The bisphenol chemical exposome: a federated knowledge-graph map from exposure to adverse outcome

Framing (non-negotiable). The unit of analysis is the chemical–target–outcome chain for 15 bisphenol compounds assayed in high-throughput screens, integrated across seven OKN knowledge graphs on shared chemical (CAS/DTXSID), gene (Entrez) and disease (MONDO) identifiers. Every association here is hypothesis-generating, not causal or clinical: tox-assay activity, curated adverse-outcome pathways, and gene→disease associations are observational evidence of plausibility and convergence, not proof that a given bisphenol causes a given disease at real-world exposures. Keep this caveat attached to every downstream claim.

Abbreviations. BPA = bisphenol A; BPS = bisphenol S; BPF = bisphenol F; BPAF = bisphenol AF; BPB/BPE/BPZ/BPAP/BPP = bisphenol B/E/Z/AP/P; TBBPA = tetrabromobisphenol A; TCBPA = tetrachlorobisphenol A; BADGE = bisphenol A diglycidyl ether; BisGMA = bisphenol A glycidyl methacrylate; CAS = Chemical Abstracts Service number; DTXSID = EPA CompTox DSSTox substance id; AC50 = half-maximal activity concentration; AOP = adverse outcome pathway; MIE = molecular initiating event; KE = key event; AO = adverse outcome; ER = estrogen receptor; AR = androgen receptor; PXR = pregnane X receptor; PPAR = peroxisome proliferator-activated receptor; TTR = transthyretin; T4 = thyroxine; GO = Gene Ontology; BP = biological process; FDR = false-discovery rate (Benjamini–Hochberg); MONDO = Mondo Disease Ontology; NASH = non-alcoholic steatohepatitis; ICE = Integrated Chemical Environment; HTS = high-throughput screening.

1. Executive summary

This report reconstructs the bisphenol chemical exposome — the path from environmental/industrial exposure to adverse health outcomes — by integrating seven biomedical knowledge graphs in the OKN federation. Starting from 15 bisphenol compounds with active high-throughput tox-screening hits (out of 32 bisphenol substances catalogued in the Integrated Chemical Environment and 17 in ToxCast), the analysis maps each chemical to its molecular targets, mechanistic pathways, curated adverse-outcome pathways, and disease associations, then ranks chemical–disease relationships by the strength of independent cross-source support.

The bisphenols converge on a coherent, biologically interpretable target set: 183 distinct human target genes (1266 active chemical–assay–gene hits) dominated by nuclear receptors — the estrogen receptor ESR1 (hit by 14/15 compounds, AC50 down to 0.035 µM), the androgen receptor AR, progesterone receptor PGR, the metabolic receptor PPARγ (AC50 down to 0.002 µM), and the xenobiotic sensor NR1I2/PXR (active in all 15). Functional enrichment confirms the mechanism: Nuclear Receptor transcription is the most enriched Reactome pathway (~20× fold), alongside SUMOylation of intracellular receptors, xenobiotic metabolism, PPARα signalling and white-adipocyte differentiation; GO-biological- process enrichment adds xenobiotic metabolism, steroid metabolism and intracellular-receptor signalling.

Mapping the target genes onto curated disease genetics (rdkg) and effect domains (ICE) yields the disease landscape of the class — strongest fold-enrichment for thyroid tumour, ischemia/reperfusion injury, non-alcoholic steatohepatitis, obesity, coronary artery disease and hormone-dependent cancers (breast, prostate, endometrial, liver) and depression. Four AOP-Wiki pathways provide formal causal scaffolds: TBBPA → transthyretin binding → thyroid-hormone disruption → decreased cognition, and three BPA pathways through the estrogen receptor and GPER to lupus, autism-like behaviour and memory impairment.

What this adds: integrating the axes into a per-pair consensus shows that BPAF and TBBPA — not BPA — carry the broadest, strongest cross-source disease support (6 Tier-A links each, up to 36 shared target genes), a data-driven signal of "regrettable substitution" that is consistent with, and extends, the experimental literature. Of 186 chemical–disease pairs scored, 21 reach Tier A (three independent evidence types, or a gene-linkage backbone with matching toxicity-domain support).

