ARIADNE — Concentric Radial Atlas

Priority Candidates

0 visible 0 total 0 validated 205 seeds 0 max CQS — coverage

SIGNATURE VIEWS — one click links physics + biology + translation. “Expert-ready” = not refuted by QA, shared genes consistent with the disease (qaGeneDisease = ok), and mechanistic fidelity MFS ≥ 0.9.

Bench-Ready

Strict / Balanced / Broad

First-Mover IP

Strict / Balanced / Broad

Restore-Safely

Strict / Balanced / Broad

Structural

Bypass Undruggable / Physics-Ideal Chokepoint / Cross-Domain Surprise

Convergence

Cohesive (z<−2)

Restraint

Handle With Care clear

Confidence & Clinical

Confidence Tier

Tier Rubric

CQS Quality Score

Genotype Fitness

Global Cross-Seed Recurrence

Clinical Pathway Family

Bridge mode — show only cross-family pairs (A ⇄ B). Select 2+ families to isolate bridges between them.

Clinical Family Legend

Provenance & Integrity

Engine Architecture

Novelty (Relationship Type)

Content QA (dossier integrity)

Gene Attribution Integrity

Wet-Lab & Physics

Wet-Lab Assay Readiness (FSS)

AlphaGenome Impact

Target Canalization (Poincaré r)

Patent Prior-Art

Approval & Trials

Independent Check (Open Targets)

Thermodynamic State (S_VNE)

Clinical Toxicity Risk (ΔS_VNE)

Thermodynamic Toxicity

Distance from the global attractor well (S_VNE ≈ 1.524). High ΔS_VNE indicates a diffuse, high-entropy state prone to off-target clinical trial failure.

Mechanism & Convergence

Shared Pathways

Convergence Node Scope

Literature Verification (RANA)

Physical Interactome Proximity

z is the interactome separation from the degree-preserving null (script 72); dyads with no real network test are excluded, not inferred. Overlap uses Menche s_AB on BioGRID.

Molecular Machine (convergence cluster)

Groups dyads by the molecular machine their shared gene belongs to (17 clusters). Counts are filter-aware: each shows how many candidates that machine has under your current filters, and they update as you change any filter.

Causal Chain Depth

Multi-Axis Chokepoint

Drug & Bottleneck

Drug Target Integrity

Mechanism Axis

Drug Specificity (Hub-Target)

Bottleneck Type (Chokepoint)

Gate 8 · Target Class Bypass

Gate 9 · Therapeutic Modality

Gate 11 · Developability & Safety

Patterns · Cross-Gate

Statistical Confidence

Presets & Translational

Presets (physics recipes)

✦ Neuro-Immunity Bridge (cGAS–STING)

✦ Lysosomal & Autophagy

✦ Bioenergetics / Mitochondrial

Hyper-Clusters (Elite 9)

Alzheimer's / Neuro Scout

Platt Lysosomal Precision

Discovery Frontier

Frances Platt Lens (LSDs)

Metabolite Bridges

Tightest Couplings

Deep Mechanisms

Allelic Calibration

Presets — one-click filter recipes

Hyper-Clusters: CQS ≥ 9, Rim r ≥ 0.5, Tight S_VNE ≤ 1.5.
Alzheimer's Scout: Neurologic, Goldilocks r, Tight S_VNE.
Platt Precision: Lysosomal + ChEMBL Drugs.
Discovery Frontier: sweet-spot bridge · multi-pathway · rim-specific · well-corroborated.

Cross-Domain Status (v2.0)

Pharos Target Development Level (v0.6.0)

Therapeutic Axis

GTEx v10 Expressed Tissue (v0.7.0)

Approved Small-Molecule Overlay (ChEMBL)

Top Convergence Nodes

Top Connected Genes

Top Repurposing Drugs

Loading... Load top_candidates_enriched.json

ariadne — a map of what diseases share

Around the rim live all the rare diseases, gathered into the families of biology they belong to — the coloured arcs, each disease a faint name and a small radial mark, together drawing the circle. In the interior float the shared origins: the genes, pathways and metabolites where two different diseases secretly meet. The centre is left empty — a quiet void for the deepest, most generic machinery of life, the hubs so universal they explain nothing on their own. The Atlas orbits that void; it never crosses it.

Every thread is a hypothesis. It leaves one disease on the rim, dives inward, and braids through the single shared origin it converges on, before rising to its partner on the far side. Where many hypotheses share the same origin, their threads weave into a bright knot — and that knot is not decoration, it is the biology itself. Each thread is a rare condition reaching toward another: a chance that an insight, a drug, a diagnosis might cross from one child's disease to another's.

