NEXUS-9 · system architecture

The pipeline is fourteen stages, not five.

The production orchestrator nexus9 dossier <target> runs intake → supervised fold → mechanism gate → design summary in one command. Around that spine sit the rest of the platform: co-evolution, contact prediction, the nine-layer ENNEAD solver, thirteen modality routes, patient and trial matching, the 24-database evidence chain, and the TJ Compassion audit. Below is every stage, the command that runs it, and what it actually returned on a real target.

The four phases

Twelve of the fourteen stages are target-specific; the other two validate the canonical design set. Some run automatically inside nexus9 dossier; the rest are separate commands in the same 39-command binary and compose into the same evidence chain.

Phase I

Resolve the target

Everything begins with a verified identifier. No target is described from memory — the sequence, organism, length and mass come from UniProt at run time, and the structural evidence is pulled from public databases before a single prediction is made.

01

Live target intake

nexus9 live <UniProt>

Fetches the canonical record from UniProt and resolves the same protein across AlphaFold, the PDB, PubChem and ClinicalTrials.gov. Unknown accessions fail loudly instead of being guessed at.

Input
UniProt accession or gene symbol
Output
ID · gene · organism · length · molecular weight
Real run (P00533): EGFR · Homo sapiens · 1,210 aa · 134,277.0 Da.
02

Deep MSA and co-evolution

nexus9 karen deepmsa <UniProt>

Builds a real deep multiple-sequence alignment from OMA orthologs and stores the co-evolutionary signal. This is the raw statistics that contact prediction and DCA consume — not a canned alignment.

Input
UniProt accession
Output
Deep MSA + stored co-evolution signal
03

Contact map

nexus9 dca compute <msa.fasta> [uniprot] nexus9 dca contacts <uniprot> nexus9 contactnet check <uniprot>

Two independent contact routes. DCA implements the Morcos 2011 direct-coupling method and ranks pairs by an average-product score; ContactNet is the neural contact predictor. They are mathematically different, so agreement between them is meaningful. DCA now auto-generates: if no coevolution data exists for a target, dca contacts fetches orthologs, builds the MSA, computes the direct couplings and saves them — no pre-computed file required.

Input
Deep MSA / sequence
Output
Ranked residue-residue contacts
On disk: 95 ContactNet checkpoints, from 364-byte seed maps to 5,644-byte trained maps.
Phase II

Solve the structure

Structure is solved two ways and labelled honestly. The supervised NIGHTINGALE path is the production route; the ENNEAD nine-layer solver is the research route. Both report TM-score and RMSD against the experimental reference, and both numbers travel with the dossier.

04

Supervised fold — NIGHTINGALE

nexus9 fold <UniProt>

Predicts the three-dimensional structure and scores it against a known experimental structure. TM-score measures fold agreement (1.0 perfect, above 0.5 the same fold); RMSD measures average atomic displacement in ångströms.

Input
UniProt accession + reference PDB
Output
Coordinates · TM-score · RMSD · fold status
Real run (P00533, corrected 2 Nov 2026 build): reference PDB 1XKK · TM 0.9706 · RMSD 1.64 Å · Status OK — supervised against a reference structure, so a reconstruction rather than a blind prediction. The figure is the peak TM along a 1,000-iteration trajectory (Papa's rule). The earlier 0.6502 was lowered by a numerically unstable 3×3 superposition in the scorer; the quaternion alignment fix raises the same run to 0.9706. The 0.7393 figure came from a stale ContactNet checkpoint and remains retired.
05

Ab initio fold — ENNEAD

nexus9 ennead fold nexus9 ennead resolve

A nine-layer constraint solver that folds without a template, then refines the result. It is research code: it can collapse and stall, and it is not the production path. It is included because a sovereign system should be able to attempt the problem without borrowing a template.

Input
Sequence / constraint set
Output
De novo coordinates · constraint residuals
06

Domains and pockets

nexus9 fold-domains <target|fasta> nexus9 pocket <target>

Large proteins are decomposed into domains and folded hierarchically; the pocket finder locates a binding cavity and recommends a degradation route. Known pocket targets: KRAS, TP53, SNCA, TAU, BCL2. Structural context is cross-checked against the PDB structural database and AlphaFold models.

