comp-014 Phase 2 — Headline findings (2026-05-06)

Phase 2 ran three parallel passes against the chokepoint inventory. Each pass produced its own structured output; this file is the human-facing synthesis.

What ran

Pass	Source	Method	Status
2a	ChEMBL	MCP bioactivity sweep across 19 chokepoint targets, filter to fungal-source compounds	✅
2b	LOTUS	REST API iterated across 56 fungal genera/species, deduplicated by InChIKey	✅ partial (LOTUS only; KNApSAcK + NPASS + TCMSP + HIT pending)
2c	PubMed	17 searches across 3 axes (chokepoint × fungus, species × indication, multilingual catch-net)	✅

Top-line numbers

• 6,798 unique fungal compounds (LOTUS, deduped by InChIKey)
• 55 fungal genera/species with compound attributions
• 9,409 raw compound-source pairs seen
• 14 high-signal PubMed papers with primary-literature evidence at chokepoint targets
• 1 ChEMBL hit with empirical sub-μM bioactivity (Fumitremorgin C → ABCG2)

Highest-leverage individual findings

Dual-chokepoint hit — Ganoderma applanatum 2,4-DAE

PubMed PMID 35750011 (Fitoterapia 2022). 2,4-dihydroxyacetophenone derivatives from Ganoderma applanatum show in vivo SUA reduction from 407 → 134 µmol/L in murine hyperuricemia model. Hits both XO and URAT1 by mechanism. Strongest single-compound finding of the breadth pass.

URAT1 direct modulator — cordycepin (Cordyceps militaris)

PubMed PMID 29422889. Mouse dose-response, mechanism proven. Cordycepin is a 3'-deoxyadenosine purine analog — structurally adjacent to the nucleotide salvage / XO substrate axis we'd predicted from the candidate-species file.

Clinical-grade evidence — Cordyceps × hyperuricemia

• Cochrane review of 5 RCTs (PMID 26457607)
• 9-RCT meta-analysis (PMID 28137532) Both Chinese-source data — confirms multilingual ingestion is load-bearing for capturing this evidence quality. PubMed indexed only English abstracts; full RCT data is in CNKI/Wanfang.

ABCG2 — Fumitremorgin C (Aspergillus fumigatus)

ChEMBL CHEMBL1232 (or similar, full ID in JSON). IC50 800 nM, pChEMBL 6.10. The canonical ABCG2/BCRP discriminator tool compound from Western pharma curation. Fungal source: A. fumigatus — which is on the WHO Fungal Priority Pathogens list and EXCLUDED by the toxicity filter. So this hit is chemistry-relevant but producer-organism-excluded: the compound has chokepoint relevance but cannot use A. fumigatus as the production chassis. Synthetic / heterologous expression route required for downstream use.

Multi-chokepoint extract — AMC-BFE

2026 paper, PMID 41905012. Cordyceps × Astragalus solid-state co-fermentation extract hitting URAT1 + GLUT9 + ABCG2 + PPARα. Adjacent to the engineered-koji co-culture thesis.

Strategic findings (negatives + reframings)

C5aR1 platform-gap (validation-experiments §1.21) confirmed empirically

Both Phase 2a (ChEMBL) and Phase 2c (PubMed) returned ZERO direct fungal-source C5aR1 antagonists. ChEMBL has only 2 named C5aR1 compounds total, neither fungal. PubMed surfaced that mushroom β-glucans actually ACTIVATE complement via CR3/iC3b — opposite-sign to what the platform gap needs. Implication: the avacopan-dependence gap will not be closed by mushroom-extract programs. Phase 5 C5aR1 work needs synthetic / non-fungal source.

NLRP3 directionality is structure-dependent and OPPOSITE-SIGN across fungal compounds

Phase 2c surfaced this clearly: Ganoderma lucidum exopolysaccharides ACTIVATE NLRP3 (cryptococcosis defense, beneficial in that context); spore powder + GLP β-glucan INHIBIT NLRP3 (CNS contexts). Phase 5 cannot claim "fungal β-glucan inhibits NLRP3" without specifying triple-helix conformation, MW, branching, sulfation. Methodology constraint for Phase 6 triage.