2. Sources used

Every row is a knowledge graph actually queried with a logged SPARQL query in this analysis (16 logged queries total; see the reproducibility record). Versions and dates are from the federation VoID metadata (get_kg_version).

Bridging note: the Entrez↔symbol resolution (spoke-okn) and the symbol→UniProt→GO/Reactome traversal (prokn) are label/identifier bridges; every cross-KG claim in the report traces to one of the logged queries above. KGs explored but not credited (e.g. Tox21, PubChem-annotations) carried no logged query and are deliberately omitted.

3. Design & rules

Chemical inventory. Bisphenols were retrieved from ICE and ToxCast by matching compound labels and their synonyms (bisphenol, sulfonyldiphenol [BPS], methylenediphenol/dihydroxydiphenyl isomers), keyed on CAS and DTXSID. This yields a 32-substance ICE census and a 17-substance ToxCast census; 15 compounds have active curated HTS hits and form the mechanistic backbone. A curated core of 13 named analogues (BPA, BPS, BPF, BPAF, BPB, BPE, BPZ, BPAP, BPP, TBBPA, TCBPA, BADGE, BisGMA) anchors the narrative.

Molecular targets & activity. For each compound, ICE "curated HTS" assay endpoints were filtered to active calls (Call = "Active"); each active endpoint contributes its Entrez target gene (assay_entrez_gene_id), its mechanistic target (throughMechanisticTarget), the toxicity / adverse-outcome domain it informs on (mayInformOn — e.g. Cancer, DART, CardioTox, Estrogen, Thyroid Hormone) and its potency (AC50, µM). Non-human assay targets and cytotoxicity ("gene None") rows were excluded from the human-target set.

Mechanistic pathways. Curated AOPs were traversed chemical ← has_chemical_entity ← stressor ← NCIT_C54571 (Stressor) ← AOP, then AOP → MIE / key events / adverse outcome. Functional programs came from prokn (gene symbol → UniProt → GO / Reactome). Disease genetics came from rdkg (gene → biolink:related_to → MONDO). Enrichment used a one-sided hypergeometric test against an explicit background (all annotated genes in the source KG) with Benjamini–Hochberg FDR.

Evidence separation. Different evidence types are kept in separate columns throughout (tox-assay activity, curated AOP, disease-gene overlap, effect domain) and never merged into a single score — the consensus counts how many independent types support each chemical–disease pair. The full replicator specification (exact predicates, IRI normalisations, backgrounds, thresholds) is in the reproducibility record.

The breadth ranking already previews the report's central asymmetry: BPAF (127 target genes) and TBBPA (105) are far more promiscuous than the parent BPA (55), and several analogues reach sub-micromolar potency.

4. Confidence tiers

Chemical–disease relationships are graded by the number of independent evidence types supporting the link (tox-assay target overlap, matching ICE toxicity domain, and a curated AOP), plus the size of the gene backbone.

Of 186 scored chemical–disease pairs: Tier A = 21, Tier B = 160, Tier C = 5. Enrichment findings (§6) carry their own statistical confidence (FDR).

5. Findings by axis

5.1 Molecular-target landscape

The landscape is dominated by nuclear receptors and hormone machinery: the estrogen receptors ESR1 (14 compounds, 21 assays, 0.035 µM) and ESR2, the androgen receptor AR, progesterone receptor PGR, the xenobiotic sensors PXR (NR1I2, all 15 compounds) and CAR (NR1I3), the metabolic receptor PPARγ (potent, 0.002 µM), plus aromatase (CYP19A1), thyroid machinery (THRα/β, deiodinases) and the oxidative-stress regulator NFE2L2/NRF2. This is a textbook endocrine-disruptor/xenobiotic signature and sets up every downstream axis.

5.2 Adverse-outcome / toxicity domains

Every compound informs on Cancer, DART (developmental & reproductive toxicity), and CardioTox domains, with substantial Estrogen, Androgen, steroid-hormone and Thyroid-hormone signal — the toxicity-domain fingerprint mirrors the receptor landscape of §5.1 and foreshadows the disease enrichment of §6.

