🧬 RCDome

A multi-database human gene atlas of regulated cell death

_{RCD ontology dome: the cupola rim carries the 53 entities colored by ontological level (L1–L5) over a constellation of 9,265 genes — interactive in the live app (3D RCDome module)}

RCDome is a gene-centric, multi-source knowledge platform that compiles, normalizes and independently validates every human gene–regulated cell death (RCD) association against the published record — integrating 20 source databases and resources into a single, traceable atlas of 53 canonicalized RCD-related entities and 9,265 genes.

Every claim is verifiable: each association carries its source provenance, evidence tier, mechanistic classification, functional role, direction, and the verified publication (PMID) that supports it — or is explicitly marked by its mechanistic class — DATABASE_ONLY (no PMID), INSUFFICIENT_DATA (PMID but no mechanistic evidence), LITERATURE_GAP, CONTEXT_GAP, or CO_OCCURRENCE.

The dome is RCDome's visual signature. Its cupola rim carries the 53 canonicalized RCD-related entities as nodes colored by their adjudicated ontological level — L1 modalities (19, gold), L2 subtypes (7, blue), L3 triggers (2, green), L4 Functional/contextual (13, purple), and L5 Generic/umbrella/taxon (12, orange) — over a constellation of the 9,265 HGNC-validated genes. The color-coding makes the ontology's core claim visible at a glance: only 19 of the 53 entities are bona fide RCD modalities (L1); the remaining 34 are subtypes, triggers, contextual programs, or umbrella concepts, and their level color encodes that distinction.

---

📦 Large assets (entity figures, the matrix CSV, multi-omics tables) are distributed via the GitHub Release — see the release assets to reproduce the full app.

📊 Key numbers (truth matrix V027)

Matrix	22,473 rows × 100 columns
Unique gene–RCD pairs	21,568 (+ 905 multi-database corroboration rows)
HGNC-validated genes	9,265
Canonicalized RCD-related entities	53 (from 62 original name strings)
Source databases and resources	20
Independent analyses	4 — DeepSeek · Elicit · Consensus · Evidence-first adjudication
Executors / regulators / markers / unresolved	225 / 4,455 / 95 / 16,793
Direction (promotes / protects)	2,102 / 2,083
PMID titles verified against PubMed	10,141 / 10,141 (100%)

Evidence-first adjudication (4th analysis). Beyond the three independent classifiers, every gene–RCD pair carries a final verbatim-grounded resolution (final_mechanistic / final_role / final_direction), arbitrated by GPT-4o + two human curators (Curator 1, Curator 2) and tracked in the Disagreement Radar (7,905 label-disagreement rows on 7,831 unique pairs, all adjudicated). This is the evidence-first layer that resolved every classifier disagreement from the verbatim — the authoritative default across the app.

🧭 Conceptual framework

The pipeline runs in one top-to-bottom workflow: 20 source databases and resources → five compilation channels (A recall · B precision · C AI screening · D manual curation · E Elicit Phase 3) → consolidation → three independent validations (DeepSeek, Elicit, Consensus) → the 100-column matrix → the RCDome web application.

🖼️ Overview infographic

_{Regulated Cell Death Atlas — one-page overview: the 53 canonicalized RCD-related entities colored by ontological level (L1–L5), the 9,265-gene atlas, the five compilation channels, and the evidence-first adjudication that produces the traceable 100-column matrix.}

---

🗂 Repository map

Folder	Purpose
docs/manuscript	The manuscript (governance V027) + the OD M&M and Results subsections
docs/supplementary	Supplementary information, tables, and variable dictionary
docs/figures	Figures 1–4 and Supplementary Figures S1–S3 (600-DPI TIFF + PNG)
docs/	The traceability & governance contract (docs/RCDome_Traceability_Contract.md)
docs/governance	Medina's AI Governance Rules + the RCDome Definition Governance (invariance, provenance P0–P3, two-call)
docs/decoder	OD inventory, draft definitions, review guide, sync tracking
data/	The truth matrix (unified_gene_rcd_matrix_V027.rds in-repo; the CSV is in the Release) + 10,141 verified PMID titles + the 53-entity decoder (RCDome_Decoder_53*), NCCD-verbatim JSON
code/extraction	Source-database extraction (channels A, B, E)
code/validation	Phase 2A/D, DeepSeek CoT, and Consensus validation
code/decoder	OD definition pipeline (build, S scoring, audits, merge, full-text RAG validation)
code/benchmarks	Agnostic benchmark scripts (v1–v11 + Consensus v1–v7)
code/merge	Matrix merge/consolidation scripts
code/figures	Figure generators (reproduce every figure)
app/	The RCDome Shiny application
results/changelogs	Version changelog

Every copied file is listed in mirror_manifest.txt with its SHA-256 hash — the full copy is integrity-checkable.

---

🧬 RCDome Decoder — 53 operational definitions

Every canonicalized RCD-related entity carries an operational definition (OD) built in a three-tier hierarchy — NCCD 2018 verbatim (Galluzzi et al. 2018, Cell Death Differ 25(3):486–541, PMID 29362479, invariant) → Rodrigues et al. 2025 verbatim (Rodrigues de Souza et al. 2025, Front Bioinform 5:1630518, our prior study) → RCDome-drafted (from the founding article) — and scored by the six-criterion Rodrigues rubric with overall score S = 0.25·RS1 + 0.25·RS2 + 0.20·RS3 + 0.15·RS4 + 0.10·RS5 + 0.05·RS6 (verified 25/25 against the multi-optosis anchors, 53/53 bitwise).

