is sound before committing to harder-to-pull sources.

  2. Phase 2b — KNApSAcK + NPASS + TCMSP + HIT pull. The Asian-hosted compound and target databases. May require scraping where bulk download is not exposed; budget includes time for that. This is where ChEMBL-coverage-gap compounds get their first target hypotheses.

  3. Phase 2c — MIBiG + antiSMASH-DB BGC scan. Genomic-encoded compound space. For each candidate species with a sequenced genome, enumerate predicted BGCs; intersect predicted-product class with chokepoint-relevant chemistry. Surfaces "compounds we should be looking for that haven't been isolated yet."

  4. Phase 3 — target mapping. ChEMBL → HIT → PubChem BioAssay → SwissTargetPrediction cascade. Output joined to Phase 2 compound table.

  5. Phase 4 — chokepoint intersection + ranked candidate list. Output: top-N candidate compound × chokepoint pairs, ranked by empirical-evidence weight + site fit + mechanism diversity.

  6. Phase 5 — multilingual deep-dive on top 3-5 species + redox/disulfide chokepoint confirm/reject. CNKI, Wanfang, J-STAGE, KISS primary literature. Two-model translation cross-check. Final Phase 5 deliverable: a single decision document — does the redox/disulfide chokepoint enter the canonical wiki inventory or not?

Phase 6 (per-compound triage) is gated on Phase 5; budget separately when it lands.

§8 Cross-references

Maintenance

  • Status promotions: when each phase completes, update YAML frontmatter status and the computational-experiments.md row simultaneously. Phase 5 may trigger updates to modality-chokepoint-matrix.md (new column) and possibly nlrp3-exploit-map.md (TXNIP layer) — those propagations are sweep-daemon territory once the page transitions out of scoped.
  • Per-phase output files: add to experiments/comp-014-medicinal-mushroom-compound-mapping/outputs/ (e.g., phase-2-compound-table.json, phase-4-chokepoint-intersection.csv). Do not gitignore.
  • Per-phase scripts: add to experiments/comp-014-medicinal-mushroom-compound-mapping/scripts/ (e.g., phase_2_aggregate.py, phase_3_target_map.py). Stdlib + documented external API calls only.
  • Translation provenance for Phase 5: for each non-English source ingested, log original-language title, both translation outputs (Model A + Model B), the disagreement annotations, and the final translated text actually used in the wiki. Same discipline as tcm-gout-compound-triage-computational.md verification-gate disclosure.