5.3 Curated adverse-outcome pathways (mechanistic chains)

The chains give formal, curated causal scaffolds for two compounds: TBBPA displaces thyroxine from transthyretin, lowering serum and neuronal T4 and altering hippocampal biology to decrease cognition (AOP 152); BPA acts through ER-α in immune cells (→ lupus exacerbation, AOP 314), through ER antagonism (→ autism-like behaviour, AOP 522) and through GPER activation with oxidative stress/neuroinflammation (→ memory impairment, AOP 535). These are the only bisphenols with curated AOPs in the federation and anchor the Tier-A neuro/thyroid links in §7.

6. Domain analyses

6.1 Functional enrichment — GO and Reactome (both families run)

Both enrichment families were run against explicit prokn backgrounds (Reactome N = 6,032 genes; GO-BP N = 7,663 genes), hypergeometric + BH FDR. GO molecular-function and cellular-component were deliberately skipped — GO-biological-process answers "what programs are engaged", which is the question here; MF/CC would add localisation detail without changing the mechanistic story. Disease-gene enrichment (§6.2) and the broad GWAS-trait layer (§6.3) are reported separately.

Of 42 candidate Reactome pathways, 33 are significant; of 47 GO-BP terms, 43. The programs are exactly those expected of endocrine-active xenobiotics: Nuclear Receptor transcription (~20× fold), SUMOylation of intracellular receptors (27×), Xenobiotics (21×), PPARA gene expression, white-adipocyte differentiation, extra-nuclear estrogen signalling, mitochondrial UPR and FOXO oxidative-stress transcription; GO adds xenobiotic and steroid metabolic process, intracellular-receptor signalling and hypoxia response.

6.2 Disease-gene enrichment (curated, rdkg)

The most specific signals (high fold) are thyroid tumour (~17×), ischemia/reperfusion and cerebral ischemia (15–17×), thyroid adenoma/cancer, NASH (10×), coronary artery disease (~6×) and obesity (~7×); the largest-overlap signals are the hormone-dependent cancers (breast, prostate, endometrial) and liver cancer, plus unipolar depression. Because common polygenic diseases carry large gene sets, they enrich at lower fold than the tightly-defined thyroid/ischemia sets — the fold column, not raw overlap, is the discriminating measure. Enrichment is associational, not causal.

6.3 Broad GWAS-trait layer (digcfdekg, comparator)

The broad, GWAS-derived gene→trait layer (digcfdekg) is reported descriptively and kept separate from the curated disease enrichment. All 183 targets carry GWAS traits (38,833 gene–trait pairs, 3,564 traits), so this layer is broad by construction and near-null under formal enrichment — as expected for a set covering a large fraction of trait-annotated genes. Its top traits nonetheless echo the mechanism: testosterone and sex-hormone-binding-globulin measurements, hypothyroidism, type-2 diabetes, total cholesterol, triglycerides, hypertension and abdominal aortic aneurysm — an endocrine–metabolic– cardiovascular signature consistent with §6.2. Treat as a corroborating context layer, not a discriminating test.

6.4 Exposure context (industrial / commercial uses)

ICE functional-use categories place the compounds in their exposure setting: BPA → binder / catalyst / hardener (polycarbonate & epoxy manufacture) plus antioxidant / UV-absorber; TBBPA and TCBPA → flame retardant; BADGE and epoxy resins → binder / hardener / monomer (food-can coatings); BPS → colorant / thermal-paper developer; the remaining analogues (BPF, BPB, BPE, BPP) → antioxidant / polymer intermediates. These uses — food-contact plastics and coatings, thermal paper, electronics flame retardants, dental resins — are the routes by which the molecular hazards above become human exposures.

7. Discussion

The axes assemble into one coherent picture, summarised in the synthesis map (Figure 7): bisphenols are broad-spectrum nuclear-receptor modulators whose molecular promiscuity (§5.1) maps onto a small number of mechanistic modules (§6.1) that in turn map onto a specific disease landscape (§6.2) and, for two compounds, onto curated causal pathways (§5.3).