ChEMBL coverage is severely Western-pharma-biased for fungal compounds

Phase 2a found ChEMBL has none of the canonical mushroom compound classes by name — ganoderic acids, ergosterol, hericenones, erinacines, inotodiol, pachymic acid, eburicoic acid, eritadenine, PSK. The CLAUDE.md global-multilingual research warning is empirically validated. LOTUS captured 6,798 unique compounds; ChEMBL captured handful of fungal compounds. The breadth signal is in LOTUS + East-Asian DBs (still pending) + multilingual literature.

Two new chokepoint candidates surfaced

Phase 2c found primary-literature evidence at chokepoints not currently in our inventory: 1. ADA (adenosine deaminase) — Ganoderma lucidum GLPP, PMID 36385640. Mechanism: ADA is upstream of xanthine oxidase in purine catabolism. Plausibly a fourth front in the urate-axis chokepoint map. 2. Mitophagy / PINK1 — Cordyceps cicadae, PMID 40334761. Adjacent to NLRP3 priming via mitochondrial dysfunction-induced ROS.

Both warrant addition to chokepoint-targets.json (Phase 4 follow-up) and possibly the canonical wiki chokepoint inventory.

Per-species breadth ranking (top from LOTUS)

Rank	Genus/species	LOTUS compound count
1	Ganoderma (genus, summed)	1,938
2	Aspergillus terreus	457
3	Lactarius	380
4	Aspergillus niger	302
5	Phellinus	292
6	Cordyceps (genus, summed)	283
7	Fomitopsis	282
8	Penicillium chrysogenum	240

The Phase 5 anchor list (18 well-studied medicinal fungi) all appear in the LOTUS pull — sanity-check passed.

Methodology gaps (deferred to Phase 2 follow-up runs)

1. East-Asian-hosted DBs not yet pulled. KNApSAcK (Japanese), NPASS (Chinese), TCMSP (Chinese), HIT (TCM ingredient targets), BATMAN-TCM. Together likely surface another 1,000-3,000 unique compounds. None of these have public REST APIs comparable to LOTUS — scraping or bulk download required.
2. LOTUS organism-attribution is via query-hint, not per-record metadata. The search/simple endpoint returns compound metadata but not organism per record; we attribute by query name. To upgrade, hit the per-compound LOTUS endpoint or use the Zenodo bulk dump.
3. 200-records-per-LOTUS-query truncation. Ganoderma lucidum returned 28MB of data suggesting we hit the limit. Re-run with limit=2000+ for top-10 species to expand coverage.
4. Toxicity filter not yet applied at the species level. All compounds pulled were retained — the filter logic in toxicity-filter.json needs to be applied as a post-processing step before Phase 4 chokepoint intersection. A. fumigatus example (Fumitremorgin C → ABCG2) is the canonical case where filter matters.
5. MIBiG + antiSMASH-DB BGC pull not run. Genomic-encoded compound space — for compounds we don't yet have isolation data for. Phase 2c (genome-upstream) deferred.

Next concrete step (Phase 3)

Phase 3 = target mapping. For each of the 6,798 LOTUS compounds, look up known bioactivity. Three sources in cascade:

1. ChEMBL by InChIKey (highest curation, lowest fungal coverage) — already partially run via Phase 2a; intersect that data with the LOTUS InChIKey list to find any cross-source hits.
2. HIT (Herb Ingredient Targets) — TCM-curated target hypotheses for compounds with TCM literature.
3. SwissTargetPrediction for orphan compounds (no empirical target) — flagged as predicted-not-empirical in the output.

Then Phase 4 intersects with chokepoint-targets.json to produce the ranked candidate list — the answer to Brian's original question.