Kytos Observatory · quantum workstream

For 40 years, KRAS was called undruggable.

The pocket that finally cracked it wasn't at the active site — it was allosteric, encoded in the protein's contact topology. We put that topology onto a real quantum processor and measured what survives.

84–101% correlation magnitude survives
(KRAS 0.84× · myosin 1.01×)

≈0 direction preserved
(Bloch cosine)

graph-v1 ibm_fez KRAS + myosin + control 20 qubits 1024 shots

The experiment

Protein contact graphs — which residues touch which — carry the topology that allosteric drug sites live in. We take each target's coarse-grained contact graph, keep the biologically meaningful core (catalytic "source" supernodes, the known allosteric pocket, and the strongest couplings between them), and submit it to Moth's graph-v1 engine as a 20-qubit coupling map.

The KRAS G12C and cardiac myosin 20-qubit coupling graphs, nodes colored by biological role — The two allosteric cores submitted to hardware — red nodes are catalytic-site supernodes, blue the known drug pockets, purple overlap both. Node→residue mapping in `metrics/graphs/`.

Watch it decohere

Each coupling map ran twice: emu (Moth's Aer emulator) and qpu (IBM ibm_fez, 1024 shots, full two-qubit tomography on every graph edge). Edge thickness is the measured correlation strength — flip the switch and watch what real hardware does to it.

Hear it decohere

The same measurement, rendered as sound. Each of the 50 graph edges plays one note; pitch is the measured correlation strength, timing is fixed by edge order. Play them in turn — the melodies differ exactly where the hardware redistributed the correlations.

emulator

ibm_fez

emulator, echoed through a quantum circuit

Third track: the emulator melody passed through Moth's retrocausal-echo engine — a multi-tap delay whose tap weights are measured on a quantum circuit (Aer simulation). Not hardware: an honest echo of what the ideal device would return.

The emulator correlation matrix of KRAS morphing toward the hardware measurement, rendered through quantum rotation gates — The emulator's correlation map morphing toward the hardware's, re-rendered through quantum rotation gates by Moth's `telablur` engine — decoherence, painted by the same class of machine that produced it.

What the hardware did to it

The numbers behind the widget — three findings, all verifiable against the raw job receipts below.

Two-qubit correlation MAGNITUDES survive real hardware: qpu/emu ratios 0.84 (KRAS) and 1.01 (myosin). Edge-agreement scores are comparable or slightly higher on hardware.
Per-qubit Bloch-vector DIRECTIONS decorrelate between emu and qpu (median cosines -0.01, -0.27) - the dominant decoherence signature is directional scrambling, not magnitude loss.
Random control (degree-matched 20-node/50-edge graph, ibm_fez db1d1h2vog1s73fhvp00): emulator edge-corr 0.184 vs KRAS 0.338 / myosin 0.419 - real topology is ~2x more correlated in emulation. On hardware the separation washes out (control qpu 0.441 vs KRAS 0.284): topology-specific fine structure is the first casualty of NISQ noise.

Technical detail — what the engine actually prepared

The engine's prepared state is NOT a pure graph state (single-qubit |r| up to 1.0 vs 0 for a graph state); emu mode also deviates from the exact CZ-product baseline (metrics/graphs/*_exact.json). Interpretation is 'Moth correlated-state protocol driven by our protein topology', not graph-state fidelity.
Role-level decomposition (source / known-site / connector edges) is inconsistent across targets and underpowered (n=5-6 known edges) - reported as noise, not signal.
KRAS role means (qpu): source 0.305 / sourceknown 0.253 / connector 0.295 / known 0.237. Myosin: source 0.414 / sourceknown 0.281 / known 0.509 / connector 0.499.

Bar charts comparing edge correlation magnitude and edge agreement between real protein graphs and a random control, on emulator vs ibm_fez — Real vs random. On the emulator the protein topologies are ~2× more correlated than a degree-matched random graph; on hardware the separation washes out.

Full tomography: correlation matrices + Bloch vectors

Per-edge two-qubit correlation heatmaps for KRAS and myosin — emulator, ibm_fez, and the difference — Per-edge |correlation vector|, emulator vs hardware vs the difference. Magnitudes transfer; the pattern does not.