How to read it

The rim is every disease, grouped by clinical family (the colour), labelled with its name and a radial tick whose length and brightness grow with how many hypotheses touch it — so recurrent "crossroads" diseases stand out and rare ones fade. Hover a family arc to light all of its connections.

The stars are the shared genes and convergence nodes. Their position carries meaning: the deeper toward the centre a star sits, the more generic it is — a promiscuous hub (like TP53 or ubiquitin) the instrument deliberately distrusts; the closer to the rim, the more specific and precious. Their brightness and size follow discovery quality — the best CQS of any hypothesis meeting there — so specific, well-scored convergences shine while the generic centre stays a hush. Hover any star to see every hypothesis that converges on it; a node is coloured by type (cyan gene, green pathway/process, amber metabolite).

The threads thicken gently with how connected their shared origin is. Hover a discovery in the left list, or on the map, and just that pair lights up: its two diseases named on the rim, its shared gene/pathway/metabolite flared and labelled, and every sibling pair that converges on the same origin glowing softly — the whole family of the same hope.

How it moves when you filter

Nothing here is fixed. The entire picture is computed live from exactly the hypotheses that pass your filters. Change a tier, a fitness floor, a family, a gene, a convergence node — and the stars re-place, the braids re-weave, and the constellation reorganises to reveal what is shared under that lens. That is the whole point: filtering is how the instrument shows you the unknown. By default the Atlas opens on its Exceptional discoveries (CQS 10–13, Tier 3). The right‑hand "Top Convergence Nodes" and "Top Connected Genes" recompute with every filter too — they are the shared origins the current view is resting on.

This is a radial specificity map, not a learned embedding. Each candidate's radius r is a monotonic function of its shared-pathway size — generic scale-free hubs collapse toward the centre (r → 0) and specific, expert-ready convergences are pushed to the rim (r → 1) — rendered in a Poincaré-style disc, combining the validated corpus (16,447 v2.1-evolved candidates) with the legacy discovery set (6,690 more) for 23,137 total. A star's depth is that specificity; its brightness is its discovery quality (CQS). To be precise: this visualizes the specificity score; it is not an optimized hyperbolic embedding of the interactome, and the distances between candidates are not metric hyperbolic distances.

Convergence Topology

A shared mechanism can look important just because it is busy — but busy often means meaningless, like “cell stress.” So for every shared mechanism we measure two things: how often it recurs (frequency) and how many different disease families it spans (breadth). Together they tell you whether a mechanism is a real discovery or just background noise. This card is that separation, made visible.

Scale-free: the top 1% of — shared mechanisms carry —% of all convergences.

Where it comes from: we count how many disease-pair hypotheses pass through each mechanism. Why it matters: a handful of hubs carry most of the traffic while thousands of mechanisms are rare specialists — this lop-sided shape is the fingerprint of the cell’s real wiring, and it is exactly why a few generic hubs must be told apart from the many specific mechanisms. That is the whole job of this card.

Master bottlenecks the molecular steps many separate diseases funnel into

Read a row: conv. = how many hypotheses pass through it · families = how many of the 21 clinical domains it touches · SRI = conv ÷ families = recurrence per domain. High SRI with few families = a genuine focused bottleneck; high families = a broad crossroads.

node	conv.	families	SRI

The gene map — the same gene, three very different meanings

A gene’s value is not how often it appears — it is how many different disease families it bridges. Sorted by that, every gene falls into one of three roles:

1. Domain masters (1 family) — validation: ARIADNE re-derives known drivers from scratch

2. Cross-domain bridges (2–4 families) — THE GOLD: one gene linking distinct diseases (hover for families)

Why this matters: a gene that bridges 2–4 families is a true pleiotropic crossroads — where a single insight (or a single drug) can explain several rare diseases at once. This is where discovery and drug-repurposing live.

3. Promiscuous hubs (≥8 families) — too broad to be a specific clue; background

Generic solvent nodes (≥10 families) — housekeeping noise, shown so you can filter it out

Processes so universal that every disease eventually touches them (cytoskeleton, neuron development). Historically people removed these with hand-made blacklists — here the maths (family breadth) demotes them automatically.

Filter the Atlas into any of these zones with Convergence Node Scope (Domain-master / Sweet-spot bridge / Cross-domain hub / Generic solvent) and Shared Pathways. Regenerated by 13_convergence_topology.py.