Input
Target or FASTA / Nonary DB record
Output
Domain folds · pocket coordinates · route
Phase III

Decide and design

The gate is the ethical centre of the system. Degradation is only valid for a toxic gain of function; for a missing or weak protein it would make the patient worse. The system refuses the wrong action before any molecule is drawn. After design, two validation stages score the nine canonical designs for binding and toxicity — both explicit that they are estimates, not experiments.

07

Mechanism gate

nexus9 therapeutic <target>

Reads the disease mechanism — gain of function, loss of function, aggregation — and decides whether degradation is valid. A loss-of-function target is protected and routed elsewhere; the refusal is code, not prose.

Input
Target + mechanism class
Output
PASSED (degradable) or BLOCKED (protected)
Real run (P00533): GAIN-OF-FUNCTION → AUTOTAC eligibility PASSED, 90% confidence. Counter-example, verified: UBE3A (Angelman) and FXN (Friedreich's ataxia) are LOSS-OF-FUNCTION → BLOCKED, no degrader issued. HPRT1 returns BLOCKED on mechanism-unknown, sending it to primary research instead of a design.
08

Modality routing

modality_db · 13 routes

Not every disease wants a degrader. The router selects among thirteen modalities: AUTOTAC, repurposing, polypharmacology, botanicals, medicinal fungi, natural proteins, phage, microbiome, CAR-T, mRNA, psychedelics, gene therapy and peptide.

Input
Mechanism + target biology
Output
Selected modality + rationale
09

Molecular design

nexus9 ligand <target> nexus9 mass-reduce <target>

Assembles a bifunctional chimera: a target-binding ligand, a linker, and an autophagy-targeting ligand that recruits p62/SQSTM1. It computes total mass, predicted logP and permeability, and states when the molecule is too large to cross membranes.

Input
Eligible target + binding site
Output
TBL + linker + ATL · MW · logP · permeability
Real run (P00533): TBL gefitinib analog 446.0 Da + PEG4 linker 194.2 Da + ATL YTK-105 257.0 Da = 897.2 Da; logP 0.9; permeability LOW (MW > 800, not CNS-penetrant) — the tool reports the limitation rather than hiding it.
10

Docking estimate

nexus9 dock

Scores all nine canonical AUTOTAC designs against their warheads' expected IC50 and ranks them into affinity classes. It is an estimate of whether the warhead binds — not a docking simulation of the full chimera.

Input
Nine canonical AUTOTAC designs
Output
IC50 estimate · affinity class 1–9
Real run: six of nine score 8/9 STRONG (osimertinib EGFR C797S 842.84 Da, IC50 6.0; enzalutamide AR/VHL 802.97 Da, IC50 2.0), three score 6/9 MODERATE. The tool's own note: IC50 is the warhead's known binding, not the full AUTOTAC — real docking needs AutoDock/Glide-class confirmation.
11

Toxicity screen

nexus9 tox

Rule-based structural alerts over the same nine designs — hERG cardiotoxicity, teratogenicity, solubility — with a 1–9 risk score. Predicted, not measured.

Input
Nine canonical AUTOTAC designs
Output
Risk score · structural alerts
Real run: all nine pass Lipinski; eight score 9/9 total risk, the AR/VHL design 5/9. Every thalidomide-class design carries a HIGH teratogenicity alert — the kind of risk a degrader programme must see before synthesis.
Phase IV

Evidence, patients, verification

A design is worthless without the evidence behind it. This phase assembles the twenty-four biomedical databases, matches the target to real patients and recruiting trials, applies the TJ Compassion / L.O.V.E. gate, and writes the dossier in native Rust.

12

Knowledge assembly — KAREN

nexus9 karen query <name> nexus9 karen pipeline <uniprot>

KAREN is the zero-search co-processor: a constant-time lookup over the assembled evidence, with live connectors to UniProt, PubChem, RCSB PDB, ClinicalTrials.gov, ChEMBL, NCBI Gene, Reactome and OpenTargets. The twenty-four biomedical databases are surfaced through it into one record.

