Explore RCDome

RCDome is a gene-centric atlas that consolidates human gene–RCD associations captured across its governed source universe into one traceable platform. It is not a single database — it is a compendium of 20 independent resources, evaluated through three complementary independent analysis layers. The core question it answers: is this gene mechanistically involved in this form of cell death — and how?

The Home page is a girded cupola: 53 entity nodes along the rim, colored by ontological level (L1 gold, L2 blue, L3 green, L4 purple, L5 orange), and a gene constellation beneath. Hover a node to see its genes; click it to open its full page.

Each gene–RCD pair is examined by one or more of three complementary methods, each kept in its own columns so provenance is never lost:

Evidence-first — the final verbatim-grounded resolution of every pair (5,138 mechanistic · 225 executors · 4,606 regulators · 95 markers).
DeepSeek — a chain-of-thought classification of all 21,568 pairs (100%), giving mechanistic (YES/NO), role, and direction.
Elicit — systematic-review extraction across 50 forms (4,601 associations).
Consensus — an independent evidence-channel verdict (100% of pairs).

The Direction/Role source selector in the Gene Catalog lets you switch between analyses to compare them directly.

Only 225 of 21,568 pairs (4.6% of role-resolved) are executors — the genes that directly execute death (pore-forming, cleavage, permeabilization). Everything else is regulation (4,606 regulators). The "Death core" button in the Gene Catalog filters to this machinery in one click.

Search — the entity jump selector and the Gene Catalog filters.
Hover — every column and control carries an explanation (dark tooltips).
Filter — by entity, biotype, tier, role, direction, mechanistic verdict, and source database.
Download — filtered table, complete matrix, entity comparison, and the death core (CSV).

Every association carries its source databases, evidence tier, and its per-analysis verdicts. Direct association evidence rests on 10,141 curated PMIDs (each with a verified title), while co-occurrence force is reported as a literature-volume count — 316,000 PMIDs where the gene and RCD term merely co-occur. Every piece of information is downloadable (filtered table, complete matrix, entity comparison, death core) and traces back to its original source — a database or a PMID. No value is accepted without verification against the compendium matrix — the Knock Test.

RCDome is governed by an explicit traceability contract: no speculation (every association traceable to a source database or PMID), no unsupported inference (model-derived classifications are separated from source evidence and provenance-traced), no overwriting (each analysis writes to its own columns), full traceability (source databases, evidence types, and PMIDs on every row), and HGNC-only gene symbols. Every AI-assisted step follows the project's AI governance rules, with an anti-hallucination framework that rejects any sentence not found verbatim in its source.

The compendium, its manuscripts, and this interface are produced under an explicit AI-governance framework: AI assists, but the human remains both the source of the science and the safeguard against error.

Where the human belongs: Introduction and Discussion — human essential (the argument, narrative, and interpretation); Methods — protocol, precise and replicable (AI-assisted polishing welcome); Results — light (the human chooses what to present and in what order).

The three phases:
1. The Spark (no AI) — the hypothesis, argument, and narrative are imagined and written by the human before any AI touches the draft.
2. The Assist (AI as tool) — AI polishes Methods, suggests (never decides), and codes/debugs (never designs the analysis).
3. The Check (vigilance) — imagination is not hallucination; the Knock Test verifies every AI claim; only "I imagined this" and "AI helped, I confirmed" enter the record — "AI suggested, I'm unsure" never does.

The classifiers were validated against fixed, method-independent ground-truth sets before full-scale application: a four-paper landmark audit (100% gene recall and role classification), a 17-pair ground-truth set, and a 26-pair benchmark (10 executors, 8 regulators, 8 negative controls). The DeepSeek chain-of-thought classifier reached 26/26 (100%) combined accuracy, 0/8 false positives; the Consensus neutral-query rule reached the same (Supplementary Notes S5 and S3). All 10,141 curated PMID titles were verified at 100% accuracy against PubMed. The Digital Molecular RCD Board and Q&A were additionally benchmarked on random, seed-reproducible samples of n = 100 across five regression checks — reference discipline, context fidelity, English-only refusal, minimal-source reporting, and insufficient-data directness — each at 100% (95% CI 96.4–100%; Supplementary Note S6). A fresh held-out re-sample (seed 20260919, n = 100, excluding the development cases) reproduced the reference-discipline and context-fidelity checks at 100% (100/100; 95% CI 96.4–100%), confirming the results generalize to unseen cases.

A label-error benchmark measured the agreement between the independent classification layers on their shared labels and arbitrated 7,905 label-disagreements (role 320, direction 935, mechanistic 6,650) on 7,831 unique pairs with an independent vote and a human curator verdict; all pairs adjudicated the delta round (Supplementary Note S2). A semantic-contradiction audit found 337 internal contradictions (237 DeepSeek and 100 Consensus) and repaired all of them by conservative downgrade-only rules, with 0 remaining on re-audit. The 53 RCD entity descriptions were verified claim-by-claim against their cited sources: of 2,812 claims, 2,672 are supported (95.0%) and 140 were removed, leaving 2,129 references with no orphan citations, no duplicate DOIs, and no broken entries (Supplementary Note S8). A clickable-reference integrity benchmark (n = 100, seed 42) confirmed 100% accuracy, 100% link resolution, and 100% correct-paper specificity (95% CI 96.3–100% each).