Candidate Quality Score (CQS v3)

A composite, additive metric (max 13) that ranks ARIADNE's discovery candidates by expert‑readiness. Higher scores indicate stronger, more specific, cross‑domain and better‑corroborated hypotheses.

Component	What it measures	Pts
1. Fitness	Overall ARIADNE genotype score	0–3
2. Specific convergence	Mechanism names a specific node (not "oxidative stress")	0–2
3. Multiple pathways	How many distinct Reactome pathways are shared	0–2
4. Pathway specificity	Smallest shared pathway size (≤20 genes = 2, ≤50 = 1)	0–2
5. Falsifiability	Has a concrete, measurable falsification test	0–1
6. Cross‑domain surprise NEW	Seed & partner in different clinical families (a non‑obvious bridge)	0–1
Hub corroboration	Independent recurrence across multiple seed runs (≥3 seeds = 2)	0–2
Gene‑attribution gate NEW	Fabricated causal gene (absent from Monarch) → hard‑caps the score at Weak	cap 2

Interpretation bands

CQS Range	Label	What to do
10–13	Exceptional	Lead with these in expert sessions
7–9	Strong	Include in review panels
4–6	Moderate	Background candidates; may need refinement
0–3	Weak	Generic convergences, artefacts, or fabricated attributions

v3 changes: the two dead placeholders (phenotype‑overlap, cross‑domain) are replaced — Cross‑domain surprise is now live (rewards bridges between different clinical families), and a Gene‑Attribution Integrity gate caps any hypothesis whose causal gene was flagged as fabricated at Weak, so a fluent‑but‑fabricated mechanism can never read as expert‑ready. Literature (RANA) verification and physical protein‑interaction (BioGRID) remain pending until those signals are logged into the discovery JSONs.

What Is ARIADNE?

A genetic algorithm whose organisms are scientific hypotheses. Given two rare diseases, ARIADNE breeds, scores, and culls candidate explanations for how they might share a biological mechanism — surfacing a ranked shortlist for expert review rather than answering from an AI model's memory. ARIADNE is a convergence engine, not a discovery engine: every disease it links is independently established — the value is that they are studied separately, in different clinical silos, yet converge on the same molecular machine, surfaced systematically with a hub-corrected topology statistic. Every cluster is a question for an expert, never a claim.

The core idea: breed hypotheses, don't look them up

A population of rough candidate hypotheses varies, each is scored against real biological databases and a panel of checks, the strong ones survive and are bred with small mutations, and the cycle repeats for several generations — good hypotheses improve, nonsense dies out.

Genome vs. genotype

The genome is the 9-slot content of one hypothesis: which two diseases, the type of mechanistic link, which direction the gene fault runs, the step-by-step causal chain, supporting evidence, and — critically — what experiment would prove it wrong.

The genotype is the 8-test scoring panel behind a candidate's Fitness score: topological proximity, evidential pluralism, directionality match, ClinGen tier, phenotype coherence, Bradford-Hill plausibility, population falsification, and novelty/parsimony. Two of these eight can veto a hypothesis outright — e.g. a "causal" variant that turns out to be common in healthy people (gnomAD) is refuted no matter how plausible the story reads.

Why grounding matters

An LLM can write a fluent, confident explanation that is entirely fabricated. ARIADNE's defense is grounding: the AI is used only to propose and phrase hypotheses; hardcoded checks against real databases do the judging. Sensor and judge are always kept separate — the same architectural discipline used across every Aitiopoietic Systems tool.

Where the facts come from

Monarch Knowledge Graph — ARIADNE's ground truth for which genes cause which diseases and symptoms.

ClinGen — expert panels grading gene–disease confidence, Definitive → Refuted.

gnomAD — variant frequency across ~800,000 people; a common "causal" variant auto-refutes a rare-disease claim.

HPO — standard symptom dictionary, used to measure phenotype overlap.

The interactome — the protein–protein interaction map behind the S_ab network-proximity signal.

A computational physics instrument, not a language model

ARIADNE does not guess. It measures disease convergence as a quantifiable property of biological graph topology. Every candidate hypothesis is subjected to 8 physical selection pressures — including network separation, population falsification, and directionality guards — that act as a thermodynamic sieve. The fitness score is not an arbitrary confidence rating; it reflects where a hypothesis sits on a measured free‑energy landscape. Across 14,763 evaluated candidate dyads, the instrument’s output obeys the same non‑equilibrium steady‑state physics as the interactome it explores. The system is not a language model writing plausible biology — it is a computational physics instrument that isolates thermodynamically favoured mechanistic convergences.