Input
Target, design, pathway
Output
24-database evidence chain
Coverage in the binary: 101 clinical phenotypes, 36 curated trials, 77 pharmacogenomic rules, 9 patient variant records.
13

Patient and trial matching

nexus9 patient · nexus9 pgx nexus9 ptrials for <patient_id>

The eleven-dimensional Nonary matcher scores a patient against recruiting trials: variant lookup, gene and indication filter, status, mutation specificity, mechanism alignment, line-of-therapy fit, phase readiness, recruitment freshness, biomarker match, endpoint relevance and safety — plus PDB structural coverage and line-timeline fit.

Input
Patient variant + indication
Output
Ranked trials · PGx risk · structural coverage
14

Verification and report

nexus9 audit compassion nexus9 dossier-pdf <target>

Every fact is stamped with its source, or labelled unverified — there is no middle state. The TJ Compassion audit scans the workspace for harm, destruction and deceit before release; the L.O.V.E. gate (Learn, Observe, Verify, Explain) runs on individual claims. The dossier is rendered to PDF natively, with no Python in the path.

Input
Full evidence chain
Output
Audited, citation-referenced dossier
Real run: the P00533 dossier states mechanism, modality, warhead, 897.2 Da design and 90% confidence — with the fold caveat and the scope note attached.

The 24 evidence databases

These are the biomedical databases the evidence chain draws from. Of the binary's 31 database modules, 24 are biomedical evidence; the other seven are pattern, game and infrastructure (arc patterns, crystal palace, modality, nonary games, physics anomalies, resonance patterns, string resonance) and are not counted as clinical evidence.

Adverse events — event and severity records Antimicrobial resistance — resistance determinants AUTOTAC designs — canonical bifunctional chimeras Biomarkers — diagnostic and prognostic markers CDx diagnostics — companion diagnostic links Clinical guidelines — treatment guidance Clinical phenotypes — 101 disease indications Clinical trials — 36 curated trials + live feed Drug interactions — CYP450 and transporter matrix Endothelial / vascular — vascular biology Human genome — 81 therapeutic genes (9 × 9) Interaction network — protein and pathway interactions Patent / FTO — freedom-to-operate landscape Microbiome — microbiome therapeutics NIGHTINGALE structural — fold and structure records Orphan disease — rare disease registry Pathways — Reactome-anchored pathways Patient variants — variant-level patient records PDB structural — experimental structures Pharmacogenomics — 77 PGx rules Pharmacology / PK — absorption and clearance Psychedelics — 10 PubChem-verified compounds Resistance — acquired resistance mechanisms Vaccine / cell therapy — immunotherapeutic routes

Engines and subsystems

The pipeline is not a single model. It is a set of engines with different mathematics, wired together so that each checks a different property of the answer.

SubsystemWhat it doesEvidence
NIGHTINGALE foldSupervised protein structure prediction with TM/RMSD scoringProduction path
ENNEAD solverNine-layer ab initio constraint solver, fold / resolve / benchResearch path
ContactNetNeural residue-contact prediction95 checkpoints
DCADirect coupling analysis, Morcos 2011, AP contact scoreIndependent of ContactNet
KARENZero-search co-processor over the evidence chain; 8 live connectorsUniProt · PubChem · PDB · ChEMBL · Gene · Reactome · OpenTargets · trials
NELLY swarm81-node PHI-constrained consensus brainPHI 1.618034; coherence > 0.90
ptrialsEleven-dimensional patient-to-trial matcher36 curated trials + live feed
QEYNOSFive-layer sovereign encryption (Argon2id · AES-256-GCM · nonary diffusion · Shamir · photonic seal)81 shards, threshold 27
SABRE-9Nonary VM, PTX/CUBIN compilation, photonic blockchainRTX 5060 path

Scale, verified from the binary: 39 advertised commands, 141 dispatch arms, 219 Rust source modules, 95 ContactNet checkpoints, 24 evidence databases — against the release build of 22 October 2026, 09:41. The README's older figures (139 commands, 192 modules, 43 checkpoints) are stale; the numbers above are what the current help and source tree report.

The rest of the binary

The dossier orchestrator is one path through a much larger program. These are the current build's other working surfaces — separate commands, not hidden steps inside nexus9 dossier.