Layer	File
Entity records	data/RCDome_Decoder_53.csv / .xlsx (+ _FINAL)
Two-call review	(intermediate worksheets — available in the GitHub Release)
Governance	docs/governance/RCDome_Definition_Governance.md (invariance · provenance P0–P3 · two-call)
Methods / Results	docs/manuscript/MandM_operational_definitions.md + docs/manuscript/Results_OD_census.md
Validation	DeepSeek full-text RAG · Consensus API · Europe PMC recovery · PubMed Knock Test (52/52)
Verbatim source	data/galluzzi2018_nccd_definitions.json (NCCD 2018 Box 1, byte-identical)

The decoder is fully audited (final readiness audit 26/26; verbatim audit 20/20 NCCD + 25/25 Rodrigues; S formula 25/25) and every entity is PMID-anchored with a full provenance chain (source databases → channels → reference PMID → author).

References.

1. Galluzzi L., Vitale I., Aaronson S. A., Abrams J. M., Adam D., Agostinis P., et al. (2018). Molecular mechanisms of cell death: recommendations of the Nomenclature Committee on Cell Death 2018. Cell Death Differ 25(3): 486-541.
2. Rodrigues de Souza E., Almeida Cordeiro Nogueira H., da Silva Francisco Junior R., Garcia A. B. and Medina-Acosta E. (2025). Integrated multi-optosis model for pan-cancer candidate biomarker and therapy target discovery. Front Bioinform 5: 1630518.
3. Rodrigues de Souza E., Almeida Cordeiro Nogueira H., Lopes V. d. S. and Medina-Acosta E. (2026). A pan-cancer multi-omic SuperLearner for regulated cell death survival topologies. Frontiers in Artificial Intelligence 9.

---

🚀 Run the app locally

Live app: https://rcdome.org/

setwd("/path/to/RCDome/app")
shiny::runApp(appDir = ".", port = 8890, launch.browser = TRUE)

The app loads data/unified_gene_rcd_matrix_V027.rds (place a copy in app/data/, or point global.R to it). The CSV (unified_gene_rcd_matrix_V027.csv) is the human-readable truth export (in the GitHub Release).

---

🔁 Reproducibility

1. Secrets — the Shiny app reads .Renviron (DEEPSEEK_API_KEY, SMTP_USER, SMTP_PASS); the code-level API keys are templated in .env.example (NCBI_KEY, DEEPSEEK_KEY, ELICIT_API_KEY, CONSENSUS_API_KEY). Keys are never committed.
2. Execution order — sources → channels (code/extraction) → Phase 2A/D & CoT & Consensus (code/validation) → merge (code/merge) → benchmarks (code/benchmarks) → figures (code/figures).
3. Benchmarks — the benchmark families are agnostic (fixed ground-truth sets defined independently of the classifiers): DeepSeek model-selection (26 pairs), Elicit governance (784 records), Consensus query (26 pairs), PMID title validation (10,141 titles), and the Digital Molecular RCD Board & Q&A regression suite (five checks, n = 100 each).
4. Smoke tests — contains automated checks (run by GitHub Actions on every push) that verify the mechanistic classification mapping, the chain-of-thought governance rules, and the benchmark scoring.

🏆 Benchmarks (headline)

Analysis	Result
DeepSeek chain-of-thought (26-pair)	100% accuracy · 0% false positives
Consensus neutral query (26-pair)	100% accuracy · 0% false positives
Elicit post-export governance	5/5 gates passed (allow-list 75.4→100% · unsupported EXECUTOR rate 53→0% · traceability 24→98%)
PMID title validation	10,141/10,141 (100%; 95% CI 99.96–100%)
Digital Molecular RCD Board & Q&A — 5 regression checks (n = 100 each): reference discipline, context fidelity, English-only, minimal source, insufficient-data	100% each (95% CI 96.4–100%)

---

🧾 Governance & traceability

This project follows Medina's AI Governance Rules in Writing Scientific Manuscripts (see docs/governance): AI assists without replacing the human; every number is verified against the matrix; no placeholders, no make-believe. The full traceability & governance contract (versioning, synchronization, forbidden actions) is in docs/RCDome_Traceability_Contract.md.

⏳ Honest status

Item	Status
Matrix V027	✅ complete
Manuscript, figures, supplementaries	✅ complete (V027)
DeepSeek CoT (full matrix)	✅ complete (21,568 pairs)
Elicit (Phase 3)	✅ complete
Consensus full run	✅ complete (21,568 / 21,568 pairs)
App modules	✅ complete (Home dome, Explore RCDome, Operational Definitions, Compare Definitions, RCD entities, Gene Catalog, Discrepancy Radar, Discovery Timeline, 3D RCDome, Digital Molecular RCD Board + Q&A, Network, About)

📄 How to cite

Manuscript: RCDome: a multi-database human gene atlas of regulated cell death. Emanuell Rodrigues de Souza, Victor dos Santos Lopes, Higor Almeida Cordeiro Nogueira, Enrique Medina-Acosta.

👥 Authors

• Emanuell Rodrigues de Souza
• Victor dos Santos Lopes
• Higor Almeida Cordeiro Nogueira
• Enrique Medina-Acosta — quique@uenf.br

📜 License

Code: MIT (LICENSE). Data and figures: CC-BY-4.0 (attribution required).

🐞 Bug reports & contact

Open an issue, or contact Enrique Medina-Acosta — quique@uenf.br