Stage A→C closed on ibm_marrakesh: same-day full pass —
raw junction gap 0.185 · Aer retention 0.749 ·
separable merge-fail 0.184.
Locked junction_ryy · 156 qubits · jobs d9qubi7pemts73crsa2g+.
Classical methods on the same frozen table disagree. Multi-logic architecture evidence — not a drug-discovery claim.
Public WOW article →
In early drug discovery, a molecule is judged by many ADMET properties at once -- not by one docking score. Those properties pull in different directions. The usual classical habit is to weight them into one number, or to apply hard filters one after another. This job asks a different architectural question: can those logics sit together inside one auditable quantum run on real hardware?
Freeze a public-proxy desirability table (9 compounds x 17 TDC-aligned ADMET endpoints), encode it as a polycontextural landscape
(9 x 17 data qubits + absorb / clear / safety junctions = 156 qubits), run coupled vs no_cross control under locked
junction_ryy geometry on IBM Fez, and read back a provisional lead ranking from the coupled arm -- always compared with classical baselines on the same numbers.
Pharma teams already live with multi-objective ADMET conflict (permeability vs toxicity, CNS exposure vs CYP liability, oral absorption vs DILI risk). The point of this exemplar is not "we found a clinical candidate." The point is a bounded, public instance of multi-logic decisioning with IBM job IDs -- the shape of an Exemplar Replay a partner can later fill with their own scores under the same schema.
Not wet-lab discovery. Not a TDC leaderboard submission. Not a claim that QPC beats every classical ADMET optimizer. Scores are curated public-proxy desirabilities for architecture comparison. See the database page for the frozen table and provenance.
no_cross). Mitigation / SAR is secondary and does not replace the primary claim.(1) Classical methods disagree on this table — why multi-logic architecture matters. (2) Stage C2 on ibm_marrakesh cleared every pre-registered bar same day — the WOW closure.
| Rank | Compound | Class | Lead score |
|---|---|---|---|
| 1 | caffeine | cns_tool | 0.736 |
| 2 | acetaminophen | analgesic | 0.669 |
| 3 | propranolol | cardiovascular | 0.588 |
| 4 | ibuprofen | nsaid | 0.572 |
| 5 | fluoxetine | cns | 0.568 |
| 6 | terfenadine | withdrawn | 0.567 |
| 7 | digoxin | cardiovascular | 0.530 |
| 8 | cisplatin | oncology | 0.432 |
| 9 | doxorubicin | oncology | 0.343 |
Reading this: favorable oral / CNS-friendly profiles rise; high-toxicity oncology agents fall. That is a sanity check on the encoding, not a clinical recommendation.
Equal-weight sum ranks: caffeine, acetaminophen, propranolol, ibuprofen, fluoxetine, digoxin, terfenadine, cisplatin, doxorubicin.
Sequential hard filters (hERG, Ames, DILI, P-gp, bioavailability gates) keep only: caffeine, propranolol, ibuprofen, fluoxetine. Acetaminophen (weighted-sum #2) is killed by the DILI gate -- the classic multi-logic trap.
Pareto front (maximize all 17): caffeine, fluoxetine, acetaminophen. Kendall tau (weighted-sum vs sequential) = 0.50 -- real disagreement, not noise.
Hardware ranking order differs from classical equal-weight in places; that disagreement is useful for the architecture conversation. It is not advertised as quantum sampling advantage.
no_cross:
gap 0.185 · retention 0.749 · separable merge-fail 0.184.
Jobs d9qubi7pemts73crsa2g+ / separable d9qubk8pdb6s73e45q3g.
Primary metrics on raw counts. Fez Stage B (~0.18) remains supporting evidence.
Public article: QPC_ADMET_WOW_ARTICLE.html.
Allowed: joint multi-logic ADMET decisioning under a published conflict table; Exemplar Replay offer; IBM audit trail.
Forbidden: "we discovered a drug," TDC leaderboard win, invented ROI, calling mitigation QEC, or silent replacement of raw architecture metrics by mitigated counts.
Frozen score table and column definitions: ADMET_LEAD_DATABASE.html