CapabilityCommandWhat it does
Ortholog routenexus9 ortholog pipelineEBI Proteins + UniProt → BLOSUM62 Smith-Waterman → MSA → DCA → ENNEAD fold → PDB. Fetches up to 500 orthologs to solve the coevolution-depth limit.
Proteome scalenexus9 proteomePartitions a 20,431-protein input across 81 Nelly nodes (27 non-empty at the 18/27 floor/ceiling) with 16 workers.
Batch reportsnexus9 gen-reportsGenerates therapeutic reports from a prepared target list.
Trainingnexus9 train · train-n9 · sovereignTernary funnel training, ContactNet N9 training, sovereign autotrain loop.
GPU / VMnexus9 sabre-ptx · sabre-gpu · compile · runNeo9 source → bytecode → PTX → CUBIN, targeting the RTX 5060.
Sovereign stacknexus9 qeynos · blockchainFive-layer encryption and the QEYNOS photonic blockchain seal.
Maintenancenexus9 surgery81-node surgery agent: scan, sweep, heal, test-fix, with rollback if cargo check fails.
Code generationnexus9 ragRAG generate / validate / improve / stats over the training corpus.

Honest boundary: the proteome and batch-report runners need prepared input files (human_proteome_training.json, top_targets_live.json). The docking and toxicity screens cover the nine canonical designs, not a newly generated chimera.

One target, actually run

Verbatim excerpts from a real run on 22 October 2026. Nothing below is illustrative copy — it is what the binary printed.

$ nexus9 dossier P00533

▌ STAGE 1: LIVE UNIPROT FETCH
  ✓ ID: P00533
  ✓ Gene: EGFR
  ✓ Organism: Homo sapiens
  ✓ Length: 1210 aa
  ✓ MW: 134277.0 Da

▌ STAGE 2: NIGHTINGALE SUPERVISED FOLD (production iterations)
  ✓ PDB: 1XKK  TM=0.9706  RMSD=1.64 Å  Status=OK

✓ STRUCTURAL EVIDENCE: TM=0.9706 (>= 0.5, correct-fold threshold).

▌ STAGE 3: THERAPEUTIC / DESIGN PIPELINE
  Mechanism: GAIN-OF-FUNCTION
  Modality: AUTOTAC degrader
  Recommendation: P00533: AUTOTAC-mediated degradation
  Confidence: 90%
  Warhead: Gefitinib analog | Design MW: 897.2 Da

$ nexus9 ligand P00533
  TBL: Gefitinib analog (446.0 Da)
  ATL: YTK-105 p62/ZZ binder (257.0 Da)
  Linker: PEG4 (tetraethylene glycol) (194.2 Da)
  TOTAL: 897.2 Da  (< 1200 Da)
  LogP: 0.9  Permeability: LOW (MW > 800 ... NOT CNS-penetrant)

Read the fold number correctly. TM 0.9706 / RMSD 1.64 Å is the supervised path, folded against PDB 1XKK, reporting the peak TM along a 1,000-iteration trajectory. It measures how well the engine reconstructs a known structure, not how well it predicts an unseen one.

Why the number moved: the earlier 0.6502 was not a folding failure. The scorer's 3×3 Kabsch alignment returned a bad frame on near-degenerate covariance; identical coordinates scored TM 0.0992. The quaternion superposition fixes that, and nexus9 fold P00533 now returns TM 0.9706 / RMSD 1.64 Å. The older 0.7393 came from a stale-checkpoint bug and remains retired. The benchmark fold (1M17) is TM 0.9959 / 0.42 Å on the corrected build.

What it does not claim

The same discipline that makes the evidence chain useful makes its limits explicit.

Ab initio folding is research, not production. ENNEAD can collapse and stall; the supervised path is the production route.

The CUDA GPU paths do not execute — they report zeros and fall back to CPU. The working GPU path is the Nonary Vulkan driver targeting the RTX 5060.

Proteome-scale and batch report generation need prepared input files (human_proteome_training.json, top_targets_live.json); they are not turnkey from an empty folder.

Nothing here is clinical advice. Every design requires experimental validation, cell-based assays and regulatory review. No regulatory approval is claimed or implied.