Three implications follow. (1) The endocrine axis is primary: ER/AR/PGR/PXR/PPARγ modulation is the convergent mechanism, linking directly to hormone-dependent cancers and metabolic disease. (2) The thyroid–neurodevelopment axis is real and compound-specific: TBBPA's curated transthyretin AOP, its deiodinase/THR targeting, and the thyroid-tumour enrichment triangulate a thyroid-disruption hazard that BPA's estrogenic profile does not share. (3) Substitution has not removed hazard: the most mechanistically connected compounds are BPAF and TBBPA, not BPA (§8).

Testable predictions. (i) BPAF and TBBPA should show breast/prostate and thyroid endpoints at potencies at or below BPA in matched assays; (ii) PPARγ-active analogues (very potent here) should score as obesogens in adipogenesis assays; (iii) compounds sharing ESR1 + Cancer-domain hits should co-cluster in hormone-dependent-cancer epidemiology. Each is a decision-relevant hypothesis for prioritising analogues for regulatory testing — flagged by evidence strength, and none of them a causal claim.

8. Comparison with prior work

Central claims were checked against the primary literature via the Paperclip full-text corpus (PMC + preprints); the per-claim record with citations is in Bisphenol-Exposome_literature_comparison.md. (A PubMed connector was reported enabled but did not surface as a callable tool in this session; because Paperclip indexes PubMed Central full text, claims were verified against primary full text rather than abstracts.)

Central claims 1–6 were verified against full article text (not abstracts). Where the KG evidence diverges from the literature it is a matter of scope, not error: Claim 8 is a coverage limitation of the curated ICE ER/AR assay set (BPS is under-represented there), not a contradiction of BPS hazard; and Claim 7's ranking is a genuinely new synthesis the source graphs enable but no single study reports. No outright contradictions of the literature were found. See Claim 7 and Claim 8 in Bisphenol-Exposome_literature_comparison.md for the full per-claim detail.

9. Full ranked results

The complete ranked chemical–disease consensus (186 pairs) is in Bisphenol-Exposome_results.xlsx (sheet Consensus chem–disease) and data/consensus_chem_disease.csv; the target, enrichment and AOP tables are in the other workbook sheets. The interactive table below is sortable (click a header), filterable (search box + the tier / category / AOP pull-downs) and paginated; the sources (n) column shows how many federation KGs corroborate each row — biobricks-ice (tox-assay target + effect domain) and rdkg (disease genetics) support every row, with biobricks-aopwiki added where a curated AOP matches.

The ranking makes the class structure explicit: BPAF and TBBPA head the Tier-A list across cancer, metabolic and (for TBBPA) thyroid/neuro outcomes, with BPB, TBBPA-DHEE and TCBPA close behind — the parent compound BPA is Tier-A only where a curated AOP corroborates the gene/effect evidence (the neuro-behavioural link). The consensus matrix (Figure 8) shows the same pattern as a chemical × disease grid.

The grid concentrates in the upper-left (BPAF, TBBPA, BPB) across hormone-dependent cancers, liver cancer and depression, thinning toward the less-assayed analogues (BPS, BPE) — a visual restatement of the coverage caveat in Claim 8.

10. Summary of findings & limitations

Findings recap. Across 15 actively-screened bisphenols, the class converges on 183 human targets dominated by nuclear receptors (ER, AR, PGR, PXR, PPARγ), engaging 33 Reactome and 43 GO-BP programs centred on nuclear-receptor transcription, xenobiotic and steroid metabolism, and adipocyte differentiation. These targets enrich (curated rdkg genetics) for a specific disease landscape — thyroid tumour, ischemia, NASH, obesity, coronary artery disease and hormone-dependent cancers — and four curated AOP-Wiki pathways give TBBPA (thyroid → cognition) and BPA (ER/GPER → lupus, autism-like behaviour, memory) formal causal scaffolds. Integrating the axes, BPAF and TBBPA carry the broadest cross-source disease support, with 21 Tier-A chemical–disease relationships overall — a data-driven "regrettable substitution" signal.

Limitations.

1. Observational, not causal. Every edge (tox-assay activity, gene→disease association, AOP) is

evidence of plausibility and convergence, not proof of causation at real-world exposures. No exposure levels, doses, or pharmacokinetics are modelled here.

