Quantum Pharma

Calibrated on 1,601 real IBM Heron jobs — not simulation. This build’s own hardware validation ran on ibm_kingston. See the corpus →

Kill the bad molecule early.

Most drug candidates fail — and the later they fail, the more they cost. Quantum Pharma runs the classical pre-screens that wash out the non-starters fast, then spends the heavy GPU step only on the survivors that earn it. Every screen is sealed.

Screen a molecule ↓ Open the live run

The problem we exist to solve

Nine in ten candidates fail.
The expensive ones fail last.

  • Failure in the clinic costs billions A molecule that washes out in Phase II took years and a fortune to get there. The cheapest failure is the one you catch at the bench.
  • The search space is astronomical There are more drug-like molecules than atoms in the solar system. You can't synthesize your way through it — you have to screen first.
  • Regulators want the trail Every screen, every decision, needs to stay inspectable. A result you can't re-derive is a result a regulator won't accept.

What it mints

Four sealed steps.
Molecule to IND.

The live Pharma mint is four sealed steps. A drug-likeness pre-screen washes out non-starters first; only the survivors go on to the GPU co-fold.

  • Fold the target Build a 3D structure with Boltz-2 (MIT) so later steps work from a real shape — not a guess.
  • Generate candidates Propose small molecules with GenMol (NVIDIA). Wash out non-starters fast — drug-likeness and structural alerts, before anything is co-folded.
  • Predict affinity Estimate binding with Boltz-2 co-fold (MIT). Predicted — not a wet-lab measurement. Kill the bad molecule early.
  • Search conformers Search the winning molecule’s low-energy shapes. The calibrated router picks the machine for the job and the sealed record names which one ran it. A simulation today, not a quantum-advantage claim.

The discovery bench

Profile a candidate. Seal the verdict.

A working pre-screen bench. Enter a molecule's descriptors and it checks Lipinski's Rule of Five and Veber's rules, plots each property against its limit, flags the ADMET liabilities, and seals the screen — the deterministic gate that decides what's worth escalating. Measured against known actives, this gate passes 93% of our own generated candidates but only 3.1% of 97 known potent KRAS G12D inhibitors — real G12D binders are beyond-Rule-of-5 by design, and a Ro5 screen rejects them. 65.1% of 415 EGFR, 59.1% of 232 MAT2A. That is the honest limit of the screen.

Descriptors · Lipinski · Veber

Lipinski Ro5 · VeberDrug-like
0 violations
In-browser demonstration. Deterministic RDKit-style pre-screen, sealed locally with WebCrypto SHA-256. Local illustration — not a live quantum run.

How a discovery run is made

Fold. Candidates.
Affinity. Conformers.

  1. Fold the target

    Boltz-2 (MIT) builds the 3D structure from the target’s sequence and its multiple-sequence alignment, and scores its own confidence residue by residue.

  2. Generate candidates

    GenMol (NVIDIA) proposes candidates. Classical gates — Lipinski, Veber, QED and structural alerts — wash out the ones that can't be oral drugs: fast, deterministic, recomputable by hand. The survivors go on to the co-fold.

  3. Predict affinity

    Boltz-2 co-fold (MIT) scores binding for the survivors and keeps the strongest. Predicted — not a wet-lab measurement.

  4. Search conformers — then seal

    A QAOA search over the winning molecule’s low-energy shapes, run as a simulation, with the calibrated router’s verdict sealed beside it. Every step gets a signature and watermark on a tamper-evident record a third party can verify. Anchored daily in Bitcoin and on Base.

Built for audit

Every screen
gets sealed.

Signature and watermark on by default — written so compliance can open the trail later. Sealing is enforced; you don't turn it on after the fact.

Open a discovery run

Walk a live fold-to-binding pass, or send a target and we will run it.

Open live discovery run