ARIADNE v0.8.1 · Carlos Arleo, Localis AI / Aitiopoietic Systems. Architecture cloned from DNI (four-sensor + perturbed-weight ensemble + hardcoded structural checks). This Atlas visualises the CQS-ranked output of the evolutionary discovery loop — it does not itself run the genetic algorithm. Signature Views apply compound, transparent predicates over the same fields the manual filters expose. “Bench-Ready · Strict” is intentionally narrow: it requires Goldilocks specificity, druggable-direct Gate-8, a direction-matched approved drug, Open-Targets tractability, and a favourable gnomAD LOEUF score. On the current corpus it yields a handful of candidates — that is the intended behaviour, not a bug. Use Bench-Ready · Broad or IP-Mapped · Balanced for a wider net.

ARIADNE Atlas — Filter & Navigation Manual

Learn how to use ARIADNE's clinical, topological, and computational physics controls to uncover hidden disease convergences, eliminate scale-free network artifacts, and surface actionable drug repurposing candidates.

1. Canvas & Visual Navigation

The central concentric diagram maps seed diseases on the inner arc and candidate partner diseases on the outer arc, grouped into 10 clinical pathway families.

Elite Champions Button: Toggles a single-click filter showing only high-confidence Tier 1 and CQS ≥ 7 candidate hypotheses.

Neighborhood Focus Mode: Click any disease node or card to isolate its immediate mechanistic neighborhood across the interactome.

Ribbon Hover & Click: Hover over any ribbon to highlight the connection; click to open the full, expert-ready discovery dossier.

2. Clinical & Quality Filters

Confidence Tier (Tiers 1–5): Ranging from Tier 1 (Definitive / Shared Gene) down to Tier 5 (gnomAD BA1 Refuted / Biologically Incoherent).

Candidate Quality Score (CQS 0–11): An additive readiness metric. Scores $\ge 7$ indicate strong, multi-pathway hypotheses; scores $\ge 9$ are exceptional candidates for expert panels.

Global Cross-Seed Recurrence (Hubs): Filters candidates proposed independently across multiple seed runs ($\ge 3$ seeds). High hubs mark massive "gravity wells" in the interactome (e.g. RASopathy, mTOR, Lysosomal).

Clinical Pathway Family: Categorizes discoveries into 10 domains (RASopathy, Epigenetic, Lysosomal, Channelopathy, Ciliopathy, DNA Repair, Collagen, Mitochondrial, Peroxisomal, Other).

3. Computational Physics & Verification Controls

Specificity Radius (r) — Poincaré-style disc:

A radial specificity coordinate (a monotonic function of shared-pathway size), not a metric hyperbolic distance or a learned embedding.

• Specific Rim ($r \ge 0.50$): specific, hierarchical pathway convergences (few shared genes on the pathway). Down-weights generic scale-free hub shortcuts (*TP53*, Ubiquitin, general inflammation).

• Central Core ($r < 0.50$): central scale-free network hubs.

Graph Laplacian Spectral Entropy ($S_{VNE}$):

• Tightly Organized Modules ($S_{VNE} \le 1.8$): Isolates low-entropy, highly compartmentalized biological subgraphs with efficient information transfer.

• High Entropy ($S_{VNE} > 1.8$): Shows diffuse or broad pathways.

RANA Literature Verification:

• Supported Only: Filters candidates whose graph assertions are cross-referenced and confirmed by OpenAlex / Semantic Scholar PubMed citations.

• Zero Contradictions: Purges any candidate containing literature-contradicted assertions.

Causal Chain Depth ($N_{nodes}$):

• Deep Multi-Step Chains ($\ge 4$ Nodes): Ensures complete molecular-to-clinical pathways ($Variant \to Gene \to Pathway \to Metabolite \to Phenotype$).

4. Mechanism-by-Mechanism Interactive Search

Instead of searching disease-by-disease, use the bottom-right sidebar cards to navigate by mechanism:

Click Top Connected Genes: Click any gene symbol (e.g., KRAS, B3GLCT, MMUT) to filter the entire Atlas for all diseases converging on that gene.

Click Top Convergence Nodes: Click any GO term, Reactome pathway, or CHEBI metabolite (e.g., GO:0030198 Extracellular Matrix) to isolate that specific biological failure point across all 10,000+ candidates.