1. Assay coverage bias. Breadth counts depend on which assays a compound was tested in. Well-studied

compounds (BPA, BPAF, TBBPA) accrue more hits; sparsely-tested analogues (notably BPS, 10 targets) look "cleaner" than the wider literature supports (Claim 8) — absence of a hit is not evidence of safety. 2b. In-vitro provenance. ICE HTS activity is a perturbation signal; an active AC50 does not establish an in-vivo adverse effect.

1. Disease enrichment is associational and gene-set-size-dependent. Common polygenic diseases enrich at

lower fold despite large overlaps; rdkg is a rare-disease-centred graph, so its common-disease gene sets are incomplete. The candidate set was pre-filtered to diseases with ≥6 target genes, which inflates the significant fraction — the fold column is the discriminating measure, not the count of significant hits.

1. Symbol/identifier bridging. Entrez→symbol (spoke-okn) and symbol→UniProt→GO/Reactome (prokn) are

label/identifier bridges; a small number of non-human or unmapped assay targets were dropped, slightly undercounting the target set.

1. AOP coverage is sparse. Only BPA and TBBPA have curated AOPs in the federation; absence of an AOP

for an analogue reflects curation status, not absence of a pathway.

1. Consensus scoring is deliberately simple and evidence-separated. It counts independent evidence

types rather than weighting them; it is a transparent ranking, not a quantitative risk model, and the tier thresholds are analyst choices.

1. Literature comparison used one connector (Paperclip full text). A PubMed connector was not callable

in-session; coverage is broad (PMC + preprints) but not exhaustive.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt, the replicator specification (rules, thresholds, joins, verified quantities, limitations), and every supporting SPARQL query verbatim with its row count and query diagram, plus pinned KG versions — is in Bisphenol-Exposome_reproducibility.md, with the exact scripts in scripts/ and intermediate extracts in data/.

12. References

Full-text verification via the Paperclip MCP connector (PMC + preprint corpora).

1. Rochester JR, Bolden AL. Bisphenol S and F: A Systematic Review and Comparison of the Hormonal Activity of Bisphenol A Substitutes. Environ Health Perspect. 2015. PMID:25775505 · doi:10.1289/ehp.1408989 — full-text-verified (PMC4492270)
2. Srebny V, et al. Beyond Estrogenicity: A Comparative Assessment of Bisphenol A and Its Alternatives in In Vitro Assays Questions Safety of Replacements. Environ Sci Technol. 2025. doi:10.1021/acs.est.5c07018 — full-text-verified (PMC12392461)
3. Thoene M, et al. Bisphenol S in Food Causes Hormonal and Obesogenic Effects Comparable to or Worse than Bisphenol A: A Literature Review. Nutrients. 2020. PMID:32092919 · doi:10.3390/nu12020532 — full-text-verified (PMC7071457)
4. Ren X-M, et al. Binding and Activity of Tetrabromobisphenol A Mono-Ether Structural Analogs to Thyroid Hormone Transport Proteins and Receptors. Environ Health Perspect. 2020. PMID:33095031 · doi:10.1289/EHP6498 — full-text-verified (PMC7584160)
5. Longo M, et al. Low-dose Bisphenol-A Promotes Epigenetic Changes at Pparγ Promoter in Adipose Precursor Cells. Nutrients. 2020. PMID:33202789 · doi:10.3390/nu12113498 — full-text-verified (PMC7696502)
6. Gao H, et al. Bisphenol A and Hormone-Associated Cancers: Current Progress and Perspectives. Medicine (Baltimore). 2015. PMID:25569652 · doi:10.1097/MD.0000000000000211 — full-text-verified (PMC4602822)
7. Wang Z, Liu H, Liu S. Low-Dose Bisphenol A Exposure: A Seemingly Instigating Carcinogenic Effect on Breast Cancer. Adv Sci (Weinh). 2016. PMID:28251049 · doi:10.1002/advs.201600248 — full-text-verified (PMC5323866)
8. Singh M, et al. Tetra methyl bisphenol F: another potential obesogen. Int J Obes (Lond). 2024. PMID:38396134 · doi:10.1038/s41366-024-01496-5 — full-text-verified (PMC11216980)