Scatter plot of encoded edge correlations vs values read back by ibm_fez through the qpixl engine, showing near-identity agreement — Write the measurement back through the hardware: the 50 edge correlations encoded as qubit angles by `qpixl` and read back on `ibm_fez` — Pearson 0.98, mean error 0.057. The values survive the machine; the directional structure is what's lost.

Per-qubit Bloch vector magnitudes on emulator vs ibm_fez for both proteins — Per-qubit Bloch |r|. Both backends produce structured, polarized single-qubit states — but their directions decorrelate (median cosine ≈ 0).

Receipts

Every claim on this page resolves to a submitted job. Each job ID links to the raw engine response committed in the repository — the file itself carries the ID, so the link is the verification.

run	moth job (qpu)	moth job (emu)	IBM job
KRAS G12C	`252b239d-30c6-4400-bf45-7f78002993a6`	`21950e34-b55c-4f54-9fdb-be42eaa36ae9`	`db1crfrid5ic73eqvk1g`
cardiac myosin	`54b278da-eb53-416b-bae7-f92f14b15769`	`f3629ed6-b52a-4b4e-9582-7c0fd567b14c`	`db1cssjid5ic73eqvlc0`
random control	`6445e00e-c282-4aa3-87ea-a156a372274f`	`f097e67a-3905-449b-b152-d44439ff76f9`	`db1d1h2vog1s73fhvp00`
correlation read-back qpixl-v1	`1260e041-28c1-45c3-b224-96755f65130b`	—	`db1e6o6egvvc73bhgtu0`

How to verify independently

With any Moth API key, the platform replays the record server-side:

curl -H "Authorization: Bearer $MOTH_API_KEY" \
  https://api.mothquantum.com/api/v1/jobs/252b239d-30c6-4400-bf45-7f78002993a6/status

KRAS G12C: 20 nodes / 50 edges extracted from the 32-supernode coarse graph (GTP-site sources + Switch-II/Sotorasib known-site supernodes + strongest couplings).
cardiac myosin: 20 nodes / 42 edges extracted from the 56-supernode coarse graph (source + known-site supernodes + strongest couplings).

The record behind the run

c008 is the latest step of the Cleveland Clinic / GQAI quantum-allostery workstream — the same publish-everything discipline as the Virtual Cell Challenge runs on this site.

c001-ctrw-full-vs-coarse 2026-09-15 Classical CTRW baseline + spectral coarse-grain Spearman gate (≥0.8).

c002-ctqw-coarse 2026-09-15 CTQW on Phase 1 coarse graphs; classical CTRW comparator; myosin Phase 1 ρ=0.823 kept visible as tightest compression margin.

c003-ctqw-compression-audit 2026-09-15 Full vs coarse CTQW; myosin P1 ρ=0.823 kept visible. Mavacamten pocket labels updated to literature contacts.

c004-randomization 2026-09-15 Edge-rewire null on full-graph CTQW; myosin P1 ρ=0.823 visible.

c005-signal-sweep 2026-09-15 Grid H×T×resolution; myosin P1 ρ=0.823 visible.

c006-circuit-packaging 2026-09-15 Qiskit packaging fidelity vs eigh; Braket/Classiq exports only.

c007-topology-levers 2026-09-15 Cutoff/multiscale/community/residual CTQW; myosin P1 ρ=0.823 visible.

c008-moth-qpu-tomography 2026-10-04 Moth graph-v1 engine, emu + qpu modes on ibm_fez. 20-qubit functional cores of KRAS G12C and cardiac myosin contact graphs. Real-hardware tomography receipts for the Moth Hack virtual hackathon (submissions close 2026-10-05 23:59 PT).

Honest caveats

What this page does not claim.

graph-v1 internals are Moth's black box; the coupling_map parameterizes their protocol, circuit not exposed.
emu vs qpu disagreement cannot be attributed purely to hardware noise without knowing the emu model.
No claim is made that hardware correlations recover the known allosteric site; the CTQW ranking analysis (c002-c007) remains the biological method - this run supplies the hardware receipt + decoherence characterization.
Pre-declared significance gate from the GQAI workstream remains unmet; nothing here changes that.

Artifacts

Code, graphs, request payloads, raw engine results, analysis, and the reproduction scripts live in the repository: experiments/cleveland/c008-moth-qpu-tomography/. The classical+CTQW allostery pipeline upstream is src/cleveland/.