Coverage census
Master seeds	—
Attempted (yield of attempted)	— · —
Productive corpus	—
Empty — Monarch gap (private singletons)	— (— private)
Validation & signal
Known links recovered (recall @10 / specific)	— · —
Null model — geometry–quality corr	— (p —)
Falsifiability · novelty	—% · —% novel
Structure & physics
Rim-discoveries cluster	— @ CQS —
Cross-domain quality (median CQS)	—
Heat–entropy law R² (free-energy well)	— (S_VNE —)
Interactome
Network-proximal s_AB (z<−2) size-matched null	—% · —× vs chance
Fabrications detected & purged	— (0% remain)

Scored candidates	— over — seeds · median — partners/seed
Tiers	—
CQS bands	—
Shared-gene vs pathway-only	—% · —%
Cross-domain · novel	—% · —%
Distinct genes · pathways touched	— · —
Fabrications detected & purged	— (—% remain)
Network-proximal (interactome, z<−2) — size-matched null	—% · —× vs chance
Network-proximal — novel only (size-matched, n=7 dyads)	—%
Network-proximal — degree-preserving null stricter control: each disease gene swapped for a same-degree random gene (removes the hub inflation)	—× (95% CI —, seed-clustered) novel-only —×
Top recurrent pathways	—

1. Positive control Does it re-find known biology?	— known shared-gene convergences recovered recall@10 —% · —% for specific genes
2. Null model Is it more than chance?	geometry–quality corr — vs shuffled 0.00 permutation p — · top-quartile co-occurrence —×
3. Specificity Specific or generic overlap?	—% via a shared pathway ≤20 genes rim (r≥0.50) —% · median pathway — genes
4. Integrity Does the LLM invent genes?	— fabricated attributions detected & purged —% remain in the corpus · — records purged
5. Falsifiability Are the hypotheses testable?	—% carry a concrete falsification test
6. Novelty Discovery vs recapitulation?	—% novel · —% curated recapitulation
7. Thermodynamics Do the graphs obey physics?	heat–entropy law corr(K_t,S_VNE) — S_VNE–complexity — · real graphs —%
8. Literature audit NEW Does the field already agree?	—% of top predictions re-derive established links Discovery Gems (literature-unstudied, not structural novelty) —% un-studied vs —% broad — the filter concentrates novelty
9. Geometry ↔ interactome NEW Does the specificity geometry track the network?	corr(r, interactome z) = — (z<0 = closer than chance) network-proximal: druggability window —% vs universal —% — the pathway coordinate independently predicts interactome proximity (concordance, not an embedding)

Productive (≥1 partner) — the corpus	— (—)
Empty — Monarch data gap	— (— of attempted)
Pending — not yet run	— (—)
Attempted so far	— · yield —

Validated by orchestrate (v2.1)	— (—)
Additionally touched by legacy discovery (v2.0)	—
Covered by either architecture	— (—)
Never covered by either	—

ARIADNE — Concentric Radial Atlas

Tier Rubric

Clinical Family Legend

Thermodynamic Toxicity

Presets — one-click filter recipes

ariadne — a map of what diseases share

How to read it

How it moves when you filter

Thermodynamic Phase Space (SVNE vs. r)

Geometry ↔ Interactome Concordance

Repurposing Drug Baskets & Synergism Hub

Middle-Node Causal Meta-Graph

Cross-Domain Pleiotropy Transition Matrix

GTEx Tissue Entropy Vulnerability Matrix

Master Dashboard

ARIADNE — Master Whitepaper

Figures

Document

Corpus at a Glance

Convergence Topology

Master bottlenecks the molecular steps many separate diseases funnel into

The gene map — the same gene, three very different meanings

Generic solvent nodes (≥10 families) — housekeeping noise, shown so you can filter it out

Structure & Physics

Topology clusters (PCA + KMeans, unsupervised)

Cross-domain vs same-domain

Thermodynamic law

Validation & Signal

Corpus & Knowledge-Graph Coverage

Two architectures — the honest coverage picture

Why “empty” is a finding, not a failure

Candidate Quality Score (CQS v3)

Interpretation bands

What Is ARIADNE?

The core idea: breed hypotheses, don't look them up

Genome vs. genotype

Why grounding matters

Where the facts come from

A computational physics instrument, not a language model

ARIADNE Atlas — Filter & Navigation Manual

1. Canvas & Visual Navigation

2. Clinical & Quality Filters

3. Computational Physics & Verification Controls

4. Mechanism-by-Mechanism Interactive Search

Thermodynamic Phase Space (S_VNE vs. r)